commit files to HF hub

CMU Advanced NLP 2024 (1) Introduction to NLP/CMU Advanced NLP 2024 (1) Introduction to NLP.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:28be5d1d73b923cf7a91a66e6d77b5862dbf89020af51b887091f9aeedfd7b94
+size 66391760

CMU Advanced NLP 2024 (1) Introduction to NLP/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=6NeTO61qc4M",
+    "title": "CMU Advanced NLP 2024 (1) Introduction to NLP"
+}

CMU Advanced NLP 2024 (10) Retrieval and RAG/CMU Advanced NLP 2024 (10) Retrieval and RAG.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81d3898858e07098de0177379421d5ba13d45835e73a1a41b3b7696c17d01774
+size 54642972

CMU Advanced NLP 2024 (10) Retrieval and RAG/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=WQYi-1mvGDM",
+    "title": "CMU Advanced NLP 2024 (10) Retrieval and RAG"
+}

CMU Advanced NLP 2024 (10) Retrieval and RAG/transcript.vtt

@@ -0,0 +1,4036 @@
+WEBVTT
+
+00:00:00.040 --> 00:00:03.880
+so today I'm going to talk about
+
+00:00:01.319 --> 00:00:06.680
+retrieval and retrieval augmented
+
+00:00:03.880 --> 00:00:09.040
+generation so if we look at our standard
+
+00:00:06.680 --> 00:00:10.880
+prompting flow normally what we do is we
+
+00:00:09.040 --> 00:00:14.160
+combine together a prompt template with
+
+00:00:10.880 --> 00:00:16.600
+an input so if we say please answer this
+
+00:00:14.160 --> 00:00:18.720
+question I think Vin Diesel has been a
+
+00:00:16.600 --> 00:00:21.000
+voice actor for several pictors in TV
+
+00:00:18.720 --> 00:00:24.000
+series do you know what their names
+
+00:00:21.000 --> 00:00:25.400
+are we could get a response from a
+
+00:00:24.000 --> 00:00:26.840
+language model but there are several
+
+00:00:25.400 --> 00:00:30.840
+problems with
+
+00:00:26.840 --> 00:00:33.680
+this the first is accuracy issues
+
+00:00:30.840 --> 00:00:36.160
+the models generally have a knowledge
+
+00:00:33.680 --> 00:00:38.879
+cut off so the parameters are usually
+
+00:00:36.160 --> 00:00:41.120
+only updated to a particular time so for
+
+00:00:38.879 --> 00:00:43.200
+example if a new Vin Diesel TV series
+
+00:00:41.120 --> 00:00:44.960
+comes out then the model that was
+
+00:00:43.200 --> 00:00:47.440
+trained up to a certain time Point won't
+
+00:00:44.960 --> 00:00:51.000
+be able to know anything about
+
+00:00:47.440 --> 00:00:53.600
+it there's also issues of private data
+
+00:00:51.000 --> 00:00:55.320
+so data stored in private text or data
+
+00:00:53.600 --> 00:00:57.840
+repositories is not suitable for
+
+00:00:55.320 --> 00:01:02.600
+training for a number of reasons number
+
+00:00:57.840 --> 00:01:05.199
+one it's not available to to particular
+
+00:01:02.600 --> 00:01:07.799
+language model training providers such
+
+00:01:05.199 --> 00:01:10.720
+as you know open AI or Google or anybody
+
+00:01:07.799 --> 00:01:13.840
+else like this the second thing is
+
+00:01:10.720 --> 00:01:16.799
+Access Control issues so even if you're
+
+00:01:13.840 --> 00:01:17.840
+within an organization that has lots of
+
+00:01:16.799 --> 00:01:20.799
+private data and you can train a
+
+00:01:17.840 --> 00:01:22.600
+language model on that certain people in
+
+00:01:20.799 --> 00:01:24.200
+the organization may have access to
+
+00:01:22.600 --> 00:01:27.640
+certain varieties of data and other
+
+00:01:24.200 --> 00:01:29.400
+people may not so it's not just solely
+
+00:01:27.640 --> 00:01:31.520
+an issue of third party providers it's
+
+00:01:29.400 --> 00:01:33.840
+an issue of organization level Access
+
+00:01:31.520 --> 00:01:36.159
+Control in
+
+00:01:33.840 --> 00:01:38.920
+general in addition there are learning
+
+00:01:36.159 --> 00:01:40.320
+failures so even for data that the model
+
+00:01:38.920 --> 00:01:42.640
+was trained on it might not be
+
+00:01:40.320 --> 00:01:44.399
+sufficient to get the right answer and
+
+00:01:42.640 --> 00:01:47.799
+this is particularly the case for very
+
+00:01:44.399 --> 00:01:52.320
+very large uh training data sets and
+
+00:01:47.799 --> 00:01:53.920
+models that are you know modestly sized
+
+00:01:52.320 --> 00:01:55.880
+because the models very often won't be
+
+00:01:53.920 --> 00:01:58.360
+able to learn from a single look at a
+
+00:01:55.880 --> 00:02:02.039
+particular fact or or whatever else like
+
+00:01:58.360 --> 00:02:02.039
+this especially if iter early in
+
+00:02:02.159 --> 00:02:08.160
+training another thing is even if the
+
+00:02:05.240 --> 00:02:10.599
+answer is correct it might not be
+
+00:02:08.160 --> 00:02:13.440
+verifiable so you might want to be very
+
+00:02:10.599 --> 00:02:15.000
+sure that the model is not making any
+
+00:02:13.440 --> 00:02:17.640
+accuracy
+
+00:02:15.000 --> 00:02:19.040
+problems and so in order to do that very
+
+00:02:17.640 --> 00:02:21.879
+often a human will want to go back to
+
+00:02:19.040 --> 00:02:21.879
+the source of the
+
+00:02:22.200 --> 00:02:27.319
+data so to solve this there's a method
+
+00:02:25.480 --> 00:02:29.200
+called retrieval augmented generation
+
+00:02:27.319 --> 00:02:30.280
+which will also be the topic of our
+
+00:02:29.200 --> 00:02:32.599
+second assignment
+
+00:02:30.280 --> 00:02:35.680
+here and the way it works is you
+
+00:02:32.599 --> 00:02:38.319
+retrieve relevant passages
+
+00:02:35.680 --> 00:02:40.680
+efficiently ones that kind of entail the
+
+00:02:38.319 --> 00:02:42.480
+answer to a question and then read the
+
+00:02:40.680 --> 00:02:46.080
+passages to answer the
+
+00:02:42.480 --> 00:02:48.599
+query so we have documents like this we
+
+00:02:46.080 --> 00:02:52.360
+have a query based on the query we form
+
+00:02:48.599 --> 00:02:55.360
+retrieval we get a whole bunch of uh
+
+00:02:52.360 --> 00:02:57.560
+passages we do reading and then we get
+
+00:02:55.360 --> 00:02:57.560
+the
+
+00:02:58.280 --> 00:03:04.440
+answer so this is in fact implemented in
+
+00:03:01.720 --> 00:03:07.599
+many or even most uh language modeling
+
+00:03:04.440 --> 00:03:09.840
+providers including open AI so to give
+
+00:03:07.599 --> 00:03:11.480
+an example I asked the question that I
+
+00:03:09.840 --> 00:03:12.879
+just said about Vin Diesel's voice
+
+00:03:11.480 --> 00:03:16.599
+acting and TV
+
+00:03:12.879 --> 00:03:19.760
+series and Chad GPT gave me an answer
+
+00:03:16.599 --> 00:03:22.440
+and you can see that J gpt's answer
+
+00:03:19.760 --> 00:03:24.720
+includes several places with quotes um
+
+00:03:22.440 --> 00:03:28.159
+they the little blue quotes
+
+00:03:24.720 --> 00:03:30.760
+there and if you click on the quote it
+
+00:03:28.159 --> 00:03:33.120
+tells you where the information Source
+
+00:03:30.760 --> 00:03:35.000
+came from and so this one says behind
+
+00:03:33.120 --> 00:03:37.760
+the voice actors been
+
+00:03:35.000 --> 00:03:39.920
+Diesel and behind the voice actors TV
+
+00:03:37.760 --> 00:03:42.959
+shows Big Mouth V
+
+00:03:39.920 --> 00:03:45.640
+diesel now if we look
+
+00:03:42.959 --> 00:03:48.640
+closer into this answer we'll see that
+
+00:03:45.640 --> 00:03:49.959
+it's not perfect even though it is uh
+
+00:03:48.640 --> 00:03:52.519
+performing retrieval augmented
+
+00:03:49.959 --> 00:03:54.840
+Generations so for example I only asked
+
+00:03:52.519 --> 00:03:57.200
+about TV series but it's giving me lots
+
+00:03:54.840 --> 00:03:59.680
+of things about movies where it says
+
+00:03:57.200 --> 00:04:01.319
+Groot in Guardians of the Galaxy volume
+
+00:03:59.680 --> 00:04:04.480
+3 2023
+
+00:04:01.319 --> 00:04:07.200
+movie and in fact uh Vin Diesel was not
+
+00:04:04.480 --> 00:04:10.920
+even voicing a character named gut here
+
+00:04:07.200 --> 00:04:13.480
+so that's definitely an accuracy
+
+00:04:10.920 --> 00:04:15.079
+mistake and separately there's a place
+
+00:04:13.480 --> 00:04:17.639
+where it says additionally though the
+
+00:04:15.079 --> 00:04:19.959
+website for big mouthless Vin Diesel it
+
+00:04:17.639 --> 00:04:22.040
+appears to be a misunderstanding or err
+
+00:04:19.959 --> 00:04:25.360
+as Nick croll is credited as the voice
+
+00:04:22.040 --> 00:04:27.800
+of Vin Diesel in that show so there
+
+00:04:25.360 --> 00:04:30.039
+actually Nick croll was acting as V
+
+00:04:27.800 --> 00:04:32.800
+diesel but that's um kind of a
+
+00:04:30.039 --> 00:04:34.600
+misunderstanding of the reader model but
+
+00:04:32.800 --> 00:04:36.600
+anyway you can get the general idea here
+
+00:04:34.600 --> 00:04:40.199
+you can also see that it's not perfect
+
+00:04:36.600 --> 00:04:42.720
+even for very strong models like GPD
+
+00:04:40.199 --> 00:04:44.800
+4 so now I'd like to go into the actual
+
+00:04:42.720 --> 00:04:46.759
+methodology that we use for this uh we
+
+00:04:44.800 --> 00:04:50.360
+have retrieval
+
+00:04:46.759 --> 00:04:53.160
+methods and for the retrieval methods we
+
+00:04:50.360 --> 00:04:55.160
+have uh quite a few different options
+
+00:04:53.160 --> 00:04:57.960
+I'm going to go through each one of them
+
+00:04:55.160 --> 00:05:00.960
+at a time so sparse retrieval document
+
+00:04:57.960 --> 00:05:04.240
+level dense retrieval token level DSE
+
+00:05:00.960 --> 00:05:08.039
+retrieval cross- encoder reranking and
+
+00:05:04.240 --> 00:05:09.320
+blackbox retrieval so blackbox retrieval
+
+00:05:08.039 --> 00:05:11.280
+I'm not really going to go into it a
+
+00:05:09.320 --> 00:05:16.000
+whole lot basically this is just asking
+
+00:05:11.280 --> 00:05:17.560
+a blackbox search engine to retrieve uh
+
+00:05:16.000 --> 00:05:20.000
+you know the relevant context and
+
+00:05:17.560 --> 00:05:22.560
+getting the top several results
+
+00:05:20.000 --> 00:05:24.039
+nonetheless this is a pretty you know
+
+00:05:22.560 --> 00:05:26.800
+reasonable method to do it if you want
+
+00:05:24.039 --> 00:05:29.080
+to do search over you know lots of data
+
+00:05:26.800 --> 00:05:32.759
+that exists on the internet already and
+
+00:05:29.080 --> 00:05:36.600
+that in is what chat jpt does it looks
+
+00:05:32.759 --> 00:05:39.240
+up on Bing by generating a query to
+
+00:05:36.600 --> 00:05:41.560
+Bing so anyway let's go into the actual
+
+00:05:39.240 --> 00:05:43.840
+methods that you develop and control
+
+00:05:41.560 --> 00:05:46.600
+yourself so the first one is sparse
+
+00:05:43.840 --> 00:05:48.479
+retrieval and the way this works is you
+
+00:05:46.600 --> 00:05:50.440
+express the query and document as a
+
+00:05:48.479 --> 00:05:53.680
+sparse word frequency Vector usually
+
+00:05:50.440 --> 00:05:58.759
+normalized by length and so if I ask uh
+
+00:05:53.680 --> 00:06:01.720
+query what is NLP we get a vector where
+
+00:05:58.759 --> 00:06:04.120
+each row the vector corresponds to a
+
+00:06:01.720 --> 00:06:07.919
+different
+
+00:06:04.120 --> 00:06:12.960
+token and we asked what is
+
+00:06:07.919 --> 00:06:16.360
+NLP and so uh the places for what NLP
+
+00:06:12.960 --> 00:06:18.199
+and is will all have a non-zero value
+
+00:06:16.360 --> 00:06:20.199
+and everything else will have a zero
+
+00:06:18.199 --> 00:06:21.720
+value and we also normalize by the
+
+00:06:20.199 --> 00:06:24.120
+length of vectors so we get something
+
+00:06:21.720 --> 00:06:24.120
+like
+
+00:06:24.840 --> 00:06:28.440
+333333 then we have a whole bunch of
+
+00:06:26.759 --> 00:06:30.720
+documents so the first document says
+
+00:06:28.440 --> 00:06:31.759
+what is life can is life someone really
+
+00:06:30.720 --> 00:06:33.960
+likes
+
+00:06:31.759 --> 00:06:36.000
+candy we also have another one that says
+
+00:06:33.960 --> 00:06:38.360
+NLP as an acronym for natural language
+
+00:06:36.000 --> 00:06:39.479
+processing so this is a pretty good uh
+
+00:06:38.360 --> 00:06:42.479
+you
+
+00:06:39.479 --> 00:06:44.840
+know answer to our
+
+00:06:42.479 --> 00:06:48.039
+question then we also have I like to do
+
+00:06:44.840 --> 00:06:49.360
+good research on NLP which is you know a
+
+00:06:48.039 --> 00:06:51.360
+nice sentiment but not a very good
+
+00:06:49.360 --> 00:06:54.400
+answer to our question I
+
+00:06:51.360 --> 00:06:59.479
+guess so if we look at the vectors here
+
+00:06:54.400 --> 00:07:03.280
+we have uh what and candy and is have uh
+
+00:06:59.479 --> 00:07:07.120
+a fairly high
+
+00:07:03.280 --> 00:07:12.520
+score and we have here NLP and is have a
+
+00:07:07.120 --> 00:07:16.479
+high score and NLP has a a nonzero
+
+00:07:12.520 --> 00:07:18.400
+score So based on this um we find the
+
+00:07:16.479 --> 00:07:20.560
+document similarity with the highest
+
+00:07:18.400 --> 00:07:22.039
+inner product or cosine similarity in
+
+00:07:20.560 --> 00:07:24.360
+the document
+
+00:07:22.039 --> 00:07:27.000
+collection and so if we take the inner
+
+00:07:24.360 --> 00:07:28.759
+product between these vectors we
+
+00:07:27.000 --> 00:07:31.280
+actually see that the first one got the
+
+00:07:28.759 --> 00:07:34.479
+highest score because of its
+
+00:07:31.280 --> 00:07:37.440
+relatively High values for the words
+
+00:07:34.479 --> 00:07:37.440
+what and
+
+00:07:38.160 --> 00:07:43.759
+is
+
+00:07:40.199 --> 00:07:46.720
+so as you can see common words like what
+
+00:07:43.759 --> 00:07:49.000
+and is can get a high score kind of
+
+00:07:46.720 --> 00:07:51.800
+regardless of whether a document is very
+
+00:07:49.000 --> 00:07:53.919
+relevant and so one way we can fix this
+
+00:07:51.800 --> 00:07:55.960
+is through something called term
+
+00:07:53.919 --> 00:07:59.479
+waiting and the way that term waiting
+
+00:07:55.960 --> 00:08:02.680
+works is in addition to having this
+
+00:07:59.479 --> 00:08:04.599
+Vector that
+
+00:08:02.680 --> 00:08:07.680
+calculates
+
+00:08:04.599 --> 00:08:10.680
+the frequency within a particular
+
+00:08:07.680 --> 00:08:13.639
+document we also have an upweighting
+
+00:08:10.680 --> 00:08:15.599
+term that gives higher weight to low
+
+00:08:13.639 --> 00:08:18.199
+frequency words because low frequency
+
+00:08:15.599 --> 00:08:20.280
+words like NLP tend to be more
+
+00:08:18.199 --> 00:08:22.759
+informative about whether the document
+
+00:08:20.280 --> 00:08:25.240
+is relevant than high frequency words
+
+00:08:22.759 --> 00:08:27.080
+like what it is because these high
+
+00:08:25.240 --> 00:08:31.320
+frequency words like what and is Could
+
+00:08:27.080 --> 00:08:34.279
+Happen kind of regardless of whether
+
+00:08:31.320 --> 00:08:36.680
+the you know document is relevant the
+
+00:08:34.279 --> 00:08:41.800
+particular terms the person is asking
+
+00:08:36.680 --> 00:08:44.000
+about so one well used and easy to
+
+00:08:41.800 --> 00:08:46.560
+understand version of this is uh tfidf
+
+00:08:44.000 --> 00:08:48.839
+or term frequency indument
+
+00:08:46.560 --> 00:08:51.200
+frequency so the way we Define term
+
+00:08:48.839 --> 00:08:52.959
+frequency is exactly what I talked about
+
+00:08:51.200 --> 00:08:56.959
+before so it's basically the frequency
+
+00:08:52.959 --> 00:08:59.839
+of the term uh T in the document d
+
+00:08:56.959 --> 00:09:01.640
+normalized by the total term frequency
+
+00:08:59.839 --> 00:09:03.680
+within the document so that that's what
+
+00:09:01.640 --> 00:09:06.800
+I already showed in the previous
+
+00:09:03.680 --> 00:09:09.360
+slide and then indument frequency is a
+
+00:09:06.800 --> 00:09:13.760
+little bit more involved but basically
+
+00:09:09.360 --> 00:09:15.760
+the way this works is we have log of the
+
+00:09:13.760 --> 00:09:18.160
+total number of documents in the
+
+00:09:15.760 --> 00:09:24.040
+collection divided
+
+00:09:18.160 --> 00:09:26.760
+by the total number of uh times this
+
+00:09:24.040 --> 00:09:30.279
+term appeared in any particular
+
+00:09:26.760 --> 00:09:33.360
+document and so if a term appears many
+
+00:09:30.279 --> 00:09:36.120
+times in any particular document it will
+
+00:09:33.360 --> 00:09:39.240
+have a low IDF score uh one that's close
+
+00:09:36.120 --> 00:09:41.519
+to zero but if it rarely appears it will
+
+00:09:39.240 --> 00:09:44.120
+have a high IDF score so basically this
+
+00:09:41.519 --> 00:09:45.040
+is upweighting our frequent terms and
+
+00:09:44.120 --> 00:09:47.560
+then for
+
+00:09:45.040 --> 00:09:51.320
+tfidf uh we basically multiply these two
+
+00:09:47.560 --> 00:09:53.120
+terms together and we upweight the low
+
+00:09:51.320 --> 00:09:55.640
+frequency
+
+00:09:53.120 --> 00:10:00.519
+words there's another version of this
+
+00:09:55.640 --> 00:10:03.640
+called bm25 that is uh widely used used
+
+00:10:00.519 --> 00:10:05.800
+um this is more involved so I'm not
+
+00:10:03.640 --> 00:10:08.120
+going to go into all of the details but
+
+00:10:05.800 --> 00:10:12.399
+basically if you remember back to the
+
+00:10:08.120 --> 00:10:13.720
+lecture on count-based language models
+
+00:10:12.399 --> 00:10:14.880
+there were a bunch of smoothing
+
+00:10:13.720 --> 00:10:18.839
+techniques for these count-based
+
+00:10:14.880 --> 00:10:21.839
+language models and this uses uh kind of
+
+00:10:18.839 --> 00:10:25.839
+a m multiplicative additive smoothing
+
+00:10:21.839 --> 00:10:27.160
+term to upway things instead of using
+
+00:10:25.839 --> 00:10:30.200
+the term
+
+00:10:27.160 --> 00:10:33.399
+frequency and uh the actual formula is
+
+00:10:30.200 --> 00:10:37.240
+here K and B are kind of
+
+00:10:33.399 --> 00:10:39.360
+hyperparameters and um average DL is
+
+00:10:37.240 --> 00:10:40.639
+average document length the details of
+
+00:10:39.360 --> 00:10:42.120
+this are not really important but
+
+00:10:40.639 --> 00:10:43.800
+basically what you should know is that
+
+00:10:42.120 --> 00:10:45.639
+this is doing some smoothing on the term
+
+00:10:43.800 --> 00:10:48.240
+frequencies and you can look in more
+
+00:10:45.639 --> 00:10:48.240
+detail if you're
+
+00:10:49.160 --> 00:10:54.920
+interested so now that we have this sort
+
+00:10:52.880 --> 00:10:57.959
+of term
+
+00:10:54.920 --> 00:11:00.320
+based uh sparse Vector we would like to
+
+00:10:57.959 --> 00:11:03.320
+use this to look up relevant documents
+
+00:11:00.320 --> 00:11:06.000
+in a collection very quickly because you
+
+00:11:03.320 --> 00:11:08.000
+know we might have a collection that's
+
+00:11:06.000 --> 00:11:09.720
+extremely large like as large as the
+
+00:11:08.000 --> 00:11:12.320
+entire internet like what Google is
+
+00:11:09.720 --> 00:11:14.160
+doing when it searches and so in order
+
+00:11:12.320 --> 00:11:16.240
+to solve this we need a data structure
+
+00:11:14.160 --> 00:11:17.279
+that allows for efficient sparse lookup
+
+00:11:16.240 --> 00:11:19.480
+of
+
+00:11:17.279 --> 00:11:23.720
+vectors and so we have all of these
+
+00:11:19.480 --> 00:11:27.279
+sparse vectors like this
+
+00:11:23.720 --> 00:11:31.240
+and we uh basically turn this into an
+
+00:11:27.279 --> 00:11:34.720
+index where we have something like a you
+
+00:11:31.240 --> 00:11:37.920
+know python style dictionary or map that
+
+00:11:34.720 --> 00:11:41.079
+has it's the key each uh word we would
+
+00:11:37.920 --> 00:11:45.000
+look like to look up and is the vector
+
+00:11:41.079 --> 00:11:48.480
+the corresponding um index of that
+
+00:11:45.000 --> 00:11:50.480
+document so for example what in our case
+
+00:11:48.480 --> 00:11:54.200
+here only appears in document one so it
+
+00:11:50.480 --> 00:11:56.279
+would point to document one candy uh
+
+00:11:54.200 --> 00:11:58.560
+also appears in document one NLP appears
+
+00:11:56.279 --> 00:11:59.839
+in two and three and so you can create
+
+00:11:58.560 --> 00:12:02.760
+this index IND like this and this is
+
+00:11:59.839 --> 00:12:02.760
+called an inverted
+
+00:12:03.079 --> 00:12:08.760
+Index this is an important application
+
+00:12:06.000 --> 00:12:11.600
+of course so there's lots of software
+
+00:12:08.760 --> 00:12:14.920
+the most kind of pical software for this
+
+00:12:11.600 --> 00:12:18.760
+is Apache Lucine so if you want to build
+
+00:12:14.920 --> 00:12:21.639
+a big index uh to look up vectors using
+
+00:12:18.760 --> 00:12:24.160
+this sparse index like this you can uh
+
+00:12:21.639 --> 00:12:24.160
+take a look at
+
+00:12:26.160 --> 00:12:30.880
+Lucy so the next thing I'd like to talk
+
+00:12:28.399 --> 00:12:33.199
+about is dense retrieval and the way
+
+00:12:30.880 --> 00:12:36.000
+dense retrieval works is you encode the
+
+00:12:33.199 --> 00:12:37.240
+document in query into a dense factor
+
+00:12:36.000 --> 00:12:40.240
+and find the nearest
+
+00:12:37.240 --> 00:12:42.160
+neighbor in order to do this encoding
+
+00:12:40.240 --> 00:12:44.639
+you can use a number of things you can
+
+00:12:42.160 --> 00:12:47.440
+use out of the box embeddings or you can
+
+00:12:44.639 --> 00:12:49.959
+use learned embeddings specifically
+
+00:12:47.440 --> 00:12:53.519
+created for the purpose of
+
+00:12:49.959 --> 00:12:56.240
+retrieving and so what we do is we take
+
+00:12:53.519 --> 00:12:57.920
+all of these uh documents here we
+
+00:12:56.240 --> 00:12:59.920
+convert them into embeddings using
+
+00:12:57.920 --> 00:13:04.040
+whatever embedding method that we want
+
+00:12:59.920 --> 00:13:05.920
+to use we then have a query and we take
+
+00:13:04.040 --> 00:13:07.720
+that query and we match it and find the
+
+00:13:05.920 --> 00:13:10.040
+nearest neighbor
+
+00:13:07.720 --> 00:13:13.120
+here so if you're just using out of the
+
+00:13:10.040 --> 00:13:14.839
+box embeddings you don't need to um you
+
+00:13:13.120 --> 00:13:15.880
+know do anything special for retrieval
+
+00:13:14.839 --> 00:13:18.440
+you can just take your favorite
+
+00:13:15.880 --> 00:13:22.800
+embeddings like the sentence BT
+
+00:13:18.440 --> 00:13:25.639
+embeddings or the open AI uh Adda
+
+00:13:22.800 --> 00:13:27.240
+embeddings or something like this but
+
+00:13:25.639 --> 00:13:29.519
+actually the type of embeddings you need
+
+00:13:27.240 --> 00:13:32.040
+for retrieval are kind of
+
+00:13:29.519 --> 00:13:33.519
+very special and because of that it's
+
+00:13:32.040 --> 00:13:36.160
+important
+
+00:13:33.519 --> 00:13:38.600
+to if you're very serious about doing a
+
+00:13:36.160 --> 00:13:39.800
+good job of retal it's important to use
+
+00:13:38.600 --> 00:13:41.360
+embeddings that were specifically
+
+00:13:39.800 --> 00:13:45.040
+tailored for
+
+00:13:41.360 --> 00:13:47.680
+retrieval and the reason why it is
+
+00:13:45.040 --> 00:13:50.079
+important to do this is severalfold but
+
+00:13:47.680 --> 00:13:53.800
+the most intuitive way to think about it
+
+00:13:50.079 --> 00:13:57.600
+is if we think about uh the things that
+
+00:13:53.800 --> 00:13:59.440
+tfidf does tfidf is giving a very high
+
+00:13:57.600 --> 00:14:03.000
+weight to
+
+00:13:59.440 --> 00:14:04.959
+contentful words and rare words and
+
+00:14:03.000 --> 00:14:06.639
+we're not guaranteed that any random
+
+00:14:04.959 --> 00:14:10.600
+embedding that we get is going to do
+
+00:14:06.639 --> 00:14:13.800
+that so for example if we just take the
+
+00:14:10.600 --> 00:14:16.160
+average word embeddings of every word in
+
+00:14:13.800 --> 00:14:20.160
+a sequence it's going to give the same
+
+00:14:16.160 --> 00:14:22.320
+weight to all of the words um in the
+
+00:14:20.160 --> 00:14:24.680
+output and in fact common words tend to
+
+00:14:22.320 --> 00:14:27.959
+have slightly higher Norms than
+
+00:14:24.680 --> 00:14:29.639
+infrequent words and so that would
+
+00:14:27.959 --> 00:14:31.880
+actually upli common wordss which is
+
+00:14:29.639 --> 00:14:34.639
+kind of exactly the opposite thing we
+
+00:14:31.880 --> 00:14:36.480
+want so how do we learn retrieval
+
+00:14:34.639 --> 00:14:39.160
+oriented
+
+00:14:36.480 --> 00:14:40.920
+embeddings the normal way we do this is
+
+00:14:39.160 --> 00:14:43.399
+we select positive and negative
+
+00:14:40.920 --> 00:14:46.839
+documents and then train using a
+
+00:14:43.399 --> 00:14:50.240
+contrastive loss and so an example of
+
+00:14:46.839 --> 00:14:52.519
+this is we have a query and then we have
+
+00:14:50.240 --> 00:14:55.519
+negative documents for that query and we
+
+00:14:52.519 --> 00:14:58.199
+have positive documents for that query
+
+00:14:55.519 --> 00:15:00.079
+and uh we form formulate a hinge loss or
+
+00:14:58.199 --> 00:15:04.000
+maybe some sort of probabilistic loss
+
+00:15:00.079 --> 00:15:06.560
+similar to the Hench loss and uh do fine
+
+00:15:04.000 --> 00:15:06.560
+tuning of the
+
+00:15:07.160 --> 00:15:13.440
+embeddings so if
+
+00:15:09.399 --> 00:15:16.320
+you have gold standard positive
+
+00:15:13.440 --> 00:15:18.800
+documents then this is relatively easy
+
+00:15:16.320 --> 00:15:21.040
+to train uh because you just need the
+
+00:15:18.800 --> 00:15:23.800
+positive documents and then you can get
+
+00:15:21.040 --> 00:15:25.959
+Negative documents in a number of ways
+
+00:15:23.800 --> 00:15:29.279
+one common way of getting negative
+
+00:15:25.959 --> 00:15:32.279
+documents is you just form a batch of
+
+00:15:29.279 --> 00:15:34.560
+data and given that batch of data you
+
+00:15:32.279 --> 00:15:37.480
+take all of the other documents in the
+
+00:15:34.560 --> 00:15:39.480
+batch um all of the documents in the
+
+00:15:37.480 --> 00:15:42.839
+batch that are positive for some other
+
+00:15:39.480 --> 00:15:46.399
+query and you use those as negative
+
+00:15:42.839 --> 00:15:49.000
+documents so you sample 32 query
+
+00:15:46.399 --> 00:15:50.759
+document pairs you use the aligned ones
+
+00:15:49.000 --> 00:15:53.759
+as positive documents and then use the
+
+00:15:50.759 --> 00:15:57.440
+31 other ones as negative documents and
+
+00:15:53.759 --> 00:16:00.279
+this is both effective and efficient
+
+00:15:57.440 --> 00:16:02.000
+because you can kind of learned from the
+
+00:16:00.279 --> 00:16:05.079
+query document pairs all at the same
+
+00:16:02.000 --> 00:16:05.079
+time in an efficient
+
+00:16:05.680 --> 00:16:13.680
+implementation however this is not
+
+00:16:09.160 --> 00:16:16.279
+enough in many cases because that will
+
+00:16:13.680 --> 00:16:19.040
+end up having lots of very kind of
+
+00:16:16.279 --> 00:16:20.440
+obviously wrong documents because you
+
+00:16:19.040 --> 00:16:23.120
+know
+
+00:16:20.440 --> 00:16:25.360
+they're documents that are relevant for
+
+00:16:23.120 --> 00:16:27.880
+a completely different query and it's
+
+00:16:25.360 --> 00:16:29.880
+kind of easy to distinguish uh between
+
+00:16:27.880 --> 00:16:32.319
+those you can just at superficial word
+
+00:16:29.880 --> 00:16:34.519
+overlap so another common thing to do
+
+00:16:32.319 --> 00:16:35.759
+when you're training these models is to
+
+00:16:34.519 --> 00:16:38.160
+get hard
+
+00:16:35.759 --> 00:16:40.680
+negatives so hard negatives are
+
+00:16:38.160 --> 00:16:44.360
+basically negative examples that look
+
+00:16:40.680 --> 00:16:49.399
+plausible but are actually wrong and
+
+00:16:44.360 --> 00:16:53.199
+so here uh this famous method called DPR
+
+00:16:49.399 --> 00:16:55.880
+is it basically learns the uh encoders
+
+00:16:53.199 --> 00:16:57.759
+based on both inbatch negatives like I
+
+00:16:55.880 --> 00:17:00.160
+mentioned before and hard negatives that
+
+00:16:57.759 --> 00:17:01.360
+were created by looking up documents
+
+00:17:00.160 --> 00:17:03.839
+with
+
+00:17:01.360 --> 00:17:06.039
+bm25 and so the ones that were looked up
+
+00:17:03.839 --> 00:17:07.640
+by bm25 you know kind of look very
+
+00:17:06.039 --> 00:17:10.039
+similar superficially but they might
+
+00:17:07.640 --> 00:17:12.400
+have you know subtle errors in them for
+
+00:17:10.039 --> 00:17:12.400
+why they're
+
+00:17:12.799 --> 00:17:17.160
+inappropriate there's also methods to
+
+00:17:15.679 --> 00:17:20.000
+learn these
+
+00:17:17.160 --> 00:17:23.199
+retrievers based on kind of not
+
+00:17:20.000 --> 00:17:26.199
+supervised data so one major bottleneck
+
+00:17:23.199 --> 00:17:29.000
+if you're taking the positive documents
+
+00:17:26.199 --> 00:17:30.440
+from Human annotations of whether
+
+00:17:29.000 --> 00:17:33.440
+something is correct or not or human
+
+00:17:30.440 --> 00:17:37.880
+clickthrough logs or other things like
+
+00:17:33.440 --> 00:17:40.640
+this is that you need that data in order
+
+00:17:37.880 --> 00:17:44.440
+to start training a bottle so uh
+
+00:17:40.640 --> 00:17:47.880
+contriver is another method that uses
+
+00:17:44.440 --> 00:17:51.520
+two random spans within a document is a
+
+00:17:47.880 --> 00:17:54.440
+positive pair and random spans from
+
+00:17:51.520 --> 00:17:56.559
+across documents is negative Pairs and
+
+00:17:54.440 --> 00:17:58.960
+so this can be used for you know very
+
+00:17:56.559 --> 00:18:00.039
+very large scale initial pre-training of
+
+00:17:58.960 --> 00:18:02.280
+the
+
+00:18:00.039 --> 00:18:04.520
+models and then after you've done that
+
+00:18:02.280 --> 00:18:06.840
+large scale initial pre-training you can
+
+00:18:04.520 --> 00:18:10.799
+then go in and fine-tune it on you know
+
+00:18:06.840 --> 00:18:10.799
+actually annotate the data to improve it
+
+00:18:12.120 --> 00:18:18.799
+further Okay so we've talked about
+
+00:18:15.159 --> 00:18:21.559
+training uh these dense product uh
+
+00:18:18.799 --> 00:18:24.559
+models these uh models that look at
+
+00:18:21.559 --> 00:18:27.720
+dense embedding overlap for nearest
+
+00:18:24.559 --> 00:18:28.919
+neighbors but the problem is in order to
+
+00:18:27.720 --> 00:18:30.919
+calculate this you would need to
+
+00:18:28.919 --> 00:18:35.159
+calculate it over a very very large
+
+00:18:30.919 --> 00:18:37.960
+document base and just taking a product
+
+00:18:35.159 --> 00:18:40.480
+between the query and all of the other
+
+00:18:37.960 --> 00:18:42.400
+documents in the document base is
+
+00:18:40.480 --> 00:18:46.080
+extremely
+
+00:18:42.400 --> 00:18:48.080
+costly and so in order to fix this there
+
+00:18:46.080 --> 00:18:49.080
+are methods for approximate nearest
+
+00:18:48.080 --> 00:18:52.280
+neighbor
+
+00:18:49.080 --> 00:18:54.200
+search and these are methods that allow
+
+00:18:52.280 --> 00:18:57.360
+you to retrieve embeddings that have the
+
+00:18:54.200 --> 00:19:00.280
+maximum inner product between them in
+
+00:18:57.360 --> 00:19:02.520
+sublinear time and because you're doing
+
+00:19:00.280 --> 00:19:03.960
+the maximum inner product this is also
+
+00:19:02.520 --> 00:19:06.600
+often called maximum inner product
+
+00:19:03.960 --> 00:19:06.600
+search or
+
+00:19:06.679 --> 00:19:12.360
+myips so I'm going to introduce on a
+
+00:19:09.440 --> 00:19:15.360
+very high level two common methods to do
+
+00:19:12.360 --> 00:19:19.320
+this the first one is locality sensitive
+
+00:19:15.360 --> 00:19:22.440
+hashen um or this can also be called
+
+00:19:19.320 --> 00:19:24.799
+kind of inverted index as well and what
+
+00:19:22.440 --> 00:19:26.840
+you do is you make partitions in
+
+00:19:24.799 --> 00:19:29.320
+continuous space and then you use it
+
+00:19:26.840 --> 00:19:31.240
+like an inverted index
+
+00:19:29.320 --> 00:19:33.679
+so let's say we have a whole bunch of
+
+00:19:31.240 --> 00:19:34.919
+embeddings uh I demonstrated two
+
+00:19:33.679 --> 00:19:36.640
+dimensional embeddings here but in
+
+00:19:34.919 --> 00:19:38.440
+reality this would be you know as large
+
+00:19:36.640 --> 00:19:41.159
+as your word
+
+00:19:38.440 --> 00:19:42.880
+embedding your query and document
+
+00:19:41.159 --> 00:19:47.120
+embedding space so this would be you
+
+00:19:42.880 --> 00:19:49.760
+know 512 or 1024 or something like that
+
+00:19:47.120 --> 00:19:53.480
+and what you do is you define a whole
+
+00:19:49.760 --> 00:19:56.720
+bunch of planes that separate these
+
+00:19:53.480 --> 00:19:59.320
+points into two spaces so if this is our
+
+00:19:56.720 --> 00:20:02.520
+first plane all the points above the
+
+00:19:59.320 --> 00:20:04.280
+plane will get a one for this partition
+
+00:20:02.520 --> 00:20:06.799
+and all the points below the plane will
+
+00:20:04.280 --> 00:20:08.840
+get a zero for this partition and we do
+
+00:20:06.799 --> 00:20:12.400
+it similarly we we create a whole bunch
+
+00:20:08.840 --> 00:20:15.840
+of them and then based on this we can
+
+00:20:12.400 --> 00:20:18.440
+now assign sparse vectors depending on
+
+00:20:15.840 --> 00:20:21.520
+each of these planes so we have uh for
+
+00:20:18.440 --> 00:20:24.000
+example the top one uh one0 0 because
+
+00:20:21.520 --> 00:20:26.400
+it's on the right side of the blue plane
+
+00:20:24.000 --> 00:20:28.760
+and the um wrong side of the red and the
+
+00:20:26.400 --> 00:20:30.679
+green planes and then for the top right
+
+00:20:28.760 --> 00:20:32.799
+we have one1 because it's on the right
+
+00:20:30.679 --> 00:20:37.159
+side of the blueing the green planes and
+
+00:20:32.799 --> 00:20:39.440
+the wrong side of the red plane and So
+
+00:20:37.159 --> 00:20:41.000
+based on this now we have a sparse
+
+00:20:39.440 --> 00:20:42.600
+vector and we already know what to do
+
+00:20:41.000 --> 00:20:44.640
+with a sparse Vector right we look it up
+
+00:20:42.600 --> 00:20:49.039
+in an inverted index just like we did
+
+00:20:44.640 --> 00:20:51.520
+for a sparse um you know sparse lookup
+
+00:20:49.039 --> 00:20:54.520
+table so that's one
+
+00:20:51.520 --> 00:20:57.799
+method another method uses a graph-based
+
+00:20:54.520 --> 00:21:01.320
+search and the basic idea behind this is
+
+00:20:57.799 --> 00:21:02.480
+that we create hubs uh and these hubs
+
+00:21:01.320 --> 00:21:05.200
+are kind
+
+00:21:02.480 --> 00:21:07.960
+of a small number of points that are
+
+00:21:05.200 --> 00:21:09.440
+close to other points in the space and
+
+00:21:07.960 --> 00:21:10.880
+so we create some hubs and then we
+
+00:21:09.440 --> 00:21:12.200
+search from there so if we have a
+
+00:21:10.880 --> 00:21:16.880
+similar
+
+00:21:12.200 --> 00:21:19.159
+looking uh set of points in the space we
+
+00:21:16.880 --> 00:21:21.520
+find these hubs which are something like
+
+00:21:19.159 --> 00:21:24.880
+cluster centroids and then based on the
+
+00:21:21.520 --> 00:21:28.559
+cluster centroids we then rule down or
+
+00:21:24.880 --> 00:21:31.200
+we greatly reduce the number of
+
+00:21:28.559 --> 00:21:33.400
+points that we need to be looking at and
+
+00:21:31.200 --> 00:21:36.960
+then we search through only those points
+
+00:21:33.400 --> 00:21:38.600
+in a more kind of extensive Manner and
+
+00:21:36.960 --> 00:21:41.840
+you can even turn this into a tree where
+
+00:21:38.600 --> 00:21:43.760
+you have hubs and then you have uh kind
+
+00:21:41.840 --> 00:21:46.600
+of mini hubs and then you have all the
+
+00:21:43.760 --> 00:21:50.200
+points so this allows you to do a kind
+
+00:21:46.600 --> 00:21:50.200
+of tree based or graph based
+
+00:21:50.600 --> 00:21:55.840
+search so obviously unless you're really
+
+00:21:54.159 --> 00:21:57.039
+excited about these algorithms this is
+
+00:21:55.840 --> 00:22:00.080
+something that you probably don't want
+
+00:21:57.039 --> 00:22:01.440
+to be implementing yourself um and the
+
+00:22:00.080 --> 00:22:03.000
+good news is there's lots of very good
+
+00:22:01.440 --> 00:22:04.480
+libraries that help you do this in fact
+
+00:22:03.000 --> 00:22:08.799
+there are so many libraries it's hard to
+
+00:22:04.480 --> 00:22:11.960
+manage them but some libraries that
+
+00:22:08.799 --> 00:22:13.799
+people very commonly use I I think face
+
+00:22:11.960 --> 00:22:17.320
+uh FIS
+
+00:22:13.799 --> 00:22:20.200
+SS is a widely used one created by uh
+
+00:22:17.320 --> 00:22:23.760
+fair and meta and chroma DB is a
+
+00:22:20.200 --> 00:22:27.720
+separate one uh that is kind of an AI
+
+00:22:23.760 --> 00:22:30.720
+native uh embedding search database so
+
+00:22:27.720 --> 00:22:30.720
+both those are good
+
+00:22:32.960 --> 00:22:41.120
+options even with intelligent training
+
+00:22:37.880 --> 00:22:42.640
+of dense embeddings however there still
+
+00:22:41.120 --> 00:22:45.600
+are
+
+00:22:42.640 --> 00:22:48.240
+problems and the biggest
+
+00:22:45.600 --> 00:22:51.720
+problem that you face when you're
+
+00:22:48.240 --> 00:22:54.000
+looking at something like uh cross
+
+00:22:51.720 --> 00:22:56.880
+encoders um that sorry when you're
+
+00:22:54.000 --> 00:23:00.240
+looking at dense embeddings is that in
+
+00:22:56.880 --> 00:23:02.159
+order to form a good dense embedding you
+
+00:23:00.240 --> 00:23:03.840
+need to kind of know in advance what
+
+00:23:02.159 --> 00:23:05.799
+you're looking for right because you're
+
+00:23:03.840 --> 00:23:09.120
+taking a long document you're condensing
+
+00:23:05.799 --> 00:23:10.679
+it down into a single embedding and or a
+
+00:23:09.120 --> 00:23:13.320
+long passage and you're condensing it
+
+00:23:10.679 --> 00:23:16.200
+down to a single embedding and so if
+
+00:23:13.320 --> 00:23:19.520
+that during that condensation process
+
+00:23:16.200 --> 00:23:21.240
+actually there's other information that
+
+00:23:19.520 --> 00:23:23.159
+is relevant to a query but you have to
+
+00:23:21.240 --> 00:23:27.600
+throw out because of the limited
+
+00:23:23.159 --> 00:23:30.600
+embedding capacity this causes you to
+
+00:23:27.600 --> 00:23:32.320
+you know essentially fail at um doing
+
+00:23:30.600 --> 00:23:34.840
+retrieval
+
+00:23:32.320 --> 00:23:38.159
+appropriately so there's a couple
+
+00:23:34.840 --> 00:23:40.880
+methods that can be used to fix this so
+
+00:23:38.159 --> 00:23:42.279
+the first method is in contrast to the
+
+00:23:40.880 --> 00:23:44.159
+buy encoder which is what I've been
+
+00:23:42.279 --> 00:23:47.000
+talking out about at this point where
+
+00:23:44.159 --> 00:23:48.520
+you kind of do full encoding of queries
+
+00:23:47.000 --> 00:23:52.120
+full encoding of documents and then do
+
+00:23:48.520 --> 00:23:53.840
+inner product search for a score uh you
+
+00:23:52.120 --> 00:23:56.760
+can use a cross encoder and the way the
+
+00:23:53.840 --> 00:23:58.559
+cross- encoder works is you append the
+
+00:23:56.760 --> 00:24:00.799
+query and document and then you run them
+
+00:23:58.559 --> 00:24:03.400
+through a model like a Transformer model
+
+00:24:00.799 --> 00:24:07.840
+and you calculate the output
+
+00:24:03.400 --> 00:24:09.880
+score so the problem with this um so
+
+00:24:07.840 --> 00:24:12.480
+this this is great uh because it gives
+
+00:24:09.880 --> 00:24:15.799
+you maximum flexibility um Transformer
+
+00:24:12.480 --> 00:24:18.799
+models are powerful you can uh assess
+
+00:24:15.799 --> 00:24:20.520
+relevance very well the problem with
+
+00:24:18.799 --> 00:24:22.200
+this is this precludes approximate
+
+00:24:20.520 --> 00:24:23.720
+nearest neighbor lookup because now
+
+00:24:22.200 --> 00:24:25.799
+you're running through you know many
+
+00:24:23.720 --> 00:24:28.880
+many nonlinearities
+
+00:24:25.799 --> 00:24:32.760
+here so this is can only be used for
+
+00:24:28.880 --> 00:24:34.360
+reranking documents um or if even if
+
+00:24:32.760 --> 00:24:36.880
+you're doing retrieval doing retrieval
+
+00:24:34.360 --> 00:24:39.679
+over a very very small number of
+
+00:24:36.880 --> 00:24:41.960
+documents but if you really want maximal
+
+00:24:39.679 --> 00:24:44.080
+accuracy I definitely would recommend uh
+
+00:24:41.960 --> 00:24:45.720
+doing something like this because it can
+
+00:24:44.080 --> 00:24:47.960
+allow you to do kind of a second pass
+
+00:24:45.720 --> 00:24:49.360
+filtering over the most relevant looking
+
+00:24:47.960 --> 00:24:52.399
+documents to identify the ones you
+
+00:24:49.360 --> 00:24:52.399
+really want to add to your
+
+00:24:54.240 --> 00:24:58.240
+context so then there are also
+
+00:24:56.760 --> 00:25:01.360
+approaches that are kind kind of in the
+
+00:24:58.240 --> 00:25:02.159
+middle of these two uh the most famous
+
+00:25:01.360 --> 00:25:05.880
+one is
+
+00:25:02.159 --> 00:25:08.320
+Kar and the I called this token level
+
+00:25:05.880 --> 00:25:10.840
+dense retrieval it's also called uh late
+
+00:25:08.320 --> 00:25:12.720
+interaction in the coold bear paper but
+
+00:25:10.840 --> 00:25:14.919
+the way it works is you use
+
+00:25:12.720 --> 00:25:18.440
+contextualized representations of all
+
+00:25:14.919 --> 00:25:19.440
+query and document tokens to compute a
+
+00:25:18.440 --> 00:25:23.559
+retrieval
+
+00:25:19.440 --> 00:25:26.919
+score and so you do offline indexing of
+
+00:25:23.559 --> 00:25:29.159
+every token in the document and then
+
+00:25:26.919 --> 00:25:31.399
+based on this offline X indexing of
+
+00:25:29.159 --> 00:25:35.320
+every token in the document you then
+
+00:25:31.399 --> 00:25:38.760
+have a query encoder and you do matching
+
+00:25:35.320 --> 00:25:41.799
+between each token in the query and the
+
+00:25:38.760 --> 00:25:43.399
+highest scoring tokens in each
+
+00:25:41.799 --> 00:25:46.320
+document
+
+00:25:43.399 --> 00:25:48.399
+and the reason why this is good is it
+
+00:25:46.320 --> 00:25:49.600
+still allows you to encode all of the
+
+00:25:48.399 --> 00:25:52.120
+tokens in the
+
+00:25:49.600 --> 00:25:55.440
+document and but each of these
+
+00:25:52.120 --> 00:25:59.679
+similarity searches is still just
+
+00:25:55.440 --> 00:26:03.559
+a kind of maximum product search and
+
+00:25:59.679 --> 00:26:06.279
+because of this this allows you to do
+
+00:26:03.559 --> 00:26:07.960
+each of these searches efficiently and
+
+00:26:06.279 --> 00:26:09.840
+doesn't preclude you from running it
+
+00:26:07.960 --> 00:26:12.919
+over an entire
+
+00:26:09.840 --> 00:26:16.399
+database the downside to this method uh
+
+00:26:12.919 --> 00:26:19.120
+may already be obvious but in the
+
+00:26:16.399 --> 00:26:22.200
+traditional bu encoder we have a single
+
+00:26:19.120 --> 00:26:26.880
+Vector for each document but here we
+
+00:26:22.200 --> 00:26:29.320
+have one vector for um each token in the
+
+00:26:26.880 --> 00:26:31.880
+document so BAS basically your vector
+
+00:26:29.320 --> 00:26:34.399
+database gets n times larger where n is
+
+00:26:31.880 --> 00:26:36.679
+the number of tokens in the document and
+
+00:26:34.399 --> 00:26:38.080
+there are certain methods to make this
+
+00:26:36.679 --> 00:26:41.559
+better like you can compress each
+
+00:26:38.080 --> 00:26:42.960
+document to a smaller number of n uh but
+
+00:26:41.559 --> 00:26:45.880
+still this is definitely going to be
+
+00:26:42.960 --> 00:26:48.399
+more costly than looking up each uh
+
+00:26:45.880 --> 00:26:50.360
+token so this is definitely something to
+
+00:26:48.399 --> 00:26:53.520
+consider if you want to get you know
+
+00:26:50.360 --> 00:26:55.159
+very good scores and Co bear is a good
+
+00:26:53.520 --> 00:26:59.600
+implementation of that to start with if
+
+00:26:55.159 --> 00:26:59.600
+you're interested in trying it out
+
+00:27:00.480 --> 00:27:07.000
+so this is a final thing this is uh
+
+00:27:03.080 --> 00:27:08.679
+something that is a little bit uh
+
+00:27:07.000 --> 00:27:10.080
+different than all the other things I I
+
+00:27:08.679 --> 00:27:12.399
+talked about before but I've used it
+
+00:27:10.080 --> 00:27:15.840
+myself and it actually can be pretty
+
+00:27:12.399 --> 00:27:18.799
+effective um it was also made at CMU so
+
+00:27:15.840 --> 00:27:24.399
+by Lal so I would like to promote our
+
+00:27:18.799 --> 00:27:26.880
+CMU work of course but um the HP idea
+
+00:27:24.399 --> 00:27:28.080
+between behind a hypothetical document
+
+00:27:26.880 --> 00:27:30.320
+embedding
+
+00:27:28.080 --> 00:27:33.440
+is that it's actually somewhat difficult
+
+00:27:30.320 --> 00:27:36.200
+to match a query and a document right
+
+00:27:33.440 --> 00:27:38.919
+because a query is a very short possibly
+
+00:27:36.200 --> 00:27:42.240
+ungrammatical output that's asking a
+
+00:27:38.919 --> 00:27:44.799
+question and then a document is a very
+
+00:27:42.240 --> 00:27:49.440
+long output that's written in a
+
+00:27:44.799 --> 00:27:50.799
+different proos style and you you know
+
+00:27:49.440 --> 00:27:53.159
+it might have lots of irrelevant
+
+00:27:50.799 --> 00:27:54.519
+information or or boiler plate or fluff
+
+00:27:53.159 --> 00:27:57.640
+or something like
+
+00:27:54.519 --> 00:28:00.640
+that so the idea behind a hypothetical
+
+00:27:57.640 --> 00:28:03.120
+document embedding is that it's e easier
+
+00:28:00.640 --> 00:28:05.279
+to match a document in a document than
+
+00:28:03.120 --> 00:28:08.159
+it is to match a query in a
+
+00:28:05.279 --> 00:28:10.159
+document but the input to our model is a
+
+00:28:08.159 --> 00:28:14.360
+query right so what do we
+
+00:28:10.159 --> 00:28:17.919
+do and so essentially what we do is we
+
+00:28:14.360 --> 00:28:20.399
+then take a large language model we feed
+
+00:28:17.919 --> 00:28:23.320
+it in a query in a prompt and say
+
+00:28:20.399 --> 00:28:25.399
+generate a document that looks like it
+
+00:28:23.320 --> 00:28:30.080
+should be the answer to this
+
+00:28:25.399 --> 00:28:32.120
+query and so so then the llm goes in and
+
+00:28:30.080 --> 00:28:34.440
+it generates a document and hopefully
+
+00:28:32.120 --> 00:28:38.440
+this document looks more similar to the
+
+00:28:34.440 --> 00:28:41.440
+documents you want to retrieve than the
+
+00:28:38.440 --> 00:28:44.039
+um than the original query does and I've
+
+00:28:41.440 --> 00:28:47.240
+actually found this to be relatively
+
+00:28:44.039 --> 00:28:51.880
+effective at improving accuracy
+
+00:28:47.240 --> 00:28:53.200
+on kind of difficult uh tasks especially
+
+00:28:51.880 --> 00:28:55.840
+ones that are out of domain from the
+
+00:28:53.200 --> 00:28:58.000
+trend models that I'm
+
+00:28:55.840 --> 00:29:01.880
+using so I've gone through a whole bunch
+
+00:28:58.000 --> 00:29:04.039
+of methods and I would like to finish up
+
+00:29:01.880 --> 00:29:05.679
+this section by giving some insight
+
+00:29:04.039 --> 00:29:11.399
+about which one you should be
+
+00:29:05.679 --> 00:29:14.559
+using so my impression right now is
+
+00:29:11.399 --> 00:29:17.760
+that a good basine to start out with is
+
+00:29:14.559 --> 00:29:20.679
+something like bm25 it's very easy to
+
+00:29:17.760 --> 00:29:23.080
+start out and compared to embedding
+
+00:29:20.679 --> 00:29:26.120
+based models it tends to be relatively
+
+00:29:23.080 --> 00:29:28.279
+robust to new domains so if you have a
+
+00:29:26.120 --> 00:29:30.559
+new domain you're more less guaranteed
+
+00:29:28.279 --> 00:29:32.240
+that bm25 will give you some performance
+
+00:29:30.559 --> 00:29:35.320
+whereas embeddings may be really good
+
+00:29:32.240 --> 00:29:38.399
+but they may be really bad uh depending
+
+00:29:35.320 --> 00:29:40.880
+on how out of domain that is compared to
+
+00:29:38.399 --> 00:29:42.799
+your underlying embedding
+
+00:29:40.880 --> 00:29:44.760
+model
+
+00:29:42.799 --> 00:29:48.039
+so however if you want to get the
+
+00:29:44.760 --> 00:29:51.080
+highest accuracy definitely tuned models
+
+00:29:48.039 --> 00:29:53.200
+are going to be better and if you're not
+
+00:29:51.080 --> 00:29:56.039
+worried about computation efficiency
+
+00:29:53.200 --> 00:29:58.480
+using something like P bear um with kind
+
+00:29:56.039 --> 00:30:01.320
+of the token level retrieval will
+
+00:29:58.480 --> 00:30:05.559
+definitely give you uh good accuracy
+
+00:30:01.320 --> 00:30:08.559
+here however there's better support for
+
+00:30:05.559 --> 00:30:12.159
+bu encoder style models um in kind of
+
+00:30:08.559 --> 00:30:15.240
+standard Vector databases like feice and
+
+00:30:12.159 --> 00:30:17.519
+uh chroma and other things like that so
+
+00:30:15.240 --> 00:30:19.799
+if you want a kind of easier method to
+
+00:30:17.519 --> 00:30:23.279
+get started very quickly then using a bu
+
+00:30:19.799 --> 00:30:23.279
+encoder is probably the best way to
+
+00:30:25.080 --> 00:30:31.080
+go okay so now moving on to actual
+
+00:30:28.279 --> 00:30:33.159
+retrieval augmented generation models we
+
+00:30:31.080 --> 00:30:38.360
+have uh retriever reader
+
+00:30:33.159 --> 00:30:40.880
+models and the way these work is you
+
+00:30:38.360 --> 00:30:43.279
+basically the simplest way they can work
+
+00:30:40.880 --> 00:30:45.799
+is you basically just chain retrieval
+
+00:30:43.279 --> 00:30:47.640
+and reading together so you use an outof
+
+00:30:45.799 --> 00:30:52.519
+thebox Retriever and an outof thebox
+
+00:30:47.640 --> 00:30:54.039
+reader model and you have your query uh
+
+00:30:52.519 --> 00:30:56.159
+you could for example look something up
+
+00:30:54.039 --> 00:30:58.039
+on Google get a whole bunch of passages
+
+00:30:56.159 --> 00:30:59.760
+and then feed them into a GP key model
+
+00:30:58.039 --> 00:31:03.919
+and get an
+
+00:30:59.760 --> 00:31:06.960
+answer this overall is quite effective
+
+00:31:03.919 --> 00:31:09.159
+um you it's easy to implement and it
+
+00:31:06.960 --> 00:31:10.600
+will give you decent results so
+
+00:31:09.159 --> 00:31:15.480
+definitely it's something to be worth
+
+00:31:10.600 --> 00:31:20.720
+thinking about uh for assignment two in
+
+00:31:15.480 --> 00:31:24.799
+the um in the class you're required to
+
+00:31:20.720 --> 00:31:26.679
+only use uh kind of public models or
+
+00:31:24.799 --> 00:31:29.760
+open source implementations so you could
+
+00:31:26.679 --> 00:31:34.360
+still replace that with Apachi Lucine
+
+00:31:29.760 --> 00:31:36.360
+and then um you know any standard llm
+
+00:31:34.360 --> 00:31:39.159
+and that could be you know llama llama
+
+00:31:36.360 --> 00:31:41.600
+Chad or M mistol or mixol or something
+
+00:31:39.159 --> 00:31:45.360
+like that so uh you could definitely
+
+00:31:41.600 --> 00:31:48.120
+feel feel free to do something like
+
+00:31:45.360 --> 00:31:51.559
+that um of course the passages are
+
+00:31:48.120 --> 00:31:53.200
+concatenated to the context and so
+
+00:31:51.559 --> 00:31:54.799
+because the passages are concatenated to
+
+00:31:53.200 --> 00:31:56.679
+context the contacts can get relatively
+
+00:31:54.799 --> 00:31:58.399
+long and expensive and other things like
+
+00:31:56.679 --> 00:32:01.960
+that but it's just something you have to
+
+00:31:58.399 --> 00:32:01.960
+deal with when you're using
+
+00:32:02.600 --> 00:32:07.480
+R there are methods also for Retriever
+
+00:32:05.799 --> 00:32:11.600
+and Generator endtoend
+
+00:32:07.480 --> 00:32:14.720
+training so this is the paper actually
+
+00:32:11.600 --> 00:32:17.600
+where the name rag came from and I'll
+
+00:32:14.720 --> 00:32:20.200
+use that as an example here uh but
+
+00:32:17.600 --> 00:32:21.600
+basically um there are several methods
+
+00:32:20.200 --> 00:32:23.399
+that propos to train the Retriever and
+
+00:32:21.600 --> 00:32:27.440
+reader to improve
+
+00:32:23.399 --> 00:32:31.240
+accuracy and specifically the rag p by
+
+00:32:27.440 --> 00:32:33.200
+Lewis at all the way it trained the um
+
+00:32:31.240 --> 00:32:35.639
+reader was to maximize generation
+
+00:32:33.200 --> 00:32:38.600
+likelihood given a single retrieved
+
+00:32:35.639 --> 00:32:40.279
+document and for the retriever it
+
+00:32:38.600 --> 00:32:41.880
+maximized overall likelihood by
+
+00:32:40.279 --> 00:32:44.480
+optimizing the mixture weight over
+
+00:32:41.880 --> 00:32:46.559
+documents so here's kind of a a
+
+00:32:44.480 --> 00:32:50.480
+schematic uh which is you have your
+
+00:32:46.559 --> 00:32:54.039
+query encoder um you run the Retriever
+
+00:32:50.480 --> 00:32:57.760
+with uh maximum inner product search it
+
+00:32:54.039 --> 00:33:00.919
+gives you several documents and each
+
+00:32:57.760 --> 00:33:05.880
+document has a score and then based on
+
+00:33:00.919 --> 00:33:09.399
+the documents and the scores you
+
+00:33:05.880 --> 00:33:11.200
+generate uh with each document in the
+
+00:33:09.399 --> 00:33:15.360
+context and
+
+00:33:11.200 --> 00:33:17.080
+then sum together the probabilities
+
+00:33:15.360 --> 00:33:18.639
+multiplied by the weights and I have the
+
+00:33:17.080 --> 00:33:20.320
+actual equations here because I think
+
+00:33:18.639 --> 00:33:23.039
+it'll be a little bit easier to
+
+00:33:20.320 --> 00:33:25.760
+understand after looking at the
+
+00:33:23.039 --> 00:33:28.360
+equations so generation is a mixture
+
+00:33:25.760 --> 00:33:31.440
+model and you pick a document and
+
+00:33:28.360 --> 00:33:36.519
+generate from the document this
+
+00:33:31.440 --> 00:33:40.080
+p z given X is the probability of
+
+00:33:36.519 --> 00:33:44.679
+picking that document given the query X
+
+00:33:40.080 --> 00:33:48.880
+and then this P Theta x z and all of the
+
+00:33:44.679 --> 00:33:51.480
+previous tokens is basically the uh
+
+00:33:48.880 --> 00:33:54.840
+probability of the next token given that
+
+00:33:51.480 --> 00:33:56.559
+you have this particular document so you
+
+00:33:54.840 --> 00:34:00.840
+can see that this is basically linearly
+
+00:33:56.559 --> 00:34:00.840
+interpr ating between the multiple
+
+00:34:01.559 --> 00:34:05.760
+documents and if we look this can be
+
+00:34:04.600 --> 00:34:09.039
+considered the Retriever and the
+
+00:34:05.760 --> 00:34:09.039
+generator the Retriever and the
+
+00:34:10.839 --> 00:34:16.119
+reader one really important thing here
+
+00:34:13.639 --> 00:34:17.760
+uh that enables endtoend training is
+
+00:34:16.119 --> 00:34:19.639
+they have this probability of the
+
+00:34:17.760 --> 00:34:22.919
+retriever be based on
+
+00:34:19.639 --> 00:34:25.480
+embeddings and so here we have the
+
+00:34:22.919 --> 00:34:29.040
+document embedding and the query
+
+00:34:25.480 --> 00:34:31.440
+embedding and the probability is
+
+00:34:29.040 --> 00:34:33.320
+proportional to the inner product of
+
+00:34:31.440 --> 00:34:36.599
+these exponentiated so you're basically
+
+00:34:33.320 --> 00:34:38.839
+taking a soft Max over uh the inner
+
+00:34:36.599 --> 00:34:40.599
+product between the
+
+00:34:38.839 --> 00:34:44.200
+two
+
+00:34:40.599 --> 00:34:47.919
+and this adjusts the retriever to give
+
+00:34:44.200 --> 00:34:49.560
+higher similarities to helpful
+
+00:34:47.919 --> 00:34:52.560
+documents
+
+00:34:49.560 --> 00:34:52.560
+and
+
+00:34:54.040 --> 00:35:02.800
+so because the prob probability of the
+
+00:34:59.800 --> 00:35:04.839
+retriever model here is included in the
+
+00:35:02.800 --> 00:35:07.160
+endtoend probability you don't actually
+
+00:35:04.839 --> 00:35:10.680
+need any annotations
+
+00:35:07.160 --> 00:35:12.839
+about which documents are useful you can
+
+00:35:10.680 --> 00:35:16.680
+just train all of this end to end and
+
+00:35:12.839 --> 00:35:19.480
+let backrop do its thing to update the
+
+00:35:16.680 --> 00:35:22.640
+uh the retriever as
+
+00:35:19.480 --> 00:35:25.000
+well one important issue when training
+
+00:35:22.640 --> 00:35:27.480
+models like this is that the search
+
+00:35:25.000 --> 00:35:30.400
+index will become stale so what do I
+
+00:35:27.480 --> 00:35:34.760
+mean by this if we go back to our
+
+00:35:30.400 --> 00:35:34.760
+previous uh thing about dense
+
+00:35:35.480 --> 00:35:43.560
+models creating this blue search index
+
+00:35:39.800 --> 00:35:45.400
+on the right side of the figure here is
+
+00:35:43.560 --> 00:35:48.680
+very costly so like let's say you want
+
+00:35:45.400 --> 00:35:50.520
+to embed a million documents or a
+
+00:35:48.680 --> 00:35:55.240
+billion documents if you're a big search
+
+00:35:50.520 --> 00:35:58.200
+engine company so doing this is very
+
+00:35:55.240 --> 00:36:00.599
+slow and
+
+00:35:58.200 --> 00:36:01.920
+in contrast doing lookup with kind of
+
+00:36:00.599 --> 00:36:04.160
+these approximate nearest neighbor
+
+00:36:01.920 --> 00:36:05.440
+searches is sublinear time or even you
+
+00:36:04.160 --> 00:36:08.119
+know log time so you can do it
+
+00:36:05.440 --> 00:36:12.319
+relatively quickly
+
+00:36:08.119 --> 00:36:15.680
+so it's fine to do lookup over this big
+
+00:36:12.319 --> 00:36:17.520
+index but if you start updating this
+
+00:36:15.680 --> 00:36:19.920
+document embedding you need to recreate
+
+00:36:17.520 --> 00:36:23.760
+the entire index and that would be you
+
+00:36:19.920 --> 00:36:27.240
+know very computationally costly so the
+
+00:36:23.760 --> 00:36:30.119
+solution to this proposed in this rag
+
+00:36:27.240 --> 00:36:33.640
+paper by Lewis at all is uh we only
+
+00:36:30.119 --> 00:36:35.640
+train the query embeddings and we keep
+
+00:36:33.640 --> 00:36:39.640
+the document embedding
+
+00:36:35.640 --> 00:36:41.920
+swix there's other Alternatives like um
+
+00:36:39.640 --> 00:36:45.000
+there was a paper called realm uh from
+
+00:36:41.920 --> 00:36:48.040
+early in retrieval base modeling and in
+
+00:36:45.000 --> 00:36:50.040
+that in that method they basically had
+
+00:36:48.040 --> 00:36:51.520
+an asynchronous process that was going
+
+00:36:50.040 --> 00:36:55.760
+through and using the most recent
+
+00:36:51.520 --> 00:36:59.960
+document in better to re-update the
+
+00:36:55.760 --> 00:37:03.359
+search index during training but that is
+
+00:36:59.960 --> 00:37:05.960
+uh you know kind of a very onerous
+
+00:37:03.359 --> 00:37:07.800
+process so I think it's quite common to
+
+00:37:05.960 --> 00:37:11.000
+use kind of a fixed document embedding
+
+00:37:07.800 --> 00:37:11.000
+in update only the
+
+00:37:12.079 --> 00:37:17.720
+queries another thing to think about is
+
+00:37:14.359 --> 00:37:21.160
+when do we do retrieval um so there's a
+
+00:37:17.720 --> 00:37:23.079
+bunch of different methods the rag paper
+
+00:37:21.160 --> 00:37:24.440
+that I mentioned before did this only
+
+00:37:23.079 --> 00:37:26.359
+once right at the very beginning of
+
+00:37:24.440 --> 00:37:29.400
+generation it grabbed a single document
+
+00:37:26.359 --> 00:37:32.560
+and generated the entire output this is
+
+00:37:29.400 --> 00:37:34.800
+the default method used by most
+
+00:37:32.560 --> 00:37:37.240
+systems however there's other options as
+
+00:37:34.800 --> 00:37:39.640
+well you can retrieve uh several times
+
+00:37:37.240 --> 00:37:43.040
+during generation as
+
+00:37:39.640 --> 00:37:44.480
+necessary and the way this works uh we
+
+00:37:43.040 --> 00:37:46.280
+can do this either by generating a
+
+00:37:44.480 --> 00:37:48.480
+search token uh saying that we should
+
+00:37:46.280 --> 00:37:50.200
+start searching or searching when the
+
+00:37:48.480 --> 00:37:52.640
+model is
+
+00:37:50.200 --> 00:37:55.920
+uncertain and another way is to do this
+
+00:37:52.640 --> 00:37:58.079
+every token so we can do this by finding
+
+00:37:55.920 --> 00:37:59.760
+similar final embeddings and using this
+
+00:37:58.079 --> 00:38:02.240
+to influence the
+
+00:37:59.760 --> 00:38:04.720
+probabilities or approximating attention
+
+00:38:02.240 --> 00:38:06.440
+with nearest neighbors so I'm going to
+
+00:38:04.720 --> 00:38:08.920
+explain about each of these in a bit
+
+00:38:06.440 --> 00:38:12.480
+more detail
+
+00:38:08.920 --> 00:38:16.119
+in so triggering retrieval with token
+
+00:38:12.480 --> 00:38:19.720
+embeddings is um was proposed by Tool
+
+00:38:16.119 --> 00:38:22.119
+forer shik all and the way it works is
+
+00:38:19.720 --> 00:38:25.000
+you generate tokens that Tri trigger
+
+00:38:22.119 --> 00:38:27.880
+retrieval or other tools so in this
+
+00:38:25.000 --> 00:38:30.079
+particular method it uh had several
+
+00:38:27.880 --> 00:38:32.000
+tools including asking a QA model or
+
+00:38:30.079 --> 00:38:34.800
+getting a calculator or having a machine
+
+00:38:32.000 --> 00:38:37.200
+translation system but with respect to
+
+00:38:34.800 --> 00:38:40.000
+retrieval augmented generation it had
+
+00:38:37.200 --> 00:38:41.560
+this essentially Wiki search
+
+00:38:40.000 --> 00:38:43.680
+functionality that would look up
+
+00:38:41.560 --> 00:38:46.680
+something in Wikipedia and then use that
+
+00:38:43.680 --> 00:38:46.680
+to influence the final
+
+00:38:46.760 --> 00:38:52.200
+probabilities
+
+00:38:48.800 --> 00:38:55.160
+and the way this was trained is training
+
+00:38:52.200 --> 00:38:59.800
+was done in an inative manner where it
+
+00:38:55.160 --> 00:38:59.800
+basically generated uh kind
+
+00:39:00.000 --> 00:39:05.680
+of examples of tools being useful and
+
+00:39:04.359 --> 00:39:09.560
+when the
+
+00:39:05.680 --> 00:39:14.160
+tools improve the probability of the
+
+00:39:09.560 --> 00:39:16.119
+following output then that would be kind
+
+00:39:14.160 --> 00:39:19.560
+of treated as a positive example and
+
+00:39:16.119 --> 00:39:21.520
+used to further train the model so this
+
+00:39:19.560 --> 00:39:23.400
+was really influential and in fact this
+
+00:39:21.520 --> 00:39:27.000
+is how things are implemented in chat
+
+00:39:23.400 --> 00:39:29.319
+GPT nowadays not only for um doing
+
+00:39:27.000 --> 00:39:33.400
+retrieval but also doing other tools
+
+00:39:29.319 --> 00:39:35.200
+like um for example uh generating code
+
+00:39:33.400 --> 00:39:37.440
+or generating images or other things
+
+00:39:35.200 --> 00:39:37.440
+like
+
+00:39:38.200 --> 00:39:45.079
+this another option is to trigger
+
+00:39:40.920 --> 00:39:48.240
+retrieval uh with uncertainty estimates
+
+00:39:45.079 --> 00:39:52.280
+so flare this is a paper by my student
+
+00:39:48.240 --> 00:39:55.160
+Jang bang um where we try to generate
+
+00:39:52.280 --> 00:39:58.560
+content and then do retrieval if the
+
+00:39:55.160 --> 00:40:01.800
+language model certainty is low so
+
+00:39:58.560 --> 00:40:05.599
+here's a schematic of how this works but
+
+00:40:01.800 --> 00:40:09.160
+basically um if we have
+
+00:40:05.599 --> 00:40:13.440
+some uh retrieved documents we can say
+
+00:40:09.160 --> 00:40:16.560
+generate a a summary about Joe Biden and
+
+00:40:13.440 --> 00:40:19.560
+when it generates a summary maybe for
+
+00:40:16.560 --> 00:40:20.960
+the first output um the language model
+
+00:40:19.560 --> 00:40:22.960
+has high
+
+00:40:20.960 --> 00:40:24.240
+confidence and because the language
+
+00:40:22.960 --> 00:40:25.359
+model has high confidence we just
+
+00:40:24.240 --> 00:40:27.520
+generate the
+
+00:40:25.359 --> 00:40:29.599
+output
+
+00:40:27.520 --> 00:40:31.839
+however in the next step if it might
+
+00:40:29.599 --> 00:40:33.599
+generate something like saying Joe Biden
+
+00:40:31.839 --> 00:40:35.680
+attended the University of Pennsylvania
+
+00:40:33.599 --> 00:40:37.160
+where he earned a law degree but the
+
+00:40:35.680 --> 00:40:39.000
+model might not be very certain about
+
+00:40:37.160 --> 00:40:41.560
+this it might have a low probability of
+
+00:40:39.000 --> 00:40:45.839
+certain important entities and So based
+
+00:40:41.560 --> 00:40:48.839
+on this uh we then form a a query where
+
+00:40:45.839 --> 00:40:52.119
+what we do is essentially we blank out
+
+00:40:48.839 --> 00:40:55.079
+the low probability parts of this and we
+
+00:40:52.119 --> 00:40:57.200
+do a search and so this is also a little
+
+00:40:55.079 --> 00:41:00.240
+bit like the hypothetical
+
+00:40:57.200 --> 00:41:02.520
+edings method where we basically create
+
+00:41:00.240 --> 00:41:04.040
+a document that we think will look
+
+00:41:02.520 --> 00:41:07.119
+similar to the document that we want to
+
+00:41:04.040 --> 00:41:09.480
+find we use that to create search
+
+00:41:07.119 --> 00:41:11.359
+results and then we generate the output
+
+00:41:09.480 --> 00:41:13.880
+and then we continue doing that and
+
+00:41:11.359 --> 00:41:15.960
+whenever we have a high confidence
+
+00:41:13.880 --> 00:41:18.800
+output like the one here we don't do any
+
+00:41:15.960 --> 00:41:20.040
+retrieval we just you know generate uh
+
+00:41:18.800 --> 00:41:21.880
+directly from the parameters of the
+
+00:41:20.040 --> 00:41:23.960
+model but whenever we have low
+
+00:41:21.880 --> 00:41:27.400
+confidence outputs we do the retrieval
+
+00:41:23.960 --> 00:41:30.400
+and base the output on this and so I I
+
+00:41:27.400 --> 00:41:33.119
+think this is uh you know a nice method
+
+00:41:30.400 --> 00:41:35.000
+that could potentially be uh used the
+
+00:41:33.119 --> 00:41:36.920
+downside to that is you might sometimes
+
+00:41:35.000 --> 00:41:38.920
+need to generate twice because you would
+
+00:41:36.920 --> 00:41:40.480
+generate the output once and then find
+
+00:41:38.920 --> 00:41:42.720
+the low confidence parts and generate
+
+00:41:40.480 --> 00:41:45.400
+again but you know if you really care
+
+00:41:42.720 --> 00:41:47.319
+about the uh kind of quality of the
+
+00:41:45.400 --> 00:41:49.640
+output this is I think a reasonable
+
+00:41:47.319 --> 00:41:49.640
+thing to
+
+00:41:50.160 --> 00:41:54.920
+do okay so now moving on to the Token by
+
+00:41:53.000 --> 00:41:59.800
+token retrieval
+
+00:41:54.920 --> 00:42:03.560
+methods the kind of original or one of
+
+00:41:59.800 --> 00:42:05.200
+the methods that popularized this idea
+
+00:42:03.560 --> 00:42:08.720
+of token by token retrieval is something
+
+00:42:05.200 --> 00:42:10.760
+called K&N LM and the way it works is it
+
+00:42:08.720 --> 00:42:13.839
+retrieves similar
+
+00:42:10.760 --> 00:42:16.680
+examples and then uses the following
+
+00:42:13.839 --> 00:42:20.880
+tokens from these
+
+00:42:16.680 --> 00:42:23.800
+examples and this is kind of like a very
+
+00:42:20.880 --> 00:42:25.839
+powerful count-based byr model in a way
+
+00:42:23.800 --> 00:42:28.440
+so if you remember back to when we were
+
+00:42:25.839 --> 00:42:32.920
+talking about count based Pam models
+
+00:42:28.440 --> 00:42:36.440
+what we would do is we would take the
+
+00:42:32.920 --> 00:42:39.400
+previous token and we would calculate
+
+00:42:36.440 --> 00:42:41.319
+the probability of the next token by
+
+00:42:39.400 --> 00:42:43.040
+summing up together all of the next
+
+00:42:41.319 --> 00:42:44.800
+tokens and dividing by the total number
+
+00:42:43.040 --> 00:42:49.240
+of times that previous token
+
+00:42:44.800 --> 00:42:52.720
+occurred and so given that background uh
+
+00:42:49.240 --> 00:42:56.760
+we can talk about how the KLM
+
+00:42:52.720 --> 00:43:00.319
+works so we have the text context X
+
+00:42:56.760 --> 00:43:02.240
+and we want to generate a Target output
+
+00:43:00.319 --> 00:43:04.839
+separately from this we have all of the
+
+00:43:02.240 --> 00:43:06.440
+training contexts so this is all of the
+
+00:43:04.839 --> 00:43:09.920
+contexts that appeared in our training
+
+00:43:06.440 --> 00:43:13.520
+data and we encode all of these training
+
+00:43:09.920 --> 00:43:15.720
+contexts specifically by calculating the
+
+00:43:13.520 --> 00:43:18.559
+representation of the final layer or
+
+00:43:15.720 --> 00:43:21.119
+near the final layer of the model and so
+
+00:43:18.559 --> 00:43:23.200
+we encode that as
+
+00:43:21.119 --> 00:43:25.240
+representations separately from that we
+
+00:43:23.200 --> 00:43:27.920
+remember the next word that appeared
+
+00:43:25.240 --> 00:43:29.720
+after this Contex
+
+00:43:27.920 --> 00:43:32.920
+so now we have a data store consisting
+
+00:43:29.720 --> 00:43:35.040
+of representations in next words we then
+
+00:43:32.920 --> 00:43:38.440
+take the representation of the current
+
+00:43:35.040 --> 00:43:40.880
+context and we calculate the distance
+
+00:43:38.440 --> 00:43:43.400
+between the current context and all of
+
+00:43:40.880 --> 00:43:47.119
+the other similar context in the
+
+00:43:43.400 --> 00:43:49.839
+database we take the nearest K so we
+
+00:43:47.119 --> 00:43:52.440
+take the top uh K examples here which
+
+00:43:49.839 --> 00:43:55.240
+would be Hawaii Illinois and
+
+00:43:52.440 --> 00:43:57.520
+Hawaii we then do uh some sort of
+
+00:43:55.240 --> 00:44:01.440
+normalization based on the
+
+00:43:57.520 --> 00:44:05.200
+distance and this gives us a probability
+
+00:44:01.440 --> 00:44:06.680
+distribution over all of the next tokens
+
+00:44:05.200 --> 00:44:10.599
+sometimes these tokens are duplicated
+
+00:44:06.680 --> 00:44:13.599
+multiple times and so we aggregate all
+
+00:44:10.599 --> 00:44:15.800
+of these counts to be Hawaii for example
+
+00:44:13.599 --> 00:44:18.839
+0.8 and Illinois
+
+00:44:15.800 --> 00:44:21.839
+0.2 and then we interpolate this with
+
+00:44:18.839 --> 00:44:24.040
+the probability given by the standard
+
+00:44:21.839 --> 00:44:26.440
+language model using an interpolation
+
+00:44:24.040 --> 00:44:28.400
+coefficient Lambda and this gives us our
+
+00:44:26.440 --> 00:44:31.000
+final
+
+00:44:28.400 --> 00:44:34.559
+probability so the nice thing about this
+
+00:44:31.000 --> 00:44:38.000
+is this allows us to explicitly ground
+
+00:44:34.559 --> 00:44:42.079
+our outputs in individual
+
+00:44:38.000 --> 00:44:45.319
+examples uh and it's a pretty effective
+
+00:44:42.079 --> 00:44:48.760
+way to improve the probability of models
+
+00:44:45.319 --> 00:44:53.839
+improve translation and other stuff like
+
+00:44:48.760 --> 00:44:56.119
+this the disadvantage of doing this is
+
+00:44:53.839 --> 00:44:59.319
+that it provides it it kind of ADD add
+
+00:44:56.119 --> 00:45:01.800
+an extra component of the model it adds
+
+00:44:59.319 --> 00:45:05.440
+extra
+
+00:45:01.800 --> 00:45:08.520
+um kind of hyperparameters like Lambda
+
+00:45:05.440 --> 00:45:11.680
+and things like this so it is a little
+
+00:45:08.520 --> 00:45:16.960
+bit finicky and it doesn't work in all
+
+00:45:11.680 --> 00:45:21.440
+situations and so another method that we
+
+00:45:16.960 --> 00:45:23.559
+uh proposed or by Manda Birch who gave
+
+00:45:21.440 --> 00:45:26.920
+the uh previous lecture on generation in
+
+00:45:23.559 --> 00:45:29.240
+this class is unlimi forer and basically
+
+00:45:26.920 --> 00:45:32.680
+what unlimi forer does is it notes that
+
+00:45:29.240 --> 00:45:36.079
+attention itself is an in inner product
+
+00:45:32.680 --> 00:45:40.440
+search and it does topk
+
+00:45:36.079 --> 00:45:42.680
+attention and the way we do this is we
+
+00:45:40.440 --> 00:45:45.160
+first process the input with a sliding
+
+00:45:42.680 --> 00:45:47.480
+window and then perform attention using
+
+00:45:45.160 --> 00:45:49.960
+a vector index so if we have a really
+
+00:45:47.480 --> 00:45:54.280
+long input that we want to encode what
+
+00:45:49.960 --> 00:45:56.559
+we do is we first encode chunks so we
+
+00:45:54.280 --> 00:46:01.960
+encode for example AB
+
+00:45:56.559 --> 00:46:03.839
+then we encode CD and we encode EF we
+
+00:46:01.960 --> 00:46:06.240
+concatenate them together into a big
+
+00:46:03.839 --> 00:46:07.800
+index of one long input so in a way that
+
+00:46:06.240 --> 00:46:10.920
+this is similar to what they did in the
+
+00:46:07.800 --> 00:46:12.720
+KLM you know concatenate all of these
+
+00:46:10.920 --> 00:46:16.520
+embeddings into a single
+
+00:46:12.720 --> 00:46:18.680
+input but the difference is that this is
+
+00:46:16.520 --> 00:46:21.640
+done with
+
+00:46:18.680 --> 00:46:24.280
+um the values that we are attending to
+
+00:46:21.640 --> 00:46:27.559
+as opposed to just the final
+
+00:46:24.280 --> 00:46:30.079
+layer and
+
+00:46:27.559 --> 00:46:33.680
+the interesting thing about this is now
+
+00:46:30.079 --> 00:46:36.200
+we have an index of one long input and
+
+00:46:33.680 --> 00:46:39.800
+when we want to do our next version of
+
+00:46:36.200 --> 00:46:42.240
+attention we do KNN search from the
+
+00:46:39.800 --> 00:46:44.280
+query we take the retrieved hidden
+
+00:46:42.240 --> 00:46:47.880
+States and then we just do attention
+
+00:46:44.280 --> 00:46:50.440
+over them so the nice thing about this
+
+00:46:47.880 --> 00:46:53.079
+is in the extreme case this makes no
+
+00:46:50.440 --> 00:46:55.240
+changes to the model what I mean by this
+
+00:46:53.079 --> 00:46:57.520
+is let's say our input was small enough
+
+00:46:55.240 --> 00:47:02.240
+that we could coded in only a single
+
+00:46:57.520 --> 00:47:06.400
+chunk and for KNN search we also did KNN
+
+00:47:02.240 --> 00:47:09.559
+search um we did you know exact Canon
+
+00:47:06.400 --> 00:47:12.400
+search over all of the embeddings in the
+
+00:47:09.559 --> 00:47:14.680
+trunk in that case this would just be
+
+00:47:12.400 --> 00:47:16.520
+normal attention it's exactly the same
+
+00:47:14.680 --> 00:47:18.640
+as normal
+
+00:47:16.520 --> 00:47:20.160
+attention however there are some
+
+00:47:18.640 --> 00:47:21.760
+approximations that go into here like
+
+00:47:20.160 --> 00:47:24.000
+when we encode chunks they might not be
+
+00:47:21.760 --> 00:47:26.359
+exactly the same as if we encoded the
+
+00:47:24.000 --> 00:47:29.839
+entire thing together and we're also
+
+00:47:26.359 --> 00:47:33.640
+chopping off some of the values with
+
+00:47:29.839 --> 00:47:35.800
+very low um kind of inner products and
+
+00:47:33.640 --> 00:47:37.400
+so because of this there are some
+
+00:47:35.800 --> 00:47:38.760
+approximations being made but in the
+
+00:47:37.400 --> 00:47:40.160
+extreme case if we made no
+
+00:47:38.760 --> 00:47:41.880
+approximations this would just be
+
+00:47:40.160 --> 00:47:44.359
+exactly the same model as we were using
+
+00:47:41.880 --> 00:47:46.160
+before so I find this pretty attractive
+
+00:47:44.359 --> 00:47:48.760
+and uh you know empirically it gives
+
+00:47:46.160 --> 00:47:51.720
+very good results over long
+
+00:47:48.760 --> 00:47:53.440
+distances and you know we can always
+
+00:47:51.720 --> 00:47:56.240
+make our approximations better and
+
+00:47:53.440 --> 00:47:57.680
+improve this model as well so I I think
+
+00:47:56.240 --> 00:48:00.960
+this is a attractive method that you
+
+00:47:57.680 --> 00:48:00.960
+might be interested in taking a look
+
+00:48:02.240 --> 00:48:06.200
+at okay for the final part of this I'd
+
+00:48:04.559 --> 00:48:08.079
+like to talk about long context
+
+00:48:06.200 --> 00:48:12.400
+Transformers and these are models that
+
+00:48:08.079 --> 00:48:15.119
+are explicitly trained in a way that
+
+00:48:12.400 --> 00:48:16.920
+allows you to attend to longer contexts
+
+00:48:15.119 --> 00:48:18.839
+in an efficient
+
+00:48:16.920 --> 00:48:21.960
+manner
+
+00:48:18.839 --> 00:48:23.680
+so one way that we can train over longer
+
+00:48:21.960 --> 00:48:25.880
+context is just append all of the
+
+00:48:23.680 --> 00:48:28.040
+context together and in fact shortly
+
+00:48:25.880 --> 00:48:32.200
+after Transformers came out uh this
+
+00:48:28.040 --> 00:48:34.280
+paper by VOA at all demonstrated that um
+
+00:48:32.200 --> 00:48:36.160
+it doing this can learn you know
+
+00:48:34.280 --> 00:48:38.119
+interesting document level phenomena so
+
+00:48:36.160 --> 00:48:40.440
+it can identify when
+
+00:48:38.119 --> 00:48:42.480
+multiple uh words refer to the same
+
+00:48:40.440 --> 00:48:43.680
+thing or co-reference and other things
+
+00:48:42.480 --> 00:48:45.640
+like
+
+00:48:43.680 --> 00:48:47.720
+this however the problem with
+
+00:48:45.640 --> 00:48:51.119
+Transformers is that computation is
+
+00:48:47.720 --> 00:48:52.799
+quadratic in the sentence length because
+
+00:48:51.119 --> 00:48:54.599
+you're multiplying all of the query
+
+00:48:52.799 --> 00:48:56.799
+vectors by all of the key
+
+00:48:54.599 --> 00:48:59.480
+vectors
+
+00:48:56.799 --> 00:49:02.799
+and that basically causes a big problem
+
+00:48:59.480 --> 00:49:02.799
+if your sequences become very
+
+00:49:03.480 --> 00:49:09.760
+long so if we go back to what we did in
+
+00:49:07.480 --> 00:49:12.400
+rnns uh from the very beginning of the
+
+00:49:09.760 --> 00:49:14.359
+class in rnns they don't have this
+
+00:49:12.400 --> 00:49:16.280
+problem because computation is linear in
+
+00:49:14.359 --> 00:49:20.440
+the length of the sequence you just pass
+
+00:49:16.280 --> 00:49:22.200
+along the RNN State and every single
+
+00:49:20.440 --> 00:49:23.839
+time you do the same computation over it
+
+00:49:22.200 --> 00:49:26.559
+so there's no quadratic term in
+
+00:49:23.839 --> 00:49:32.400
+calculating rnns
+
+00:49:26.559 --> 00:49:34.880
+another thing is that when doing rnns
+
+00:49:32.400 --> 00:49:37.680
+you can actually P State infinitely
+
+00:49:34.880 --> 00:49:39.040
+during the forward pass by just
+
+00:49:37.680 --> 00:49:40.240
+calculating the hidden State and then
+
+00:49:39.040 --> 00:49:42.119
+throwing away the rest of the
+
+00:49:40.240 --> 00:49:43.359
+computation graph that was used in
+
+00:49:42.119 --> 00:49:45.160
+calculating that hidden State and
+
+00:49:43.359 --> 00:49:48.319
+there's no approximation that goes on
+
+00:49:45.160 --> 00:49:49.680
+there so unlike on in un liform that I
+
+00:49:48.319 --> 00:49:51.640
+was talking about before where we needed
+
+00:49:49.680 --> 00:49:54.119
+to make approximations none need to be
+
+00:49:51.640 --> 00:49:56.400
+made in this
+
+00:49:54.119 --> 00:50:00.200
+case however there is a problem with
+
+00:49:56.400 --> 00:50:02.040
+doing back propop uh because in order to
+
+00:50:00.200 --> 00:50:05.839
+do back propop normally you maintain the
+
+00:50:02.040 --> 00:50:09.720
+entire you know state of the computation
+
+00:50:05.839 --> 00:50:12.400
+graph and so there a common method to
+
+00:50:09.720 --> 00:50:15.280
+fix this is basically you pass along the
+
+00:50:12.400 --> 00:50:16.920
+RNN state from the previous sentence but
+
+00:50:15.280 --> 00:50:19.240
+you just don't do backdrop into the
+
+00:50:16.920 --> 00:50:21.200
+previous sentence and this is called
+
+00:50:19.240 --> 00:50:24.040
+truncated backrop or truncated back
+
+00:50:21.200 --> 00:50:27.280
+propagation through time and this allows
+
+00:50:24.040 --> 00:50:30.160
+you to essentially train models with
+
+00:50:27.280 --> 00:50:32.319
+infinite context um or at least models
+
+00:50:30.160 --> 00:50:33.720
+that can pass along context infinitely
+
+00:50:32.319 --> 00:50:36.359
+even if you're not back propping into
+
+00:50:33.720 --> 00:50:36.359
+they Cod ear
+
+00:50:37.480 --> 00:50:43.520
+there so of course a problem with this
+
+00:50:40.720 --> 00:50:45.880
+over long contexts is recurrents uh
+
+00:50:43.520 --> 00:50:47.520
+recurrent models can be slow due to the
+
+00:50:45.880 --> 00:50:51.400
+kind of sequential dependence they're
+
+00:50:47.520 --> 00:50:54.280
+not ideal for um you know running on
+
+00:50:51.400 --> 00:50:57.359
+gpus or things like that and this is
+
+00:50:54.280 --> 00:51:01.960
+improved by recent architectures like
+
+00:50:57.359 --> 00:51:05.359
+Mamba and RW KV which are more conducive
+
+00:51:01.960 --> 00:51:07.079
+to GPU Based training um while still
+
+00:51:05.359 --> 00:51:08.599
+maintaining linear time complexity and
+
+00:51:07.079 --> 00:51:11.480
+so I'm looking forward to talking about
+
+00:51:08.599 --> 00:51:11.480
+that more in a future
+
+00:51:13.000 --> 00:51:17.559
+class so actually if we take this idea
+
+00:51:15.880 --> 00:51:20.440
+of truncated back propagation through
+
+00:51:17.559 --> 00:51:22.359
+time this can also be applied to
+
+00:51:20.440 --> 00:51:25.440
+Transformers and there's a really nice
+
+00:51:22.359 --> 00:51:27.880
+paper Transformer XEL also created by
+
+00:51:25.440 --> 00:51:31.119
+kungai who was formerly at
+
+00:51:27.880 --> 00:51:33.119
+CMU and what this does is this attempts
+
+00:51:31.119 --> 00:51:35.760
+to fix vectors from the previous
+
+00:51:33.119 --> 00:51:39.440
+sentence so if we have a standard
+
+00:51:35.760 --> 00:51:40.720
+Transformer uh in a Transformer XL
+
+00:51:39.440 --> 00:51:44.640
+normally what we do in the standard
+
+00:51:40.720 --> 00:51:48.480
+Transformer is each Vector attends back
+
+00:51:44.640 --> 00:51:50.920
+to all the other vectors in the current
+
+00:51:48.480 --> 00:51:53.839
+context what Transformer XEL does
+
+00:51:50.920 --> 00:51:56.359
+instead is when you have a new segment
+
+00:51:53.839 --> 00:51:58.960
+that you want to do backrop
+
+00:51:56.359 --> 00:52:01.200
+into um you have a new segment that you
+
+00:51:58.960 --> 00:52:03.960
+want to basically train over you also
+
+00:52:01.200 --> 00:52:06.400
+attend to all of the previous tokens in
+
+00:52:03.960 --> 00:52:07.640
+the previous segment but you don't do
+
+00:52:06.400 --> 00:52:10.319
+back propop into
+
+00:52:07.640 --> 00:52:12.079
+them so this is essentially truncated
+
+00:52:10.319 --> 00:52:14.480
+backpropagation through time from the
+
+00:52:12.079 --> 00:52:17.760
+Transformer
+
+00:52:14.480 --> 00:52:19.520
+perspective this is also really nice
+
+00:52:17.760 --> 00:52:21.200
+because what it allows you to do is if
+
+00:52:19.520 --> 00:52:25.880
+you have a multi-layer
+
+00:52:21.200 --> 00:52:27.720
+Transformer it allows you to attend far
+
+00:52:25.880 --> 00:52:30.520
+back so if you look at the last layer
+
+00:52:27.720 --> 00:52:33.520
+it's attending um to things in the
+
+00:52:30.520 --> 00:52:36.599
+previous context window but the second
+
+00:52:33.520 --> 00:52:39.760
+to last layer is attending to things in
+
+00:52:36.599 --> 00:52:41.520
+the um not just one context window
+
+00:52:39.760 --> 00:52:44.079
+before but multiple context windows
+
+00:52:41.520 --> 00:52:45.760
+before and actually this allows you to
+
+00:52:44.079 --> 00:52:47.880
+very effectively attend a very long
+
+00:52:45.760 --> 00:52:51.720
+context because each time kind of the
+
+00:52:47.880 --> 00:52:54.799
+context expands in an exponential
+
+00:52:51.720 --> 00:52:56.520
+manner so um recently there's a popular
+
+00:52:54.799 --> 00:52:57.799
+model called mistol that I'm sure a lot
+
+00:52:56.520 --> 00:52:59.480
+of people have heard about and this is
+
+00:52:57.799 --> 00:53:01.920
+using sliding window attention which is
+
+00:52:59.480 --> 00:53:04.160
+essentially the same mechanism proposed
+
+00:53:01.920 --> 00:53:09.240
+by Transformer XEL so this method is
+
+00:53:04.160 --> 00:53:09.240
+still uh used in uh very practical
+
+00:53:10.400 --> 00:53:17.359
+systems another paper that has been
+
+00:53:13.440 --> 00:53:19.319
+pretty influential in this general area
+
+00:53:17.359 --> 00:53:21.079
+is something called sparse
+
+00:53:19.319 --> 00:53:23.359
+Transformers and the way sparse
+
+00:53:21.079 --> 00:53:25.960
+Transformers work is instead of
+
+00:53:23.359 --> 00:53:29.520
+attending to every single previous state
+
+00:53:25.960 --> 00:53:32.640
+you attend to every n previous
+
+00:53:29.520 --> 00:53:34.599
+States and what this allows you to do is
+
+00:53:32.640 --> 00:53:37.119
+this allows you to essentially create
+
+00:53:34.599 --> 00:53:40.319
+something like the strided uh
+
+00:53:37.119 --> 00:53:42.079
+convolutions or um pyramidal recurrent
+
+00:53:40.319 --> 00:53:45.520
+neural networks that I talked about
+
+00:53:42.079 --> 00:53:49.760
+earlier um so what this looks like
+
+00:53:45.520 --> 00:53:51.079
+essentially is you have um this like if
+
+00:53:49.760 --> 00:53:54.880
+you have a particular state it might
+
+00:53:51.079 --> 00:53:56.480
+attend to all of the previous end tokens
+
+00:53:54.880 --> 00:54:00.240
+but then it
+
+00:53:56.480 --> 00:54:04.400
+also attends to all of the
+
+00:54:00.240 --> 00:54:06.880
+previous um kind of M chunks so you kind
+
+00:54:04.400 --> 00:54:08.920
+of have a combination of local and
+
+00:54:06.880 --> 00:54:11.640
+Global
+
+00:54:08.920 --> 00:54:14.760
+attention or not local and Global but
+
+00:54:11.640 --> 00:54:16.760
+local and kind of longer range attention
+
+00:54:14.760 --> 00:54:18.760
+and this can be very effective because
+
+00:54:16.760 --> 00:54:22.319
+you can attend to you know much longer
+
+00:54:18.760 --> 00:54:24.079
+context with a minimal increase in a
+
+00:54:22.319 --> 00:54:26.520
+computational
+
+00:54:24.079 --> 00:54:28.720
+complexity
+
+00:54:26.520 --> 00:54:31.160
+so another method that's a little bit
+
+00:54:28.720 --> 00:54:32.960
+like this uh or it's very similar in
+
+00:54:31.160 --> 00:54:34.359
+spirit but slightly different in
+
+00:54:32.960 --> 00:54:35.599
+implementation is something called the
+
+00:54:34.359 --> 00:54:37.520
+compressive
+
+00:54:35.599 --> 00:54:40.400
+Transformer and in the compressive
+
+00:54:37.520 --> 00:54:43.000
+Transformer you also have this idea of a
+
+00:54:40.400 --> 00:54:44.319
+local memory and then a longer term
+
+00:54:43.000 --> 00:54:47.200
+compressed
+
+00:54:44.319 --> 00:54:50.799
+memory but you have an explicit
+
+00:54:47.200 --> 00:54:54.319
+compression step that
+
+00:54:50.799 --> 00:54:58.079
+directly essentially generates this uh
+
+00:54:54.319 --> 00:55:00.960
+compressed mem M itself and so this is a
+
+00:54:58.079 --> 00:55:04.119
+little bit more flexible I guess it
+
+00:55:00.960 --> 00:55:06.280
+allows you to take all of the you know
+
+00:55:04.119 --> 00:55:09.000
+relevant things from your local memory
+
+00:55:06.280 --> 00:55:12.000
+and compress it down so it's another
+
+00:55:09.000 --> 00:55:12.000
+method that's worth thinking
+
+00:55:12.760 --> 00:55:18.400
+about finally uh there are some very
+
+00:55:15.799 --> 00:55:20.200
+interesting methods that do low rank
+
+00:55:18.400 --> 00:55:23.039
+approximations for
+
+00:55:20.200 --> 00:55:25.920
+Transformers and so calculating the
+
+00:55:23.039 --> 00:55:29.119
+attention Matrix is expensive but this
+
+00:55:25.920 --> 00:55:31.640
+is a matrix and because it's a matrix we
+
+00:55:29.119 --> 00:55:32.640
+can also approximate it with a lower
+
+00:55:31.640 --> 00:55:35.480
+rank
+
+00:55:32.640 --> 00:55:38.559
+Matrix and there's a couple methods that
+
+00:55:35.480 --> 00:55:40.599
+do things uh like this uh the first one
+
+00:55:38.559 --> 00:55:42.680
+is something called Blind forer which
+
+00:55:40.599 --> 00:55:44.520
+adds low rank linear projections into
+
+00:55:42.680 --> 00:55:47.319
+the model at appropriate
+
+00:55:44.520 --> 00:55:50.359
+places and um there's another one called
+
+00:55:47.319 --> 00:55:52.200
+NR forer which approximates using the ni
+
+00:55:50.359 --> 00:55:54.440
+run method which is based on sampling
+
+00:55:52.200 --> 00:55:56.520
+Landmark points but basically the
+
+00:55:54.440 --> 00:56:00.319
+general IDE aide behind this is normally
+
+00:55:56.520 --> 00:56:03.400
+we do this kind of softmax over you know
+
+00:56:00.319 --> 00:56:06.240
+a very large attention Vector but
+
+00:56:03.400 --> 00:56:08.440
+instead we can approximate the softmax
+
+00:56:06.240 --> 00:56:11.520
+by having some low rank vectors kind of
+
+00:56:08.440 --> 00:56:12.799
+like what we used in Laura and uh
+
+00:56:11.520 --> 00:56:16.440
+nonetheless get a reasonable
+
+00:56:12.799 --> 00:56:16.440
+approximation of the softmax used
+
+00:56:17.799 --> 00:56:24.039
+inion okay so we're nearing the end of
+
+00:56:21.520 --> 00:56:26.000
+what I want to talk about today and
+
+00:56:24.039 --> 00:56:29.720
+finally the thing that I'd like to talk
+
+00:56:26.000 --> 00:56:33.240
+about is benchmarks for long PEX models
+
+00:56:29.720 --> 00:56:35.000
+and there's a few benchmarks one very
+
+00:56:33.240 --> 00:56:37.359
+well-known one is something called long
+
+00:56:35.000 --> 00:56:40.599
+range Arena this is a composite
+
+00:56:37.359 --> 00:56:43.000
+Benchmark containing mostly non NLP
+
+00:56:40.599 --> 00:56:45.280
+tasks and it's definitely used for long
+
+00:56:43.000 --> 00:56:46.760
+sequence modeling but the results on the
+
+00:56:45.280 --> 00:56:49.400
+long range Arena actually tend to
+
+00:56:46.760 --> 00:56:51.599
+diverge uh somewhat from the results
+
+00:56:49.400 --> 00:56:54.440
+that you get for longdistance language
+
+00:56:51.599 --> 00:56:56.520
+modeling so in addition to this another
+
+00:56:54.440 --> 00:56:58.400
+benchmark that I uh personally like and
+
+00:56:56.520 --> 00:57:01.960
+have used a bit is something called
+
+00:56:58.400 --> 00:57:05.720
+Scrolls which uh combines together a
+
+00:57:01.960 --> 00:57:07.960
+whole bunch of kind of QA style or
+
+00:57:05.720 --> 00:57:10.920
+summarization style tasks that have very
+
+00:57:07.960 --> 00:57:13.280
+long contexts including over narratives
+
+00:57:10.920 --> 00:57:15.680
+or books or government reports or other
+
+00:57:13.280 --> 00:57:17.280
+things like that so you can also take a
+
+00:57:15.680 --> 00:57:20.680
+look at this if you're interested in
+
+00:57:17.280 --> 00:57:20.680
+kind of benchmarking longer range
+
+00:57:21.839 --> 00:57:28.280
+models okay the final thing I'd like to
+
+00:57:24.559 --> 00:57:30.280
+talk about is now that we have retriever
+
+00:57:28.280 --> 00:57:31.680
+models we have reader models we maybe
+
+00:57:30.280 --> 00:57:34.000
+even have reader models that can
+
+00:57:31.680 --> 00:57:35.520
+effectively use very long contexts like
+
+00:57:34.000 --> 00:57:37.880
+the ones that we retrieve over whole
+
+00:57:35.520 --> 00:57:39.240
+documents how do we effectively use them
+
+00:57:37.880 --> 00:57:43.640
+in our
+
+00:57:39.240 --> 00:57:46.680
+models so there was a very nice paper um
+
+00:57:43.640 --> 00:57:48.880
+by Nelson Leo at Stanford that about a
+
+00:57:46.680 --> 00:57:51.160
+phenomenon that was kinded lost in the
+
+00:57:48.880 --> 00:57:53.079
+middle and basically what it does is it
+
+00:57:51.160 --> 00:57:55.119
+demonstrates that many many different
+
+00:57:53.079 --> 00:57:57.720
+models including state-of-the-art model
+
+00:57:55.119 --> 00:58:00.799
+models pay less attention to things in
+
+00:57:57.720 --> 00:58:03.960
+the middle of long context windows and
+
+00:58:00.799 --> 00:58:06.760
+so if we have an answer and we put it in
+
+00:58:03.960 --> 00:58:09.200
+you know the first position in Doc in
+
+00:58:06.760 --> 00:58:12.280
+you know a concatenated context or the
+
+00:58:09.200 --> 00:58:13.799
+20th position in a concatenated context
+
+00:58:12.280 --> 00:58:15.240
+it tends to attend more to the ones at
+
+00:58:13.799 --> 00:58:18.359
+the beginning or the
+
+00:58:15.240 --> 00:58:19.480
+end in contrast the ones in the middle
+
+00:58:18.359 --> 00:58:22.760
+kind of get
+
+00:58:19.480 --> 00:58:26.680
+lost hence the name lost in the middle
+
+00:58:22.760 --> 00:58:29.520
+and the problem with this is you know if
+
+00:58:26.680 --> 00:58:32.480
+we are doing something like retrieval in
+
+00:58:29.520 --> 00:58:34.160
+Reading then that's maybe not such a
+
+00:58:32.480 --> 00:58:35.680
+huge problem because we could just put
+
+00:58:34.160 --> 00:58:37.680
+you know the highest scoring documents
+
+00:58:35.680 --> 00:58:39.920
+at the beginning that might even be more
+
+00:58:37.680 --> 00:58:42.440
+effective than uh you know concatenating
+
+00:58:39.920 --> 00:58:44.160
+lots of low scoring documents together
+
+00:58:42.440 --> 00:58:45.559
+but if we want to read a really long
+
+00:58:44.160 --> 00:58:48.839
+document and synthesize something
+
+00:58:45.559 --> 00:58:52.200
+without doing kind of another uh scoring
+
+00:58:48.839 --> 00:58:54.200
+step uh that can be an issue and also
+
+00:58:52.200 --> 00:58:56.359
+you know our retriever is not perfect so
+
+00:58:54.200 --> 00:58:58.799
+we would like the model to the reader
+
+00:58:56.359 --> 00:59:00.520
+model to do a good job with the outputs
+
+00:58:58.799 --> 00:59:04.839
+that it
+
+00:59:00.520 --> 00:59:06.359
+has so there are methods uh to ensure
+
+00:59:04.839 --> 00:59:09.440
+use of relevant
+
+00:59:06.359 --> 00:59:12.119
+context so of course better retrievers
+
+00:59:09.440 --> 00:59:14.880
+make more relevant context you can do
+
+00:59:12.119 --> 00:59:16.240
+you know reranking or other things like
+
+00:59:14.880 --> 00:59:17.280
+that and only include the context that
+
+00:59:16.240 --> 00:59:19.680
+looks most
+
+00:59:17.280 --> 00:59:22.880
+relevant um or you know refine your
+
+00:59:19.680 --> 00:59:25.200
+reader model but there's also methods
+
+00:59:22.880 --> 00:59:28.720
+that can decide whether contact should
+
+00:59:25.200 --> 00:59:32.400
+be used in the first place so um there
+
+00:59:28.720 --> 00:59:35.440
+are methods uh to decide whether to use
+
+00:59:32.400 --> 00:59:37.559
+whether to include passages or not and
+
+00:59:35.440 --> 00:59:39.920
+also uh recently we proposed a method to
+
+00:59:37.559 --> 00:59:42.640
+filter down to parts of retrieve
+
+00:59:39.920 --> 00:59:44.920
+passages uh to have only appropriate
+
+00:59:42.640 --> 00:59:47.480
+content and this is a model uh that we
+
+00:59:44.920 --> 00:59:49.319
+called filco it basically filters the
+
+00:59:47.480 --> 00:59:52.160
+context down to the most relevant
+
+00:59:49.319 --> 00:59:53.920
+content that we think is appropriate and
+
+00:59:52.160 --> 00:59:56.960
+that allows us to get better results
+
+00:59:53.920 --> 00:59:56.960
+when it's fed to the
+
+00:59:57.079 --> 01:00:03.640
+generator so that's all I have for today
+
+01:00:00.319 --> 01:00:06.200
+um thank you for watching the video and
+
+01:00:03.640 --> 01:00:08.599
+for people in the class I'll be happy to
+
+01:00:06.200 --> 01:00:13.079
+take questions on Piaza or during the
+
+01:00:08.599 --> 01:00:13.079
+office hours that I had planned thanks a
+
+01:00:15.319 --> 01:00:18.319
+lot

CMU Advanced NLP 2024 (11) Distillation, Quantization, and Pruning/CMU Advanced NLP 2024 (11) Distillation Quantization and Pruning.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a58b185362d8bffc64edfe4141f67f0804b6efa04d4bfbba63316ce1b5dd8fe
+size 65064579

CMU Advanced NLP 2024 (11) Distillation, Quantization, and Pruning/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=s9yyH3RPhdM",
+    "title": "CMU Advanced NLP 2024 (11) Distillation, Quantization, and Pruning"
+}

CMU Advanced NLP 2024 (12) Reinforcement Learning/CMU Advanced NLP 2024 (12) Reinforcement Learning.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1bb70e0fa6c406fd7c2d8d736e10c2e652b52b3e65757930cea4fb235a50ffb3
+size 72409479

CMU Advanced NLP 2024 (12) Reinforcement Learning/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=NX0l1M0NWPM",
+    "title": "CMU Advanced NLP 2024 (12) Reinforcement Learning"
+}

CMU Advanced NLP 2024 (13) Debugging and Interpretation/CMU Advanced NLP 2024 (13) Debugging and Interpretation.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7de94ea1f378595ae1d957526a4f6aa5a2c75c49db3b40a36e1f0e5ab2a17152
+size 82237142

CMU Advanced NLP 2024 (13) Debugging and Interpretation/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=c4UwOq2J9mQ",
+    "title": "CMU Advanced NLP 2024 (13) Debugging and Interpretation"
+}

CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts/CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22c0ccf96269123bc7e4f6390f0c0220a4e05848e941a58ed5e57085ae2d8432
+size 80561402

CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=MueCRSZ3RQ0",
+    "title": "CMU Advanced NLP 2024 (14) Ensembling and Mixture of Experts"
+}

CMU Advanced NLP 2024 (15) A Tour of Modern Large Language Models/CMU Advanced NLP 2024 (15) A Tour of Modern Large Language Models.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5c3f854f59933275f3273f0b77753dee254d48e166d37f6f3190a5423767201
+size 79708142

CMU Advanced NLP 2024 (15) A Tour of Modern Large Language Models/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=2rOSrDtg7HQ",
+    "title": "CMU Advanced NLP 2024 (15) A Tour of Modern Large Language Models"
+}

CMU Advanced NLP 2024 (17) Code Generation/CMU Advanced NLP 2024 (17) Code Generation.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7fcb735ceea4c24db426084df97f450a16142c64c4736ab6e403bb13741c8350
+size 63648833

CMU Advanced NLP 2024 (17) Code Generation/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=bN2ZZieBXsE",
+    "title": "CMU Advanced NLP 2024 (17) Code Generation"
+}

CMU Advanced NLP 2024 (18) Knowledge and Language Models/CMU Advanced NLP 2024 (18) Knowledge and Language Models.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b246f116f9c543f9cc995a334954a8947064a1bc3950d9acdc34b8bf42b8771
+size 74113017

CMU Advanced NLP 2024 (18) Knowledge and Language Models/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=IwEYCbdgJ9U",
+    "title": "CMU Advanced NLP 2024 (18) Knowledge and Language Models"
+}

CMU Advanced NLP 2024 (2) Word Representation and Text Classification/CMU Advanced NLP 2024 (2) Word Representation and Text Classification.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:35d3a9cb9cc7d1aedba6a0742b46cfd3e24c4999c46b59ae1df22321775c9102
+size 82455565

CMU Advanced NLP 2024 (2) Word Representation and Text Classification/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=wa61zdcKWyU",
+    "title": "CMU Advanced NLP 2024 (2) Word Representation and Text Classification"
+}

CMU Advanced NLP 2024 (20) Tool Use and Language Agents/CMU Advanced NLP 2024 (20) Tool Use and Language Agents.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:639a8f338d12c0f77948889dffe7bb1bbe4e9ad4cb6f4aa806babd15d679af88
+size 83218086

CMU Advanced NLP 2024 (20) Tool Use and Language Agents/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=d0QSnLjlgzc",
+    "title": "CMU Advanced NLP 2024 (20) Tool Use and Language Agents"
+}

CMU Advanced NLP 2024 (21) Complex Reasoning/CMU Advanced NLP 2024 (21) Complex Reasoning.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:10a2dabeb41186cd432caae205d3e22b8ad34e91253d174abfdadaa82ea581f2
+size 56293331

CMU Advanced NLP 2024 (21) Complex Reasoning/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=mPd2hFmzjWE",
+    "title": "CMU Advanced NLP 2024 (21) Complex Reasoning"
+}

CMU Advanced NLP 2024 (21) Complex Reasoning/transcript.srt

@@ -0,0 +1,5007 @@
+1
+00:00:00,280 --> 00:00:05,120
+so I'd like to go ahead with uh complex
+
+2
+00:00:02,399 --> 00:00:08,719
+reasoning and we've talked a little bit
+
+3
+00:00:05,120 --> 00:00:10,719
+about uh reasoning in language models uh
+
+4
+00:00:08,719 --> 00:00:12,160
+up until now and so I'm going to be
+
+5
+00:00:10,719 --> 00:00:15,280
+talking about stuff that we didn't talk
+
+6
+00:00:12,160 --> 00:00:17,240
+about yet um this might be a little bit
+
+7
+00:00:15,280 --> 00:00:19,199
+short because of that because I'm not
+
+8
+00:00:17,240 --> 00:00:20,640
+talking about like programs because we
+
+9
+00:00:19,199 --> 00:00:22,080
+talked about that in the code generation
+
+10
+00:00:20,640 --> 00:00:24,199
+class and we already talked a little bit
+
+11
+00:00:22,080 --> 00:00:26,320
+about some of the basics here but um you
+
+12
+00:00:24,199 --> 00:00:30,119
+know if we have time at the end I'd be
+
+13
+00:00:26,320 --> 00:00:30,840
+happy to discuss free form also so what
+
+14
+00:00:30,119 --> 00:00:34,320
+is
+
+15
+00:00:30,840 --> 00:00:35,920
+reasoning um the basic idea is using
+
+16
+00:00:34,320 --> 00:00:37,680
+evidence and logic to arrive at
+
+17
+00:00:35,920 --> 00:00:40,200
+conclusions and make
+
+18
+00:00:37,680 --> 00:00:43,760
+judgments
+
+19
+00:00:40,200 --> 00:00:48,039
+and what is it in language models is a
+
+20
+00:00:43,760 --> 00:00:49,399
+little bit um you know less clear uh but
+
+21
+00:00:48,039 --> 00:00:52,680
+if we talk about it kind of like from
+
+22
+00:00:49,399 --> 00:00:56,280
+the philosophical standpoint um there
+
+23
+00:00:52,680 --> 00:00:58,399
+are two varieties of this one is formal
+
+24
+00:00:56,280 --> 00:01:01,680
+uh reasoning and formal reasoning is
+
+25
+00:00:58,399 --> 00:01:04,239
+mostly based on strict truth values so
+
+26
+00:01:01,680 --> 00:01:05,920
+it's kind of like um you can definitely
+
+27
+00:01:04,239 --> 00:01:08,360
+say this is true you can definitely say
+
+28
+00:01:05,920 --> 00:01:11,680
+this is not true
+
+29
+00:01:08,360 --> 00:01:13,799
+and in real life there's very little
+
+30
+00:01:11,680 --> 00:01:15,759
+actual formal reasoning outside of like
+
+31
+00:01:13,799 --> 00:01:17,960
+for example mathematics or maybe you
+
+32
+00:01:15,759 --> 00:01:20,240
+know algorithms computer science and
+
+33
+00:01:17,960 --> 00:01:21,759
+other things like that um and then
+
+34
+00:01:20,240 --> 00:01:23,240
+separately from that we have informal
+
+35
+00:01:21,759 --> 00:01:27,040
+reasoning based on experience and
+
+36
+00:01:23,240 --> 00:01:30,439
+intuition and actually um this is this
+
+37
+00:01:27,040 --> 00:01:32,360
+was uh rather elusive uh until
+
+38
+00:01:30,439 --> 00:01:33,720
+large language models you know people
+
+39
+00:01:32,360 --> 00:01:35,560
+were working on it but it was really
+
+40
+00:01:33,720 --> 00:01:38,119
+hard and this is like one of the big
+
+41
+00:01:35,560 --> 00:01:41,479
+breakthroughs I think of the past few
+
+42
+00:01:38,119 --> 00:01:46,799
+years um I should note that this uh
+
+43
+00:01:41,479 --> 00:01:48,520
+paper here uh hang and Chan is a kind of
+
+44
+00:01:46,799 --> 00:01:50,119
+review survey paper of reasoning in
+
+45
+00:01:48,520 --> 00:01:51,520
+large language models it's on the
+
+46
+00:01:50,119 --> 00:01:54,719
+references so if you're interested you
+
+47
+00:01:51,520 --> 00:01:57,600
+can take a look at that too um but
+
+48
+00:01:54,719 --> 00:01:59,200
+there's three kinds of reasoning uh
+
+49
+00:01:57,600 --> 00:02:00,840
+there's many kinds of reasoning but
+
+50
+00:01:59,200 --> 00:02:03,280
+there's three kinds of reasoning in
+
+51
+00:02:00,840 --> 00:02:06,240
+particular that I'd like to talk about
+
+52
+00:02:03,280 --> 00:02:08,840
+um from the point of view of today and
+
+53
+00:02:06,240 --> 00:02:10,360
+the first one is uh deductive reasoning
+
+54
+00:02:08,840 --> 00:02:13,080
+and deductive reasoning is basically
+
+55
+00:02:10,360 --> 00:02:16,040
+using logic to go from a premise to a
+
+56
+00:02:13,080 --> 00:02:18,440
+conclusion and this is largely what
+
+57
+00:02:16,040 --> 00:02:19,879
+people not entirely but largely what
+
+58
+00:02:18,440 --> 00:02:22,400
+people talk about when they think about
+
+59
+00:02:19,879 --> 00:02:25,879
+formal reasoning and so basically you
+
+60
+00:02:22,400 --> 00:02:28,640
+have several premises um like all
+
+61
+00:02:25,879 --> 00:02:32,120
+mammals have kidneys and all whales are
+
+62
+00:02:28,640 --> 00:02:35,239
+mammals and then from this uh you can go
+
+63
+00:02:32,120 --> 00:02:35,239
+to all whales have
+
+64
+00:02:35,440 --> 00:02:40,640
+kidneys then separately there's
+
+65
+00:02:38,000 --> 00:02:44,040
+inductive reasoning and inductive
+
+66
+00:02:40,640 --> 00:02:46,040
+reasoning is um from
+
+67
+00:02:44,040 --> 00:02:48,480
+observation uh predict a likely
+
+68
+00:02:46,040 --> 00:02:50,080
+conclusion or predict a likely kind of
+
+69
+00:02:48,480 --> 00:02:53,640
+generalized
+
+70
+00:02:50,080 --> 00:02:55,360
+conclusion um so this is one example uh
+
+71
+00:02:53,640 --> 00:02:56,920
+when we see a creature with wings it is
+
+72
+00:02:55,360 --> 00:02:58,599
+usually a bird we see a creature with
+
+73
+00:02:56,920 --> 00:03:00,400
+wings the creature is likely to be a
+
+74
+00:02:58,599 --> 00:03:02,879
+bird so it's kind of this is kind of
+
+75
+00:03:00,400 --> 00:03:05,319
+like a soft version of deduction another
+
+76
+00:03:02,879 --> 00:03:07,440
+common thing is like every single
+
+77
+00:03:05,319 --> 00:03:10,760
+creature I have seen with wings is a
+
+78
+00:03:07,440 --> 00:03:12,480
+bird and then you can kind of um induce
+
+79
+00:03:10,760 --> 00:03:16,799
+that all
+
+80
+00:03:12,480 --> 00:03:19,159
+uh like all uh creatures with wings are
+
+81
+00:03:16,799 --> 00:03:21,120
+birds but that might not be true it's
+
+82
+00:03:19,159 --> 00:03:23,879
+not necessarily logically entailed but
+
+83
+00:03:21,120 --> 00:03:27,560
+you you make that kind
+
+84
+00:03:23,879 --> 00:03:31,000
+of logical conclusion uh without it
+
+85
+00:03:27,560 --> 00:03:32,840
+being formally uh correct or verif
+
+86
+00:03:31,000 --> 00:03:34,720
+and then the final one is abductive
+
+87
+00:03:32,840 --> 00:03:38,000
+reasoning and so this is from an
+
+88
+00:03:34,720 --> 00:03:40,760
+observation we predict the most likely
+
+89
+00:03:38,000 --> 00:03:42,760
+explanation and so for example if we
+
+90
+00:03:40,760 --> 00:03:44,480
+have something like the car cannot start
+
+91
+00:03:42,760 --> 00:03:48,319
+and there is a puddle of liquid under
+
+92
+00:03:44,480 --> 00:03:50,200
+the engine um then we might have a
+
+93
+00:03:48,319 --> 00:03:53,360
+likely explanation that the car has a
+
+94
+00:03:50,200 --> 00:03:55,280
+leak in the radiator so we're going from
+
+95
+00:03:53,360 --> 00:03:58,760
+kind of uh the
+
+96
+00:03:55,280 --> 00:04:00,879
+car you know these these things and then
+
+97
+00:03:58,760 --> 00:04:02,280
+we try to predict the reason why this
+
+98
+00:04:00,879 --> 00:04:05,040
+happens so we're trying to predict like
+
+99
+00:04:02,280 --> 00:04:07,360
+reverse pausal links
+
+100
+00:04:05,040 --> 00:04:08,480
+essentially um there's other types of re
+
+101
+00:04:07,360 --> 00:04:10,400
+reasoning that I'm not going to talk
+
+102
+00:04:08,480 --> 00:04:12,159
+about as much like analogical reasoning
+
+103
+00:04:10,400 --> 00:04:14,079
+and and things like this but uh these
+
+104
+00:04:12,159 --> 00:04:15,440
+are the three main ones I want to talk
+
+105
+00:04:14,079 --> 00:04:17,720
+about
+
+106
+00:04:15,440 --> 00:04:22,040
+today uh one thing I should point out is
+
+107
+00:04:17,720 --> 00:04:24,400
+like even in philosophy or you know
+
+108
+00:04:22,040 --> 00:04:26,240
+like even when you read descriptions
+
+109
+00:04:24,400 --> 00:04:29,280
+about these various types of reasoning
+
+110
+00:04:26,240 --> 00:04:31,880
+the types are a little bit vague so um
+
+111
+00:04:29,280 --> 00:04:35,280
+take these is like
+
+112
+00:04:31,880 --> 00:04:37,240
+general not you know General directions
+
+113
+00:04:35,280 --> 00:04:39,400
+and not strict rules because like which
+
+114
+00:04:37,240 --> 00:04:42,120
+falls on under which category also can
+
+115
+00:04:39,400 --> 00:04:44,880
+be a little bit uh you know unclear uh
+
+116
+00:04:42,120 --> 00:04:44,880
+according to various
+
+117
+00:04:45,479 --> 00:04:53,440
+definitions cool um so first before
+
+118
+00:04:49,840 --> 00:04:55,720
+getting into formal reasoning methods
+
+119
+00:04:53,440 --> 00:04:57,759
+are before getting into the bulk of the
+
+120
+00:04:55,720 --> 00:05:00,000
+talk which is going to be about llms I
+
+121
+00:04:57,759 --> 00:05:02,479
+want to talk about some pre-m reasoning
+
+122
+00:05:00,000 --> 00:05:03,720
+methods and the first one is kind of
+
+123
+00:05:02,479 --> 00:05:05,160
+like formal reasoning within
+
+124
+00:05:03,720 --> 00:05:07,320
+computational
+
+125
+00:05:05,160 --> 00:05:09,840
+semantics and this has been around for a
+
+126
+00:05:07,320 --> 00:05:12,479
+really long time um it's also kind of
+
+127
+00:05:09,840 --> 00:05:15,000
+what powered the things that worked over
+
+128
+00:05:12,479 --> 00:05:21,039
+knowledge bases and other things like
+
+129
+00:05:15,000 --> 00:05:23,639
+this um and the way it works is it does
+
+130
+00:05:21,039 --> 00:05:27,600
+derivational um
+
+131
+00:05:23,639 --> 00:05:31,800
+reasoning by uh sorry I can't read that
+
+132
+00:05:27,600 --> 00:05:34,720
+in the back um by starting out with
+
+133
+00:05:31,800 --> 00:05:36,080
+certain premises and getting to um
+
+134
+00:05:34,720 --> 00:05:40,000
+getting to final
+
+135
+00:05:36,080 --> 00:05:43,039
+conclusions so there's ways that you can
+
+136
+00:05:40,000 --> 00:05:44,060
+write this I think you might have
+
+137
+00:05:43,039 --> 00:05:47,080
+seen
+
+138
+00:05:44,060 --> 00:05:50,479
+[Music]
+
+139
+00:05:47,080 --> 00:05:54,240
+um you might have seen
+
+140
+00:05:50,479 --> 00:05:58,319
+uh this in uh another like math class or
+
+141
+00:05:54,240 --> 00:06:02,440
+something but uh we we have symbols like
+
+142
+00:05:58,319 --> 00:06:02,440
+all and um
+
+143
+00:06:03,039 --> 00:06:08,280
+exist let's
+
+144
+00:06:04,960 --> 00:06:10,960
+see yeah we have things like all and
+
+145
+00:06:08,280 --> 00:06:13,319
+exist and like all
+
+146
+00:06:10,960 --> 00:06:16,240
+X
+
+147
+00:06:13,319 --> 00:06:20,479
+die means
+
+148
+00:06:16,240 --> 00:06:23,919
+like every Everything has died and this
+
+149
+00:06:20,479 --> 00:06:27,360
+uh implies that Mia and Zed have
+
+150
+00:06:23,919 --> 00:06:30,440
+died um
+
+151
+00:06:27,360 --> 00:06:32,240
+so yeah this is a actually maybe I'll
+
+152
+00:06:30,440 --> 00:06:33,280
+not I'll not go through this one and let
+
+153
+00:06:32,240 --> 00:06:37,639
+me go
+
+154
+00:06:33,280 --> 00:06:40,440
+through um go to this one so like it
+
+155
+00:06:37,639 --> 00:06:40,440
+would be something
+
+156
+00:06:40,639 --> 00:06:45,080
+like uh for
+
+157
+00:06:42,960 --> 00:06:47,480
+all
+
+158
+00:06:45,080 --> 00:06:50,669
+X um
+
+159
+00:06:47,480 --> 00:06:50,669
+[Music]
+
+160
+00:06:52,039 --> 00:07:00,400
+mamal well X
+
+161
+00:06:56,759 --> 00:07:03,520
+implies have
+
+162
+00:07:00,400 --> 00:07:07,560
+X kidney or something like
+
+163
+00:07:03,520 --> 00:07:09,280
+that and then you would have other rules
+
+164
+00:07:07,560 --> 00:07:11,879
+and you can go through uh through
+
+165
+00:07:09,280 --> 00:07:14,440
+derivations and and other things like
+
+166
+00:07:11,879 --> 00:07:16,120
+this
+
+167
+00:07:14,440 --> 00:07:19,280
+um
+
+168
+00:07:16,120 --> 00:07:21,560
+my favorite reference for this is this
+
+169
+00:07:19,280 --> 00:07:24,599
+Blackburn and buz book right here it's
+
+170
+00:07:21,560 --> 00:07:26,400
+really well written um and it has like
+
+171
+00:07:24,599 --> 00:07:28,039
+lots of good examples it also explains
+
+172
+00:07:26,400 --> 00:07:30,440
+how you go through derivations and other
+
+173
+00:07:28,039 --> 00:07:34,360
+stuff like that
+
+174
+00:07:30,440 --> 00:07:35,759
+um and actually neural networks can do
+
+175
+00:07:34,360 --> 00:07:37,039
+this variety of reasoning through Chain
+
+176
+00:07:35,759 --> 00:07:38,599
+of Thought and other things I'm going to
+
+177
+00:07:37,039 --> 00:07:40,120
+talk about today but it's a very rough
+
+178
+00:07:38,599 --> 00:07:43,960
+approximation and it doesn't work
+
+179
+00:07:40,120 --> 00:07:47,199
+particularly well for saying like all
+
+180
+00:07:43,960 --> 00:07:51,240
+you know all people
+
+181
+00:07:47,199 --> 00:07:53,599
+are of a uh like things that apply to
+
+182
+00:07:51,240 --> 00:07:57,240
+all people or things that apply to sets
+
+183
+00:07:53,599 --> 00:08:00,039
+or other things like this so within
+
+184
+00:07:57,240 --> 00:08:02,879
+prologue you could
+
+185
+00:08:00,039 --> 00:08:06,520
+take a knowledge base and ask the
+
+186
+00:08:02,879 --> 00:08:11,960
+knowledge base like do
+
+187
+00:08:06,520 --> 00:08:12,800
+all people who work at CMU as professors
+
+188
+00:08:11,960 --> 00:08:15,840
+have a
+
+189
+00:08:12,800 --> 00:08:18,080
+PhD and you could like actually examine
+
+190
+00:08:15,840 --> 00:08:20,639
+that based on the knowledge base uh
+
+191
+00:08:18,080 --> 00:08:23,520
+whereas even if you had
+
+192
+00:08:20,639 --> 00:08:25,800
+a language model that had access to
+
+193
+00:08:23,520 --> 00:08:27,280
+everybody's CVS it wouldn't necessarily
+
+194
+00:08:25,800 --> 00:08:28,599
+be able to answer that question and it
+
+195
+00:08:27,280 --> 00:08:31,440
+especially wouldn't be able to answer
+
+196
+00:08:28,599 --> 00:08:31,440
+that question if you were
+
+197
+00:08:32,320 --> 00:08:37,760
+um it wouldn't be able to answer that
+
+198
+00:08:34,640 --> 00:08:42,880
+question if there were like multiple
+
+199
+00:08:37,760 --> 00:08:46,480
+steps so did all people who are working
+
+200
+00:08:42,880 --> 00:08:50,959
+at CMU get their PHD after
+
+201
+00:08:46,480 --> 00:08:52,959
+19 90 or something like that um so and
+
+202
+00:08:50,959 --> 00:08:54,680
+the answer to that is obviously no but
+
+203
+00:08:52,959 --> 00:08:56,519
+uh this would be able to find the
+
+204
+00:08:54,680 --> 00:08:58,120
+counter evidence to that whereas LMS
+
+205
+00:08:56,519 --> 00:09:00,000
+would not be guaranteed to be able to do
+
+206
+00:08:58,120 --> 00:09:02,800
+that
+
+207
+00:09:00,000 --> 00:09:04,279
+so I I think this is really uh like a
+
+208
+00:09:02,800 --> 00:09:06,760
+nice thing to know but there's a couple
+
+209
+00:09:04,279 --> 00:09:09,600
+problems with it the first thing is this
+
+210
+00:09:06,760 --> 00:09:12,519
+really only traffics in like strictly
+
+211
+00:09:09,600 --> 00:09:17,880
+true or strictly false statements um and
+
+212
+00:09:12,519 --> 00:09:20,560
+that's a really big issue um so like if
+
+213
+00:09:17,880 --> 00:09:22,959
+anything's soft you start uh this sort
+
+214
+00:09:20,560 --> 00:09:24,320
+of formal reasoning starts breaking down
+
+215
+00:09:22,959 --> 00:09:25,880
+the second thing which actually is a
+
+216
+00:09:24,320 --> 00:09:28,959
+really big problem is once you start
+
+217
+00:09:25,880 --> 00:09:30,600
+dealing with more complex things you
+
+218
+00:09:28,959 --> 00:09:32,560
+don't ize it but there's always like
+
+219
+00:09:30,600 --> 00:09:34,560
+exceptions and exceptions to exceptions
+
+220
+00:09:32,560 --> 00:09:36,240
+and other things like that and actually
+
+221
+00:09:34,560 --> 00:09:38,320
+becomes very computationally expensive
+
+222
+00:09:36,240 --> 00:09:41,640
+to prove anything that's kind of like
+
+223
+00:09:38,320 --> 00:09:43,279
+non-trivial um and so because of that uh
+
+224
+00:09:41,640 --> 00:09:45,839
+I'm not actually going to cover it in
+
+225
+00:09:43,279 --> 00:09:47,880
+the lecture today but recently there are
+
+226
+00:09:45,839 --> 00:09:50,880
+um kind of search algorithms through
+
+227
+00:09:47,880 --> 00:09:54,279
+proof spaces that use uh like neural
+
+228
+00:09:50,880 --> 00:09:55,880
+models to do to speed up the search by
+
+229
+00:09:54,279 --> 00:09:58,120
+picking the best and most promising
+
+230
+00:09:55,880 --> 00:10:00,800
+hypotheses and uh for example Sean
+
+231
+00:09:58,120 --> 00:10:02,800
+wellik uh here at CMU is working on that
+
+232
+00:10:00,800 --> 00:10:04,800
+for neural theorem proving where you
+
+233
+00:10:02,800 --> 00:10:05,959
+have uh like mathematical theorem
+
+234
+00:10:04,800 --> 00:10:08,079
+proving and then you use a neural
+
+235
+00:10:05,959 --> 00:10:13,120
+network to pick the best uh paths
+
+236
+00:10:08,079 --> 00:10:14,880
+through logical uh operations so um
+
+237
+00:10:13,120 --> 00:10:19,279
+that's kind of a combination of the more
+
+238
+00:10:14,880 --> 00:10:22,920
+classical and uh modern
+
+239
+00:10:19,279 --> 00:10:26,240
+methods then another thing that's useful
+
+240
+00:10:22,920 --> 00:10:28,079
+to talk about I think this isn't very
+
+241
+00:10:26,240 --> 00:10:31,640
+popular right now but I think it might
+
+242
+00:10:28,079 --> 00:10:34,360
+be become more popular uh in the future
+
+243
+00:10:31,640 --> 00:10:36,120
+is we start hitting the limits of uh you
+
+244
+00:10:34,360 --> 00:10:38,560
+know what we can fit into long context
+
+245
+00:10:36,120 --> 00:10:40,040
+Windows uh for neural models and stuff
+
+246
+00:10:38,560 --> 00:10:42,600
+like this is memory
+
+247
+00:10:40,040 --> 00:10:48,600
+networks and basically the way that
+
+248
+00:10:42,600 --> 00:10:50,639
+memory networks work is they have write
+
+249
+00:10:48,600 --> 00:10:51,399
+they have the ability to write and read
+
+250
+00:10:50,639 --> 00:10:55,639
+from
+
+251
+00:10:51,399 --> 00:10:57,360
+memory and so this figure is a little
+
+252
+00:10:55,639 --> 00:11:00,440
+bit complex here but
+
+253
+00:10:57,360 --> 00:11:02,880
+basically you have a query and then you
+
+254
+00:11:00,440 --> 00:11:04,560
+get the embedding of the query um you
+
+255
+00:11:02,880 --> 00:11:06,760
+take the inner product you get the soft
+
+256
+00:11:04,560 --> 00:11:09,720
+Max of the inner product so this looks
+
+257
+00:11:06,760 --> 00:11:11,040
+like a tension you look up embeddings
+
+258
+00:11:09,720 --> 00:11:12,839
+and you take the weighted Su of the
+
+259
+00:11:11,040 --> 00:11:14,560
+embeddings and you get the like summary
+
+260
+00:11:12,839 --> 00:11:17,680
+of the memor so this is basically
+
+261
+00:11:14,560 --> 00:11:20,320
+attention over a big memory
+
+262
+00:11:17,680 --> 00:11:22,120
+base but then uh memory networks also
+
+263
+00:11:20,320 --> 00:11:24,000
+have the ability to go in and update the
+
+264
+00:11:22,120 --> 00:11:26,639
+memory so they also H have write
+
+265
+00:11:24,000 --> 00:11:30,360
+operations so you can read and write
+
+266
+00:11:26,639 --> 00:11:34,320
+from uh from the memory
+
+267
+00:11:30,360 --> 00:11:36,279
+base and so the reason why I say this
+
+268
+00:11:34,320 --> 00:11:40,480
+might become more popular is like one of
+
+269
+00:11:36,279 --> 00:11:42,200
+the big issues with large language
+
+270
+00:11:40,480 --> 00:11:45,320
+models nowadays is they don't get like
+
+271
+00:11:42,200 --> 00:11:47,320
+to continually update their memory um
+
+272
+00:11:45,320 --> 00:11:50,279
+and like one way you can do that is you
+
+273
+00:11:47,320 --> 00:11:52,160
+can just add text to the memory but
+
+274
+00:11:50,279 --> 00:11:54,000
+there are certain limits to that right
+
+275
+00:11:52,160 --> 00:11:56,360
+uh you know text isn't necessarily the
+
+276
+00:11:54,000 --> 00:11:58,959
+best way to encode all of the things
+
+277
+00:11:56,360 --> 00:12:01,880
+that you've seen in the past so I I feel
+
+278
+00:11:58,959 --> 00:12:03,360
+like this kind of architecture might be
+
+279
+00:12:01,880 --> 00:12:04,920
+um how to pin these sorts of
+
+280
+00:12:03,360 --> 00:12:06,480
+architectures onto language models might
+
+281
+00:12:04,920 --> 00:12:08,639
+be an interesting research direction for
+
+282
+00:12:06,480 --> 00:12:08,639
+the
+
+283
+00:12:08,680 --> 00:12:15,360
+future um another thing which I am not
+
+284
+00:12:12,600 --> 00:12:16,720
+going to talk about very much uh but
+
+285
+00:12:15,360 --> 00:12:20,560
+because we kind of already talked about
+
+286
+00:12:16,720 --> 00:12:23,560
+it in the code Generation Um area but
+
+287
+00:12:20,560 --> 00:12:26,959
+it's actually been around for a while is
+
+288
+00:12:23,560 --> 00:12:30,600
+solving questions with sort of symbolic
+
+289
+00:12:26,959 --> 00:12:36,480
+reasoning and the way it works
+
+290
+00:12:30,600 --> 00:12:41,320
+is for example you would have a
+
+291
+00:12:36,480 --> 00:12:43,639
+um you would have a text here and based
+
+292
+00:12:41,320 --> 00:12:47,440
+on the text you can run these sort of
+
+293
+00:12:43,639 --> 00:12:50,440
+symbolic operations like find and filter
+
+294
+00:12:47,440 --> 00:12:52,720
+and find the max number and relocate and
+
+295
+00:12:50,440 --> 00:12:54,480
+other things like this and this
+
+296
+00:12:52,720 --> 00:12:58,040
+explicitly
+
+297
+00:12:54,480 --> 00:12:59,880
+manipulates uh kind of the attention and
+
+298
+00:12:58,040 --> 00:13:02,519
+the um
+
+299
+00:12:59,880 --> 00:13:03,839
+you can do things like filtering down to
+
+300
+00:13:02,519 --> 00:13:08,600
+find the
+
+301
+00:13:03,839 --> 00:13:11,040
+most uh like highest largest number for
+
+302
+00:13:08,600 --> 00:13:12,800
+example or other things like this and
+
+303
+00:13:11,040 --> 00:13:14,160
+this is kind of interesting because like
+
+304
+00:13:12,800 --> 00:13:17,240
+some of the things that neural networks
+
+305
+00:13:14,160 --> 00:13:20,360
+are bad at are like finding the largest
+
+306
+00:13:17,240 --> 00:13:21,600
+number in a big data set or um finding
+
+307
+00:13:20,360 --> 00:13:23,360
+all of the things where something
+
+308
+00:13:21,600 --> 00:13:26,240
+applies and throwing out all of the
+
+309
+00:13:23,360 --> 00:13:27,959
+things where something doesn't apply so
+
+310
+00:13:26,240 --> 00:13:29,560
+again this isn't used super widely in
+
+311
+00:13:27,959 --> 00:13:31,959
+larged language models right now because
+
+312
+00:13:29,560 --> 00:13:33,920
+I feel like um people have been focusing
+
+313
+00:13:31,959 --> 00:13:36,440
+on prompting
+
+314
+00:13:33,920 --> 00:13:38,880
+techniques uh in order to do this sort
+
+315
+00:13:36,440 --> 00:13:41,199
+of reasoning but I think this is another
+
+316
+00:13:38,880 --> 00:13:43,320
+thing that's worth thinking about taking
+
+317
+00:13:41,199 --> 00:13:45,079
+a close another look at and seeing if
+
+318
+00:13:43,320 --> 00:13:47,440
+there are ways to incorporate it with
+
+319
+00:13:45,079 --> 00:13:49,320
+the current models because like
+
+320
+00:13:47,440 --> 00:13:50,720
+basically what I wanted to say is like
+
+321
+00:13:49,320 --> 00:13:52,279
+all of the things that I decided to
+
+322
+00:13:50,720 --> 00:13:54,560
+introduce here in this section are
+
+323
+00:13:52,279 --> 00:13:57,600
+things that current models are still not
+
+324
+00:13:54,560 --> 00:14:00,800
+particularly good at like Reon taking
+
+325
+00:13:57,600 --> 00:14:03,079
+many steps over sets of
+
+326
+00:14:00,800 --> 00:14:05,079
+inputs um reading and writing from
+
+327
+00:14:03,079 --> 00:14:09,839
+memory so that you can remember things
+
+328
+00:14:05,079 --> 00:14:11,720
+over long periods and also um filtering
+
+329
+00:14:09,839 --> 00:14:13,399
+down large pieces of text into smaller
+
+330
+00:14:11,720 --> 00:14:16,040
+pieces of text to find relevant
+
+331
+00:14:13,399 --> 00:14:17,560
+information so um if any of those things
+
+332
+00:14:16,040 --> 00:14:19,880
+sound interesting you can take a look at
+
+333
+00:14:17,560 --> 00:14:22,800
+this but um after this I'd like to go
+
+334
+00:14:19,880 --> 00:14:24,399
+kind of into the you know main event
+
+335
+00:14:22,800 --> 00:14:27,759
+where I talk about the stuff that people
+
+336
+00:14:24,399 --> 00:14:31,040
+are actually using a lot no it is um any
+
+337
+00:14:27,759 --> 00:14:31,040
+questions about these three
+
+338
+00:14:33,000 --> 00:14:39,120
+okay cool um so now I'd like to go into
+
+339
+00:14:36,399 --> 00:14:40,639
+Chain of Thought and variance and I
+
+340
+00:14:39,120 --> 00:14:42,279
+actually have already talked about Chain
+
+341
+00:14:40,639 --> 00:14:44,199
+of Thought in fact we've mentioned it a
+
+342
+00:14:42,279 --> 00:14:47,720
+couple times um but just you know to
+
+343
+00:14:44,199 --> 00:14:49,399
+remind everybody the basic idea is um
+
+344
+00:14:47,720 --> 00:14:52,880
+compared to standard prompting where we
+
+345
+00:14:49,399 --> 00:14:55,519
+have like a question um and an answer in
+
+346
+00:14:52,880 --> 00:14:58,480
+Chain of Thought we have a question and
+
+347
+00:14:55,519 --> 00:15:01,040
+then we have a derivation for the
+
+348
+00:14:58,480 --> 00:15:02,440
+questions so like uh Roger started with
+
+349
+00:15:01,040 --> 00:15:06,120
+five
+
+350
+00:15:02,440 --> 00:15:09,040
+balls two can uh five balls two cans of
+
+351
+00:15:06,120 --> 00:15:13,839
+three tennis balls each is six tenis 5
+
+352
+00:15:09,040 --> 00:15:15,639
+plus 6al 11 the answer is 11 so um you
+
+353
+00:15:13,839 --> 00:15:17,519
+add this to the prompt and by adding
+
+354
+00:15:15,639 --> 00:15:19,240
+this to the prompt you get the model to
+
+355
+00:15:17,519 --> 00:15:22,600
+uh also do these derivations at test
+
+356
+00:15:19,240 --> 00:15:25,199
+time and this greatly improves some
+
+357
+00:15:22,600 --> 00:15:27,759
+tasks it improves tasks where we can't
+
+358
+00:15:25,199 --> 00:15:30,040
+like immediately predict the answer
+
+359
+00:15:27,759 --> 00:15:32,000
+directly and then I also previously
+
+360
+00:15:30,040 --> 00:15:33,440
+talked about zero shot Chain of Thought
+
+361
+00:15:32,000 --> 00:15:35,880
+uh reasoning where we just prompt the
+
+362
+00:15:33,440 --> 00:15:38,480
+model to with something like let's think
+
+363
+00:15:35,880 --> 00:15:42,680
+step by step and then the model becomes
+
+364
+00:15:38,480 --> 00:15:46,240
+able to do this uh Chain of Thought
+
+365
+00:15:42,680 --> 00:15:48,279
+reasoning okay so that was review and
+
+366
+00:15:46,240 --> 00:15:51,680
+now I'd like to talk about some of like
+
+367
+00:15:48,279 --> 00:15:53,560
+more advanced methods that people use
+
+368
+00:15:51,680 --> 00:15:55,079
+for uh reasoning as
+
+369
+00:15:53,560 --> 00:15:58,040
+well
+
+370
+00:15:55,079 --> 00:15:59,959
+and this is by no means an exhaustive
+
+371
+00:15:58,040 --> 00:16:01,800
+list they're just of the ones that I
+
+372
+00:15:59,959 --> 00:16:03,319
+found interesting so if you know other
+
+373
+00:16:01,800 --> 00:16:04,839
+ones that you'd like to talk about or
+
+374
+00:16:03,319 --> 00:16:07,720
+introduce to the class or something like
+
+375
+00:16:04,839 --> 00:16:10,600
+that I'd also be happy to uh to hear uh
+
+376
+00:16:07,720 --> 00:16:14,120
+which ones you like or have heard about
+
+377
+00:16:10,600 --> 00:16:16,920
+but the first one is um self-as and one
+
+378
+00:16:14,120 --> 00:16:20,959
+of the issues with large language models
+
+379
+00:16:16,920 --> 00:16:23,240
+nowadays is that they're not uh very
+
+380
+00:16:20,959 --> 00:16:25,519
+good at asking follow-up questions or
+
+381
+00:16:23,240 --> 00:16:27,839
+maybe not that they're not very good at
+
+382
+00:16:25,519 --> 00:16:31,160
+it but just they're not trained to do it
+
+383
+00:16:27,839 --> 00:16:32,880
+so like if you play around with chat GPT
+
+384
+00:16:31,160 --> 00:16:35,240
+I have never had chat GPT ask me a
+
+385
+00:16:32,880 --> 00:16:36,680
+follow-up question I don't think it's
+
+386
+00:16:35,240 --> 00:16:38,319
+like it's not because large language
+
+387
+00:16:36,680 --> 00:16:41,920
+models aren't capable of doing it it's
+
+388
+00:16:38,319 --> 00:16:43,519
+just that they like the open AI must
+
+389
+00:16:41,920 --> 00:16:45,000
+think it's a bad user experience to have
+
+390
+00:16:43,519 --> 00:16:47,680
+a language model that asks you follow up
+
+391
+00:16:45,000 --> 00:16:51,319
+questions that's only like you know
+
+392
+00:16:47,680 --> 00:16:53,160
+reason I can think about it um but
+
+393
+00:16:51,319 --> 00:16:56,199
+basically what self ask does is it
+
+394
+00:16:53,160 --> 00:17:00,000
+explicitly prompts the model to ask to
+
+395
+00:16:56,199 --> 00:17:02,360
+ask if there are followup questions so
+
+396
+00:17:00,000 --> 00:17:05,799
+here's an example on the left where the
+
+397
+00:17:02,360 --> 00:17:11,240
+question is uh who lived longer Theodore
+
+398
+00:17:05,799 --> 00:17:12,640
+haer or Harry vau wat uh Watkins and
+
+399
+00:17:11,240 --> 00:17:15,240
+basically it says are follow-up
+
+400
+00:17:12,640 --> 00:17:17,679
+questions needed here yes and then the
+
+401
+00:17:15,240 --> 00:17:20,319
+followup is how old was Theodore hacker
+
+402
+00:17:17,679 --> 00:17:23,640
+when he died and the intermediate answer
+
+403
+00:17:20,319 --> 00:17:26,959
+is Theodore hacker was 65 years old how
+
+404
+00:17:23,640 --> 00:17:29,000
+old was Harry Von Watkins um Harry Von
+
+405
+00:17:26,959 --> 00:17:32,400
+Watkins was 69 years old but so the
+
+406
+00:17:29,000 --> 00:17:35,240
+final answer is Harry Bon Watkins and um
+
+407
+00:17:32,400 --> 00:17:37,520
+in this particular paper this is just
+
+408
+00:17:35,240 --> 00:17:42,520
+like another variety of Chain of Thought
+
+409
+00:17:37,520 --> 00:17:44,720
+it's like not using it to incorporate
+
+410
+00:17:42,520 --> 00:17:47,400
+any external information or anything
+
+411
+00:17:44,720 --> 00:17:48,720
+like that it's just trying to more
+
+412
+00:17:47,400 --> 00:17:52,360
+directly
+
+413
+00:17:48,720 --> 00:17:53,840
+elicit um information from the model um
+
+414
+00:17:52,360 --> 00:17:55,360
+but nonetheless they demonstrate that
+
+415
+00:17:53,840 --> 00:17:57,760
+this is useful and then there's also
+
+416
+00:17:55,360 --> 00:18:00,120
+other methods that actually try to look
+
+417
+00:17:57,760 --> 00:18:02,240
+up information explicit to answer these
+
+418
+00:18:00,120 --> 00:18:05,280
+questions um which are even more
+
+419
+00:18:02,240 --> 00:18:05,280
+powerful than what we have
+
+420
+00:18:05,720 --> 00:18:13,200
+here um so that's what I'd like to
+
+421
+00:18:09,960 --> 00:18:16,919
+introduce next and basically the idea um
+
+422
+00:18:13,200 --> 00:18:19,760
+here is this is a method that instead of
+
+423
+00:18:16,919 --> 00:18:22,880
+just doing Chain of Thought it retrieves
+
+424
+00:18:19,760 --> 00:18:25,480
+relevant sentences when you're doing the
+
+425
+00:18:22,880 --> 00:18:28,919
+Chain of Thought So like
+
+426
+00:18:25,480 --> 00:18:30,880
+here um
+
+427
+00:18:28,919 --> 00:18:32,960
+uh we have the followup are follow-ups
+
+428
+00:18:30,880 --> 00:18:35,159
+needed here yes and then this is the
+
+429
+00:18:32,960 --> 00:18:36,880
+followup but if the model itself doesn't
+
+430
+00:18:35,159 --> 00:18:39,440
+know how old somebody was when they died
+
+431
+00:18:36,880 --> 00:18:40,760
+then it won't be able to answer this so
+
+432
+00:18:39,440 --> 00:18:44,400
+what they do in order to make this
+
+433
+00:18:40,760 --> 00:18:47,200
+happen is they um do bm25 based
+
+434
+00:18:44,400 --> 00:18:49,520
+retrieval over Wikipedia for each of the
+
+435
+00:18:47,200 --> 00:18:51,760
+Chain of Thought uh answers and then
+
+436
+00:18:49,520 --> 00:18:53,400
+they use the retrieved uh I think it's
+
+437
+00:18:51,760 --> 00:18:56,039
+like 10 documents or something like that
+
+438
+00:18:53,400 --> 00:18:59,640
+multiple retriev documents to prompt the
+
+439
+00:18:56,039 --> 00:19:03,200
+model um to basically follow up with its
+
+440
+00:18:59,640 --> 00:19:05,440
+Chain of Thought so this is another uh
+
+441
+00:19:03,200 --> 00:19:07,880
+variety of things that you can do in
+
+442
+00:19:05,440 --> 00:19:07,880
+order to
+
+443
+00:19:10,720 --> 00:19:16,120
+improve
+
+444
+00:19:13,120 --> 00:19:16,120
+cool
+
+445
+00:19:16,400 --> 00:19:21,440
+um then another one that I'd like to
+
+446
+00:19:18,960 --> 00:19:22,559
+talk about is U multilingual Chain of
+
+447
+00:19:21,440 --> 00:19:24,039
+Thought reasoning I'm going to be
+
+448
+00:19:22,559 --> 00:19:28,000
+talking more about multilingual things
+
+449
+00:19:24,039 --> 00:19:29,960
+in the multilingual class in a week but
+
+450
+00:19:28,000 --> 00:19:33,559
+the interesting thing about multilingual
+
+451
+00:19:29,960 --> 00:19:37,200
+Chain of Thought is we have a design
+
+452
+00:19:33,559 --> 00:19:41,280
+decision right like do we want to just
+
+453
+00:19:37,200 --> 00:19:44,000
+answer questions in the language that we
+
+454
+00:19:41,280 --> 00:19:46,679
+are asking questions in like so if I ask
+
+455
+00:19:44,000 --> 00:19:48,080
+a question in Japanese am I going to
+
+456
+00:19:46,679 --> 00:19:49,840
+have it go through the whole chain of
+
+457
+00:19:48,080 --> 00:19:52,720
+thought process in Japanese and then
+
+458
+00:19:49,840 --> 00:19:55,840
+answer my question in Japanese or do I
+
+459
+00:19:52,720 --> 00:19:57,120
+want it to uh somehow go through English
+
+460
+00:19:55,840 --> 00:19:59,159
+because the model has been trained on
+
+461
+00:19:57,120 --> 00:20:00,640
+lots of English and it has better
+
+462
+00:19:59,159 --> 00:20:02,120
+it's like a better way to take advantage
+
+463
+00:20:00,640 --> 00:20:04,840
+of its reasoning
+
+464
+00:20:02,120 --> 00:20:07,200
+capabilities does anyone have a idea
+
+465
+00:20:04,840 --> 00:20:07,200
+about the
+
+466
+00:20:07,960 --> 00:20:12,480
+answer who thinks it's better to do it
+
+467
+00:20:10,240 --> 00:20:15,360
+entirely in the the language that the
+
+468
+00:20:12,480 --> 00:20:15,360
+question is asked
+
+469
+00:20:15,640 --> 00:20:20,080
+in and who thinks it's better to do
+
+470
+00:20:17,919 --> 00:20:23,000
+something in
+
+471
+00:20:20,080 --> 00:20:28,200
+English
+
+472
+00:20:23,000 --> 00:20:29,159
+okay so um basically the answer is do it
+
+473
+00:20:28,200 --> 00:20:31,440
+in English
+
+474
+00:20:29,159 --> 00:20:34,120
+um and maybe this
+
+475
+00:20:31,440 --> 00:20:35,799
+is it might be a little bit dependent on
+
+476
+00:20:34,120 --> 00:20:39,840
+the language but all of the languages
+
+477
+00:20:35,799 --> 00:20:42,880
+they tested it's essentially uh that's
+
+478
+00:20:39,840 --> 00:20:44,919
+the conclusion that they came to and
+
+479
+00:20:42,880 --> 00:20:47,679
+it's pretty Stark in this particular
+
+480
+00:20:44,919 --> 00:20:50,640
+paper this might change a little bit
+
+481
+00:20:47,679 --> 00:20:52,960
+with um with more powerful models but I
+
+482
+00:20:50,640 --> 00:20:57,360
+still would be very surprised if this is
+
+483
+00:20:52,960 --> 00:21:00,440
+not like if this doesn't hold still so
+
+484
+00:20:57,360 --> 00:21:04,440
+you can see it's like approximately on
+
+485
+00:21:00,440 --> 00:21:08,200
+average uh 7even Point increase in the
+
+486
+00:21:04,440 --> 00:21:11,720
+results and just to to be clear here um
+
+487
+00:21:08,200 --> 00:21:13,600
+we have native uh Chain of Thought So
+
+488
+00:21:11,720 --> 00:21:16,039
+This is doing Chain of Thought in the in
+
+489
+00:21:13,600 --> 00:21:17,799
+the language itself this is doing Chain
+
+490
+00:21:16,039 --> 00:21:19,240
+of Thought in English but then answering
+
+491
+00:21:17,799 --> 00:21:22,200
+in the language itself and this is just
+
+492
+00:21:19,240 --> 00:21:23,799
+like translating everything into
+
+493
+00:21:22,200 --> 00:21:27,440
+English
+
+494
+00:21:23,799 --> 00:21:30,159
+um you can try this out too like if you
+
+495
+00:21:27,440 --> 00:21:31,840
+uh if you speak another Lang you can um
+
+496
+00:21:30,159 --> 00:21:34,200
+try to do it myself when I try it in
+
+497
+00:21:31,840 --> 00:21:36,200
+Japanese it's very clear that like the
+
+498
+00:21:34,200 --> 00:21:38,640
+model seems more intelligent in English
+
+499
+00:21:36,200 --> 00:21:41,559
+it just can seems like it can do other
+
+500
+00:21:38,640 --> 00:21:43,120
+things even though like intelligence uh
+
+501
+00:21:41,559 --> 00:21:44,640
+shouldn't be a function of the language
+
+502
+00:21:43,120 --> 00:21:47,120
+that you're asking a question in right
+
+503
+00:21:44,640 --> 00:21:49,679
+like the model should have the ability
+
+504
+00:21:47,120 --> 00:21:51,440
+to answer questions but it because
+
+505
+00:21:49,679 --> 00:21:53,000
+that's how humans work right our
+
+506
+00:21:51,440 --> 00:21:54,520
+intelligence is kind of separated from
+
+507
+00:21:53,000 --> 00:21:57,039
+our language how well we can express
+
+508
+00:21:54,520 --> 00:22:00,480
+ourselves is a little bit different but
+
+509
+00:21:57,039 --> 00:22:02,320
+um yeah for the final appli this was it
+
+510
+00:22:00,480 --> 00:22:04,840
+translated back to the original language
+
+511
+00:22:02,320 --> 00:22:09,440
+and then evaluated for translate English
+
+512
+00:22:04,840 --> 00:22:12,559
+I'm not 100% sure about this I think it
+
+513
+00:22:09,440 --> 00:22:13,840
+was not so that might be a confounding
+
+514
+00:22:12,559 --> 00:22:16,799
+factor for this one but it's not a
+
+515
+00:22:13,840 --> 00:22:20,039
+confounding factor for this one anyway
+
+516
+00:22:16,799 --> 00:22:20,039
+yeah any other
+
+517
+00:22:20,679 --> 00:22:23,919
+questions Okay
+
+518
+00:22:24,200 --> 00:22:29,559
+cool so this is a pretty interesting
+
+519
+00:22:26,799 --> 00:22:32,000
+result here um
+
+520
+00:22:29,559 --> 00:22:34,120
+and the next kind of series of results
+
+521
+00:22:32,000 --> 00:22:35,360
+are going to be based on the uh that I'm
+
+522
+00:22:34,120 --> 00:22:36,919
+going to talk about are going to be
+
+523
+00:22:35,360 --> 00:22:39,240
+based on the quality of the reasoning
+
+524
+00:22:36,919 --> 00:22:43,480
+chains that the model uses in Chain of
+
+525
+00:22:39,240 --> 00:22:45,520
+Thought and this one is a simple
+
+526
+00:22:43,480 --> 00:22:46,600
+heuristic for improving the quality of
+
+527
+00:22:45,520 --> 00:22:49,279
+the reasoning
+
+528
+00:22:46,600 --> 00:22:50,640
+chains and um yeah one thing I should
+
+529
+00:22:49,279 --> 00:22:52,480
+mention is that the quality of the
+
+530
+00:22:50,640 --> 00:22:55,760
+reasoning chain is definitely connected
+
+531
+00:22:52,480 --> 00:22:58,080
+to the uh quality of the output like
+
+532
+00:22:55,760 --> 00:23:00,159
+some that's not necessarily the case
+
+533
+00:22:58,080 --> 00:23:04,679
+right it could just say a whole bunch of
+
+534
+00:23:00,159 --> 00:23:07,799
+you know false like uh actually no maybe
+
+535
+00:23:04,679 --> 00:23:07,799
+I'll I'll skip this
+
+536
+00:23:08,200 --> 00:23:14,919
+one and go and and explain this one next
+
+537
+00:23:11,919 --> 00:23:14,919
+so
+
+538
+00:23:15,159 --> 00:23:19,039
+um yeah actually sorry the or the
+
+539
+00:23:17,600 --> 00:23:20,520
+explanation ordering for this is a
+
+540
+00:23:19,039 --> 00:23:25,360
+little bit hard but yeah I'll explain
+
+541
+00:23:20,520 --> 00:23:26,840
+this one next so um very quickly um
+
+542
+00:23:25,360 --> 00:23:29,640
+there's two ways that you could be
+
+543
+00:23:26,840 --> 00:23:32,880
+reasoning one way you could be reasoning
+
+544
+00:23:29,640 --> 00:23:35,000
+is doing an explanation first and then
+
+545
+00:23:32,880 --> 00:23:36,720
+uh predicting the answer the other way
+
+546
+00:23:35,000 --> 00:23:39,080
+you could do it is predicting the answer
+
+547
+00:23:36,720 --> 00:23:43,039
+and then do it um then giving the
+
+548
+00:23:39,080 --> 00:23:45,559
+explanation and in general if you have a
+
+549
+00:23:43,039 --> 00:23:47,919
+reasonably strong model uh you know any
+
+550
+00:23:45,559 --> 00:23:50,679
+of the modern kind of Frontier level
+
+551
+00:23:47,919 --> 00:23:52,240
+models right now doing the explanation
+
+552
+00:23:50,679 --> 00:23:54,039
+first and then making the prediction is
+
+553
+00:23:52,240 --> 00:23:56,880
+better and the reason why is because
+
+554
+00:23:54,039 --> 00:23:59,240
+Chain of Thought works and the model is
+
+555
+00:23:56,880 --> 00:24:02,960
+able to break down the quest um the
+
+556
+00:23:59,240 --> 00:24:07,279
+questions into kind of
+
+557
+00:24:02,960 --> 00:24:10,159
+simpler uh it's able to break down the
+
+558
+00:24:07,279 --> 00:24:11,520
+like the answer into like simp simpler
+
+559
+00:24:10,159 --> 00:24:14,080
+questions for like mathematical
+
+560
+00:24:11,520 --> 00:24:15,679
+reasoning or something like that um and
+
+561
+00:24:14,080 --> 00:24:18,039
+then give me the answer so like for
+
+562
+00:24:15,679 --> 00:24:20,000
+example for text DCI 002 which was State
+
+563
+00:24:18,039 --> 00:24:22,679
+ofth art at the time of this writing you
+
+564
+00:24:20,000 --> 00:24:24,360
+see a fivepoint boost from using um
+
+565
+00:24:22,679 --> 00:24:29,080
+explanation first and then prediction
+
+566
+00:24:24,360 --> 00:24:30,640
+after that um and in accur
+
+567
+00:24:29,080 --> 00:24:34,039
+but for the weaker models that was not
+
+568
+00:24:30,640 --> 00:24:36,039
+the case so if you were using um GPD 3
+
+569
+00:24:34,039 --> 00:24:38,720
+that wasn't trained for Chain of Thought
+
+570
+00:24:36,039 --> 00:24:40,600
+or you were using opt uh that was not
+
+571
+00:24:38,720 --> 00:24:42,640
+the case but nowadays I think basically
+
+572
+00:24:40,600 --> 00:24:45,279
+all models uh doing the explanation
+
+573
+00:24:42,640 --> 00:24:48,120
+first and then the prediction is
+
+574
+00:24:45,279 --> 00:24:49,640
+better um so going
+
+575
+00:24:48,120 --> 00:24:51,640
+back
+
+576
+00:24:49,640 --> 00:24:53,559
+um another thing that people have
+
+577
+00:24:51,640 --> 00:24:55,120
+noticed is like if your explanation is
+
+578
+00:24:53,559 --> 00:24:56,520
+wrong your prediction also tends to be
+
+579
+00:24:55,120 --> 00:24:58,120
+wrong so if you make mistakes in
+
+580
+00:24:56,520 --> 00:25:00,520
+intermediate steps of your explanation
+
+581
+00:24:58,120 --> 00:25:03,679
+it's tends to mess up your final
+
+582
+00:25:00,520 --> 00:25:06,000
+prediction um so like one of the
+
+583
+00:25:03,679 --> 00:25:09,320
+interesting ways that people have found
+
+584
+00:25:06,000 --> 00:25:11,559
+to improve the final the explanation
+
+585
+00:25:09,320 --> 00:25:13,880
+quality is they just observe that if the
+
+586
+00:25:11,559 --> 00:25:18,840
+explanations are longer they tend to be
+
+587
+00:25:13,880 --> 00:25:20,960
+better it's uh kind of interesting but
+
+588
+00:25:18,840 --> 00:25:23,000
+like if they give you more reasoning
+
+589
+00:25:20,960 --> 00:25:25,000
+steps this tends to be more accurate and
+
+590
+00:25:23,000 --> 00:25:27,320
+they actually demonstrate that in this
+
+591
+00:25:25,000 --> 00:25:29,200
+paper where here's a simple reasoning
+
+592
+00:25:27,320 --> 00:25:31,720
+chain here's a more complex reasoning
+
+593
+00:25:29,200 --> 00:25:35,480
+chain and you actually see for exactly
+
+594
+00:25:31,720 --> 00:25:36,760
+the same problem they get about a 15%
+
+595
+00:25:35,480 --> 00:25:38,360
+boost and these are kind of like
+
+596
+00:25:36,760 --> 00:25:39,960
+naturally occurring reasoning chains
+
+597
+00:25:38,360 --> 00:25:41,520
+they didn't like train the model to give
+
+598
+00:25:39,960 --> 00:25:43,919
+you longer reasoning chains or anything
+
+599
+00:25:41,520 --> 00:25:45,279
+like that but amongst the naturally
+
+600
+00:25:43,919 --> 00:25:46,840
+occurring reasoning chains the longer
+
+601
+00:25:45,279 --> 00:25:50,480
+ones tend to be
+
+602
+00:25:46,840 --> 00:25:53,159
+better and this fact could be simply
+
+603
+00:25:50,480 --> 00:25:54,679
+used to improve accuracy um and so the
+
+604
+00:25:53,159 --> 00:25:57,360
+way they did this is they just sampled
+
+605
+00:25:54,679 --> 00:25:59,279
+multiple reasoning paths and then they
+
+606
+00:25:57,360 --> 00:26:00,840
+performed self consistency over the
+
+607
+00:25:59,279 --> 00:26:03,000
+longer reasoning paths so if you
+
+608
+00:26:00,840 --> 00:26:05,240
+remember what self consistency is it's
+
+609
+00:26:03,000 --> 00:26:07,240
+basically like you do majority voting
+
+610
+00:26:05,240 --> 00:26:09,679
+over the answers for multiple reasoning
+
+611
+00:26:07,240 --> 00:26:13,880
+paths so they threw out the lower
+
+612
+00:26:09,679 --> 00:26:13,880
+quality ones and that improved overall
+
+613
+00:26:14,399 --> 00:26:20,279
+accuracy so um yeah that's a thing that
+
+614
+00:26:18,000 --> 00:26:20,279
+you can
+
+615
+00:26:21,039 --> 00:26:25,960
+do
+
+616
+00:26:23,120 --> 00:26:28,880
+um so yeah going back to systematic
+
+617
+00:26:25,960 --> 00:26:31,360
+studies of reasoning in llms
+
+618
+00:26:28,880 --> 00:26:33,559
+um one of the big results that's
+
+619
+00:26:31,360 --> 00:26:35,880
+actually really important to know about
+
+620
+00:26:33,559 --> 00:26:39,039
+is th this sort of Chain of Thought
+
+621
+00:26:35,880 --> 00:26:41,080
+reasoning um is considered to be an
+
+622
+00:26:39,039 --> 00:26:43,520
+emergent ability
+
+623
+00:26:41,080 --> 00:26:47,080
+in uh large language models and what we
+
+624
+00:26:43,520 --> 00:26:49,360
+mean by an emergent ability is it's or
+
+625
+00:26:47,080 --> 00:26:53,679
+what what the the name emergent ability
+
+626
+00:26:49,360 --> 00:26:56,399
+typically refers to is that it is
+
+627
+00:26:53,679 --> 00:26:58,640
+something that increases dramatically as
+
+628
+00:26:56,399 --> 00:27:01,679
+the model size gets uh up up to a
+
+629
+00:26:58,640 --> 00:27:03,200
+certain point so these actually I'm I'm
+
+630
+00:27:01,679 --> 00:27:06,080
+really sorry I cut off the thing on the
+
+631
+00:27:03,200 --> 00:27:07,360
+bottom here this is like open AI does
+
+632
+00:27:06,080 --> 00:27:08,520
+this all the time to not tell you how
+
+633
+00:27:07,360 --> 00:27:11,399
+many parameters they have in their
+
+634
+00:27:08,520 --> 00:27:12,760
+models but I did not do it intentionally
+
+635
+00:27:11,399 --> 00:27:15,360
+here because I think it's actually in
+
+636
+00:27:12,760 --> 00:27:17,320
+here in the paper um but like these ones
+
+637
+00:27:15,360 --> 00:27:19,399
+over here are kind of the like 175
+
+638
+00:27:17,320 --> 00:27:20,640
+billion parameter models and like the
+
+639
+00:27:19,399 --> 00:27:24,520
+the larger
+
+640
+00:27:20,640 --> 00:27:25,960
+models um and what you see is like up
+
+641
+00:27:24,520 --> 00:27:29,919
+until a certain point you get basically
+
+642
+00:27:25,960 --> 00:27:33,919
+zero accuracy and then uh the outputs
+
+643
+00:27:29,919 --> 00:27:37,000
+improve and so for a while people were
+
+644
+00:27:33,919 --> 00:27:39,240
+really like confused about this like why
+
+645
+00:27:37,000 --> 00:27:41,440
+why does this happen it feels like magic
+
+646
+00:27:39,240 --> 00:27:44,279
+that you get a really you know powerful
+
+647
+00:27:41,440 --> 00:27:46,679
+model and then suddenly it gets better
+
+648
+00:27:44,279 --> 00:27:49,799
+uh uh like at the very
+
+649
+00:27:46,679 --> 00:27:52,159
+end but actually there's a much simpler
+
+650
+00:27:49,799 --> 00:27:53,760
+solution there's not not that much magic
+
+651
+00:27:52,159 --> 00:27:55,960
+to this
+
+652
+00:27:53,760 --> 00:27:58,399
+and we've known about this for a little
+
+653
+00:27:55,960 --> 00:28:00,919
+while but this paper from 2023 really
+
+654
+00:27:58,399 --> 00:28:02,360
+like expressed it very clearly um so I
+
+655
+00:28:00,919 --> 00:28:04,360
+highly recommend you take a look at this
+
+656
+00:28:02,360 --> 00:28:07,720
+if you're interested in kind of like the
+
+657
+00:28:04,360 --> 00:28:10,159
+emerg abilities and language models but
+
+658
+00:28:07,720 --> 00:28:15,039
+basically the the thing about emergent
+
+659
+00:28:10,159 --> 00:28:19,720
+abilities is that they're mostly
+
+660
+00:28:15,039 --> 00:28:20,720
+a matter of how you um how you measure
+
+661
+00:28:19,720 --> 00:28:22,519
+your
+
+662
+00:28:20,720 --> 00:28:27,640
+models
+
+663
+00:28:22,519 --> 00:28:30,120
+accuracy and so let's say as your model
+
+664
+00:28:27,640 --> 00:28:30,120
+gets better
+
+665
+00:28:39,039 --> 00:28:45,600
+it gets gradually better at predicting
+
+666
+00:28:41,200 --> 00:28:45,600
+the like a reasonable next
+
+667
+00:28:47,799 --> 00:28:54,760
+token so this is like a I don't know
+
+668
+00:28:50,919 --> 00:28:59,120
+like 200 million parameter model 500
+
+669
+00:28:54,760 --> 00:29:03,240
+million 1 billion 3 billion
+
+670
+00:28:59,120 --> 00:29:06,600
+7 billion and like 70 billion or
+
+671
+00:29:03,240 --> 00:29:09,600
+something like that um and so this is
+
+672
+00:29:06,600 --> 00:29:12,640
+like the next token prediction accuracy
+
+673
+00:29:09,600 --> 00:29:14,320
+um or like the the accuracy of
+
+674
+00:29:12,640 --> 00:29:16,279
+predicting a reasonable next token that
+
+675
+00:29:14,320 --> 00:29:18,880
+won't make result in your reasoning
+
+676
+00:29:16,279 --> 00:29:20,000
+chain being wrong and making a mistake
+
+677
+00:29:18,880 --> 00:29:24,200
+and
+
+678
+00:29:20,000 --> 00:29:26,200
+so if you have an accuracy like this in
+
+679
+00:29:24,200 --> 00:29:28,880
+order to get the correct answer like
+
+680
+00:29:26,200 --> 00:29:30,559
+let's say there's about five or eight
+
+681
+00:29:28,880 --> 00:29:33,519
+places where you could possibly make a
+
+682
+00:29:30,559 --> 00:29:35,080
+mistake in the derivation like one
+
+683
+00:29:33,519 --> 00:29:36,760
+common places to make a mistake in a
+
+684
+00:29:35,080 --> 00:29:38,519
+derivation for math for example are
+
+685
+00:29:36,760 --> 00:29:40,200
+where you predict a number like where
+
+686
+00:29:38,519 --> 00:29:42,679
+you predict the result of an equation
+
+687
+00:29:40,200 --> 00:29:44,120
+and you might have five reasoning steps
+
+688
+00:29:42,679 --> 00:29:47,720
+where you might predict the result of an
+
+689
+00:29:44,120 --> 00:29:53,039
+equation um and so if we do
+
+690
+00:29:47,720 --> 00:29:53,039
+this let's exponentiate all of these by
+
+691
+00:29:54,799 --> 00:29:58,799
+five um
+
+692
+00:30:06,640 --> 00:30:16,120
+uh write python code to exp
+
+693
+00:30:11,200 --> 00:30:16,120
+she these numbers by
+
+694
+00:30:19,600 --> 00:30:27,559
+five I'm wszy enough that I just ask
+
+695
+00:30:22,159 --> 00:30:27,559
+chat GP chat GP to do this for me now
+
+696
+00:30:30,080 --> 00:30:32,919
+and so if we do
+
+697
+00:30:35,399 --> 00:30:39,840
+this do go go chat
+
+698
+00:30:50,000 --> 00:30:58,360
+GPD so now we are getting something that
+
+699
+00:30:54,760 --> 00:30:58,360
+looks like zero
+
+700
+00:31:02,159 --> 00:31:07,960
+um basically zero basically
+
+701
+00:31:05,639 --> 00:31:10,960
+zero
+
+702
+00:31:07,960 --> 00:31:10,960
+uh
+
+703
+00:31:13,399 --> 00:31:16,399
+3%
+
+704
+00:31:16,799 --> 00:31:22,440
+23%
+
+705
+00:31:19,080 --> 00:31:22,440
+9% and
+
+706
+00:31:22,559 --> 00:31:28,720
+90% so what you can see is there's
+
+707
+00:31:26,639 --> 00:31:30,600
+actually a pretty steady GR gradation of
+
+708
+00:31:28,720 --> 00:31:33,120
+like the next token prediction accuracy
+
+709
+00:31:30,600 --> 00:31:36,600
+here but if you need to predict multiple
+
+710
+00:31:33,120 --> 00:31:38,919
+tokens correct then it looks like it's
+
+711
+00:31:36,600 --> 00:31:41,240
+doing basically nothing until you get up
+
+712
+00:31:38,919 --> 00:31:43,600
+to like 75% next token accuracy and then
+
+713
+00:31:41,240 --> 00:31:45,320
+it starts taking off so that's like uh
+
+714
+00:31:43,600 --> 00:31:46,960
+what happens in emergent abilities and
+
+715
+00:31:45,320 --> 00:31:49,159
+you'll notice that most things that are
+
+716
+00:31:46,960 --> 00:31:50,880
+talking about emergent abilities are
+
+717
+00:31:49,159 --> 00:31:53,559
+usually talking about some sort of Chain
+
+718
+00:31:50,880 --> 00:31:55,799
+of Thought or some sort of reasoning uh
+
+719
+00:31:53,559 --> 00:31:58,480
+reasoning accuracy even if that's not
+
+720
+00:31:55,799 --> 00:32:00,480
+the case um even if they're just
+
+721
+00:31:58,480 --> 00:32:02,639
+predicting a single token it can still
+
+722
+00:32:00,480 --> 00:32:05,399
+happen because
+
+723
+00:32:02,639 --> 00:32:08,559
+basically the probability of a single
+
+724
+00:32:05,399 --> 00:32:11,639
+token can continue to go up smoothly but
+
+725
+00:32:08,559 --> 00:32:13,240
+you only get the the token correct after
+
+726
+00:32:11,639 --> 00:32:14,760
+the probability starts getting higher
+
+727
+00:32:13,240 --> 00:32:18,320
+than all the others and that's also a
+
+728
+00:32:14,760 --> 00:32:21,279
+discontinuous function so um so
+
+729
+00:32:18,320 --> 00:32:23,080
+basically what this paper shows is like
+
+730
+00:32:21,279 --> 00:32:26,440
+even if you have like the probability of
+
+731
+00:32:23,080 --> 00:32:28,679
+the correct token going um the correct
+
+732
+00:32:26,440 --> 00:32:30,639
+token going up gradually uh you can see
+
+733
+00:32:28,679 --> 00:32:33,440
+this emergent ability based on how you
+
+734
+00:32:30,639 --> 00:32:37,279
+uh measure it so um that's an important
+
+735
+00:32:33,440 --> 00:32:38,960
+thing to realize about uh this another
+
+736
+00:32:37,279 --> 00:32:41,080
+correl of this is like let's say you
+
+737
+00:32:38,960 --> 00:32:44,679
+want to do interesting experiments about
+
+738
+00:32:41,080 --> 00:32:45,960
+reasoning on um on smaller models like
+
+739
+00:32:44,679 --> 00:32:47,279
+let's say you want to train a smaller
+
+740
+00:32:45,960 --> 00:32:49,159
+model and see how it improves on
+
+741
+00:32:47,279 --> 00:32:52,159
+reasoning I would definitely encourage
+
+742
+00:32:49,159 --> 00:32:54,799
+you to measure not only accuracy because
+
+743
+00:32:52,159 --> 00:32:57,279
+you might see like very little change in
+
+744
+00:32:54,799 --> 00:32:58,720
+accuracy but also measure like log
+
+745
+00:32:57,279 --> 00:33:00,360
+likelihood of reasoning chains or
+
+746
+00:32:58,720 --> 00:33:02,960
+something like that because you'll see a
+
+747
+00:33:00,360 --> 00:33:02,960
+a smoother
+
+748
+00:33:03,799 --> 00:33:09,080
+curve cool um any questions about
+
+749
+00:33:11,039 --> 00:33:17,240
+this okay um sounds
+
+750
+00:33:14,720 --> 00:33:20,559
+good so I I talked a little bit about
+
+751
+00:33:17,240 --> 00:33:23,120
+this um one one of the things here that
+
+752
+00:33:20,559 --> 00:33:25,320
+I didn't talk about is this paper
+
+753
+00:33:23,120 --> 00:33:28,159
+measures not just the accuracy of the
+
+754
+00:33:25,320 --> 00:33:30,880
+answer with chain of thoughts um but it
+
+755
+00:33:28,159 --> 00:33:35,840
+also measures the factuality of the
+
+756
+00:33:30,880 --> 00:33:40,480
+explanation so basically um whether the
+
+757
+00:33:35,840 --> 00:33:40,480
+explanation is a good explanation for
+
+758
+00:33:40,760 --> 00:33:47,240
+the um whether the explanation is a good
+
+759
+00:33:43,960 --> 00:33:50,039
+explanation for the actual
+
+760
+00:33:47,240 --> 00:33:51,919
+derivation um and also the consistency
+
+761
+00:33:50,039 --> 00:33:53,480
+of the answer in the explanation to
+
+762
+00:33:51,919 --> 00:33:56,120
+figure out whether the answer and the
+
+763
+00:33:53,480 --> 00:33:58,200
+explanation um match up with each other
+
+764
+00:33:56,120 --> 00:33:59,600
+and they they did this with some uh
+
+765
+00:33:58,200 --> 00:34:02,320
+synthetic data sets where you could
+
+766
+00:33:59,600 --> 00:34:07,120
+actually measure the um the re the
+
+767
+00:34:02,320 --> 00:34:10,399
+reasoning steps uh by using math so um
+
+768
+00:34:07,120 --> 00:34:13,560
+what they were able to find is basically
+
+769
+00:34:10,399 --> 00:34:15,760
+the answer and the explanation um
+
+770
+00:34:13,560 --> 00:34:17,639
+when the answer in the explanation
+
+771
+00:34:15,760 --> 00:34:22,079
+tended to be consistent especially for
+
+772
+00:34:17,639 --> 00:34:23,760
+the stronger models and let's see yeah
+
+773
+00:34:22,079 --> 00:34:25,399
+the the answer in the explanation tended
+
+774
+00:34:23,760 --> 00:34:28,440
+to be consistent especially for the
+
+775
+00:34:25,399 --> 00:34:30,879
+stronger models and um
+
+776
+00:34:28,440 --> 00:34:33,000
+that also meant that if you had higher
+
+777
+00:34:30,879 --> 00:34:35,839
+factuality in the explanation that
+
+778
+00:34:33,000 --> 00:34:38,240
+translates into higher um you know
+
+779
+00:34:35,839 --> 00:34:40,520
+factuality of the accuracy of the actual
+
+780
+00:34:38,240 --> 00:34:43,159
+prediction um I would bet that these
+
+781
+00:34:40,520 --> 00:34:45,240
+numbers are even higher uh nowadays I
+
+782
+00:34:43,159 --> 00:34:49,040
+bet the consistency is even higher uh
+
+783
+00:34:45,240 --> 00:34:49,040
+with more modern models than Tex avenci
+
+784
+00:34:49,399 --> 00:34:53,200
+002 and the re the reason being is like
+
+785
+00:34:51,839 --> 00:34:54,760
+number one models are stronger number
+
+786
+00:34:53,200 --> 00:34:56,560
+two all models are like trained for
+
+787
+00:34:54,760 --> 00:35:00,960
+Chain of Thought pretty aggressively now
+
+788
+00:34:56,560 --> 00:35:00,960
+so uh that would make the difference
+
+789
+00:35:02,200 --> 00:35:08,640
+there cool um so the the other thing I'd
+
+790
+00:35:07,000 --> 00:35:09,359
+like to talk about is training for Chain
+
+791
+00:35:08,640 --> 00:35:13,079
+of
+
+792
+00:35:09,359 --> 00:35:17,440
+Thought um so there's a fair amount of
+
+793
+00:35:13,079 --> 00:35:19,200
+work in this general direction um from
+
+794
+00:35:17,440 --> 00:35:23,040
+my point of view there's basically two
+
+795
+00:35:19,200 --> 00:35:25,800
+ways that people do this nowadays um the
+
+796
+00:35:23,040 --> 00:35:28,960
+first way is usually through generating
+
+797
+00:35:25,800 --> 00:35:33,480
+lots of synthetic data that represents
+
+798
+00:35:28,960 --> 00:35:37,800
+chains of thoughts and then using that
+
+799
+00:35:33,480 --> 00:35:39,520
+to um to train models and this is the
+
+800
+00:35:37,800 --> 00:35:41,839
+most famous version of this although
+
+801
+00:35:39,520 --> 00:35:44,079
+this paper cites a lot of uh a lot of
+
+802
+00:35:41,839 --> 00:35:45,760
+other ones but basically they generate a
+
+803
+00:35:44,079 --> 00:35:48,280
+large and diverse uh Chain of Thought
+
+804
+00:35:45,760 --> 00:35:51,240
+data set from GPT 3.5 and
+
+805
+00:35:48,280 --> 00:35:53,200
+gp4 um it includes 5 million complex
+
+806
+00:35:51,240 --> 00:35:55,640
+instructions I think they generated 1
+
+807
+00:35:53,200 --> 00:35:59,000
+million from GPD 4 and 4 million from uh
+
+808
+00:35:55,640 --> 00:36:01,640
+GPT 3.5 just because generating long
+
+809
+00:35:59,000 --> 00:36:06,520
+sequences from gp4 is expensive and they
+
+810
+00:36:01,640 --> 00:36:09,640
+didn't want to do that many um and
+
+811
+00:36:06,520 --> 00:36:11,760
+then they uh achieved corresponding high
+
+812
+00:36:09,640 --> 00:36:13,200
+accuracy on Chain of Thought related
+
+813
+00:36:11,760 --> 00:36:16,200
+things compared to other data sets so
+
+814
+00:36:13,200 --> 00:36:17,760
+compared to like alpaka which is much uh
+
+815
+00:36:16,200 --> 00:36:21,760
+smaller and doesn't have as much Chain
+
+816
+00:36:17,760 --> 00:36:24,079
+of Thought and also um uh vicuna which
+
+817
+00:36:21,760 --> 00:36:26,640
+is similarly less focused on chain of
+
+818
+00:36:24,079 --> 00:36:29,359
+thought they were able to do uh a good
+
+819
+00:36:26,640 --> 00:36:31,599
+job
+
+820
+00:36:29,359 --> 00:36:33,640
+um this paper was by Microsoft and they
+
+821
+00:36:31,599 --> 00:36:36,960
+didn't actually release the Orca data
+
+822
+00:36:33,640 --> 00:36:39,400
+set um for whatever reason uh legal
+
+823
+00:36:36,960 --> 00:36:41,400
+legal or competitive reasons or whatever
+
+824
+00:36:39,400 --> 00:36:43,000
+but there's another open Orca data set
+
+825
+00:36:41,400 --> 00:36:44,359
+that you can download and use uh that
+
+826
+00:36:43,000 --> 00:36:47,480
+attempts to replicate it and it's
+
+827
+00:36:44,359 --> 00:36:50,440
+reasonably good so uh you you can uh
+
+828
+00:36:47,480 --> 00:36:50,440
+keep that in mind if you're
+
+829
+00:36:50,800 --> 00:36:59,520
+interested um this is another really
+
+830
+00:36:53,280 --> 00:36:59,520
+interesting paper on uh trying to create
+
+831
+00:37:00,160 --> 00:37:05,760
+assessments automatic assessments of how
+
+832
+00:37:03,440 --> 00:37:09,880
+good chains of thought are and what they
+
+833
+00:37:05,760 --> 00:37:13,079
+do essentially is it's relatively simple
+
+834
+00:37:09,880 --> 00:37:15,200
+they get human feedback on each step of
+
+835
+00:37:13,079 --> 00:37:17,760
+a derivation so they just basically ask
+
+836
+00:37:15,200 --> 00:37:20,599
+people is this step of the derivation
+
+837
+00:37:17,760 --> 00:37:22,160
+good and uh if the answer is yes then
+
+838
+00:37:20,599 --> 00:37:24,760
+they give it a a smiley face if the
+
+839
+00:37:22,160 --> 00:37:26,440
+answer is no they give it a frowny face
+
+840
+00:37:24,760 --> 00:37:28,560
+and they use this to train a reward
+
+841
+00:37:26,440 --> 00:37:32,000
+model where the reward model basically
+
+842
+00:37:28,560 --> 00:37:34,760
+predicts whether each uh thing of the um
+
+843
+00:37:32,000 --> 00:37:36,800
+each step of the derivation is good and
+
+844
+00:37:34,760 --> 00:37:38,160
+so we have two examples over here I know
+
+845
+00:37:36,800 --> 00:37:41,160
+this is really small you might be able
+
+846
+00:37:38,160 --> 00:37:43,200
+to see it um either in the paper on uh
+
+847
+00:37:41,160 --> 00:37:46,359
+the slides on the website but what we
+
+848
+00:37:43,200 --> 00:37:49,000
+can see here is that it assesses each of
+
+849
+00:37:46,359 --> 00:37:52,680
+these steps and uh checks that the
+
+850
+00:37:49,000 --> 00:37:55,760
+answer is good um but it's also able to
+
+851
+00:37:52,680 --> 00:37:57,119
+identify places where uh like steps are
+
+852
+00:37:55,760 --> 00:37:59,560
+incorrect and then the final answer
+
+853
+00:37:57,119 --> 00:38:02,560
+becomes Incorrect and then they use this
+
+854
+00:37:59,560 --> 00:38:04,440
+for training um a Chain of Thought style
+
+855
+00:38:02,560 --> 00:38:06,319
+model so they have the model generate
+
+856
+00:38:04,440 --> 00:38:08,520
+chains of thought and they assess them
+
+857
+00:38:06,319 --> 00:38:10,079
+with the reward model and upweight
+
+858
+00:38:08,520 --> 00:38:12,160
+answers that have good chains of thought
+
+859
+00:38:10,079 --> 00:38:15,680
+and so the good thing about this is they
+
+860
+00:38:12,160 --> 00:38:17,440
+actually don't need um they don't need
+
+861
+00:38:15,680 --> 00:38:20,160
+the correct answers to train the model
+
+862
+00:38:17,440 --> 00:38:21,640
+this way and because they don't need the
+
+863
+00:38:20,160 --> 00:38:23,920
+correct answers to train the model this
+
+864
+00:38:21,640 --> 00:38:26,640
+way they can also train the model on
+
+865
+00:38:23,920 --> 00:38:29,200
+lots of other questions the reason why
+
+866
+00:38:26,640 --> 00:38:31,520
+this works is because like Chain of
+
+867
+00:38:29,200 --> 00:38:34,880
+Thought makes it easier to generate each
+
+868
+00:38:31,520 --> 00:38:36,720
+of the steps in the derivation it's also
+
+869
+00:38:34,880 --> 00:38:38,640
+easier to assess whether an individual
+
+870
+00:38:36,720 --> 00:38:40,000
+step in a derivation is wrong then
+
+871
+00:38:38,640 --> 00:38:42,960
+assess whether the answer is correct
+
+872
+00:38:40,000 --> 00:38:45,319
+overall so um this feedback signal is
+
+873
+00:38:42,960 --> 00:38:48,640
+easier to get model provided than it is
+
+874
+00:38:45,319 --> 00:38:51,160
+for um uh like getting feedback on the
+
+875
+00:38:48,640 --> 00:38:53,839
+answer itself yeah failure in one step
+
+876
+00:38:51,160 --> 00:38:56,920
+causes all the other steps to fail yep
+
+877
+00:38:53,839 --> 00:38:57,960
+you just assess the next steps based on
+
+878
+00:38:56,920 --> 00:39:00,079
+the assumption
+
+879
+00:38:57,960 --> 00:39:02,920
+the or do
+
+880
+00:39:00,079 --> 00:39:05,240
+you I I don't think
+
+881
+00:39:02,920 --> 00:39:07,599
+they I don't think they do that I think
+
+882
+00:39:05,240 --> 00:39:10,119
+they um it it's a good question I'm not
+
+883
+00:39:07,599 --> 00:39:12,160
+100% sure about this but I think they um
+
+884
+00:39:10,119 --> 00:39:14,280
+assess each one of the steps
+
+885
+00:39:12,160 --> 00:39:15,920
+independently um and it's not
+
+886
+00:39:14,280 --> 00:39:17,480
+necessarily the case that like failing
+
+887
+00:39:15,920 --> 00:39:19,000
+on this step means the step is wrong
+
+888
+00:39:17,480 --> 00:39:21,319
+right it could be just not using it at
+
+889
+00:39:19,000 --> 00:39:25,240
+all also
+
+890
+00:39:21,319 --> 00:39:25,240
+so um
+
+891
+00:39:25,440 --> 00:39:31,119
+cool so a final thing like to talk about
+
+892
+00:39:28,160 --> 00:39:34,640
+which I think is kind of interesting um
+
+893
+00:39:31,119 --> 00:39:37,040
+is abductive reasoning uh or learning
+
+894
+00:39:34,640 --> 00:39:40,040
+explanations from
+
+895
+00:39:37,040 --> 00:39:40,040
+data
+
+896
+00:39:46,359 --> 00:39:49,359
+and
+
+897
+00:39:52,440 --> 00:39:57,119
+sorry
+
+898
+00:39:54,480 --> 00:40:00,760
+so basically the idea is can we find a
+
+899
+00:39:57,119 --> 00:40:03,599
+rule that underes a pattern in data
+
+900
+00:40:00,760 --> 00:40:06,680
+and here are some examples of this the
+
+901
+00:40:03,599 --> 00:40:11,680
+basic idea is if we have
+
+902
+00:40:06,680 --> 00:40:16,599
+examples um which are like if I put
+
+903
+00:40:11,680 --> 00:40:19,960
+a cylinder and a square a cylinder and a
+
+904
+00:40:16,599 --> 00:40:22,119
+cube on uh this pink block I get a noise
+
+905
+00:40:19,960 --> 00:40:25,440
+if I put just a cylinder on the pink
+
+906
+00:40:22,119 --> 00:40:29,359
+block I don't get a noise and you want
+
+907
+00:40:25,440 --> 00:40:31,800
+to discover underlying rules based on
+
+908
+00:40:29,359 --> 00:40:33,160
+the data that you observed and so why
+
+909
+00:40:31,800 --> 00:40:34,720
+would you want to do this there's a
+
+910
+00:40:33,160 --> 00:40:38,000
+couple reasons why you would want to do
+
+911
+00:40:34,720 --> 00:40:41,560
+this um the first reason why you would
+
+912
+00:40:38,000 --> 00:40:42,920
+like to do this is because um you might
+
+913
+00:40:41,560 --> 00:40:45,119
+want something that you can explain to
+
+914
+00:40:42,920 --> 00:40:47,760
+humans right you can explain I this
+
+915
+00:40:45,119 --> 00:40:51,240
+underlying pattern um exists in this
+
+916
+00:40:47,760 --> 00:40:55,119
+data it explains why the
+
+917
+00:40:51,240 --> 00:40:57,319
+data you know appears as it does appear
+
+918
+00:40:55,119 --> 00:40:59,240
+and then humans can go in and analyze it
+
+919
+00:40:57,319 --> 00:41:02,079
+or something like that so recently
+
+920
+00:40:59,240 --> 00:41:03,880
+there's been a big focus on like using
+
+921
+00:41:02,079 --> 00:41:06,480
+large language models for scientific
+
+922
+00:41:03,880 --> 00:41:08,240
+inquiry and other things like that by
+
+923
+00:41:06,480 --> 00:41:10,920
+coming up with good explanations for why
+
+924
+00:41:08,240 --> 00:41:12,160
+data is the way it is so if we were able
+
+925
+00:41:10,920 --> 00:41:15,599
+to do that that would be really
+
+926
+00:41:12,160 --> 00:41:19,280
+interesting another thing is um language
+
+927
+00:41:15,599 --> 00:41:22,960
+models are not particularly good
+
+928
+00:41:19,280 --> 00:41:24,760
+at coming up with they're not
+
+929
+00:41:22,960 --> 00:41:29,480
+particularly good at being consistent
+
+930
+00:41:24,760 --> 00:41:33,640
+about difficult tasks across very large
+
+931
+00:41:29,480 --> 00:41:35,319
+you know numbers of examples so if you
+
+932
+00:41:33,640 --> 00:41:37,920
+could look at like all of the data at
+
+933
+00:41:35,319 --> 00:41:41,240
+once infer general rules from them put
+
+934
+00:41:37,920 --> 00:41:43,480
+those rules in a prompt and then apply
+
+935
+00:41:41,240 --> 00:41:44,960
+that prompt to make predictions on new
+
+936
+00:41:43,480 --> 00:41:47,880
+examples you might be able to raise your
+
+937
+00:41:44,960 --> 00:41:49,760
+overall accuracy as well so it's kind of
+
+938
+00:41:47,880 --> 00:41:52,480
+like you know that's how humans learn as
+
+939
+00:41:49,760 --> 00:41:55,560
+well right we don't like just memorize
+
+940
+00:41:52,480 --> 00:41:57,400
+each example um if we just look at a few
+
+941
+00:41:55,560 --> 00:41:59,040
+examples then we might you know not
+
+942
+00:41:57,400 --> 00:42:02,560
+generalize well to new examples so we
+
+943
+00:41:59,040 --> 00:42:06,359
+kind of tried to abstract away general
+
+944
+00:42:02,560 --> 00:42:08,160
+rules um so this is also similar to
+
+945
+00:42:06,359 --> 00:42:10,200
+program induction from input output
+
+946
+00:42:08,160 --> 00:42:12,240
+examples which I talked during the code
+
+947
+00:42:10,200 --> 00:42:14,040
+uh generation class so you have like
+
+948
+00:42:12,240 --> 00:42:16,200
+input output examples and from them you
+
+949
+00:42:14,040 --> 00:42:18,119
+would like to come up with uh general
+
+950
+00:42:16,200 --> 00:42:19,920
+rules but this is a little bit more
+
+951
+00:42:18,119 --> 00:42:21,920
+General it doesn't necessarily need to
+
+952
+00:42:19,920 --> 00:42:24,160
+be a program that you're inducing it
+
+953
+00:42:21,920 --> 00:42:25,920
+could be you know a grammar or it could
+
+954
+00:42:24,160 --> 00:42:29,119
+be an explanation or it could be
+
+955
+00:42:25,920 --> 00:42:29,119
+anything else like this
+
+956
+00:42:30,079 --> 00:42:34,680
+um so there's a bit of work on rule
+
+957
+00:42:31,960 --> 00:42:36,800
+induction with llms it's pretty recent
+
+958
+00:42:34,680 --> 00:42:40,200
+work uh but I think it's pretty
+
+959
+00:42:36,800 --> 00:42:43,400
+interesting so the first one is um
+
+960
+00:42:40,200 --> 00:42:45,119
+hypothesis generation or the first step
+
+961
+00:42:43,400 --> 00:42:47,839
+um of this particular work here is
+
+962
+00:42:45,119 --> 00:42:53,280
+hypothesis generation and basically what
+
+963
+00:42:47,839 --> 00:42:55,480
+it does is it takes all of these uh you
+
+964
+00:42:53,280 --> 00:42:58,119
+know input output examples and from
+
+965
+00:42:55,480 --> 00:43:01,680
+these input output examples it predicts
+
+966
+00:42:58,119 --> 00:43:04,720
+these uh rules like the answer is always
+
+967
+00:43:01,680 --> 00:43:06,720
+one or uh you want to pick the smallest
+
+968
+00:43:04,720 --> 00:43:10,839
+one or you want to pick the first
+
+969
+00:43:06,720 --> 00:43:12,880
+element and then you evaluate it um and
+
+970
+00:43:10,839 --> 00:43:14,359
+so you pick the smallest one and you can
+
+971
+00:43:12,880 --> 00:43:16,040
+either evaluate it using another
+
+972
+00:43:14,359 --> 00:43:19,040
+language model or you can evaluate it
+
+973
+00:43:16,040 --> 00:43:21,280
+using symbolic uh using a symbolic
+
+974
+00:43:19,040 --> 00:43:23,359
+evaluator um if it's a program you could
+
+975
+00:43:21,280 --> 00:43:24,680
+use a symbolic evaluator if it's a
+
+976
+00:43:23,359 --> 00:43:28,559
+language model you could just ask the
+
+977
+00:43:24,680 --> 00:43:30,960
+language model to pick you know
+
+978
+00:43:28,559 --> 00:43:33,400
+an answer one always or pick the
+
+979
+00:43:30,960 --> 00:43:35,400
+smallest one or pick the first element
+
+980
+00:43:33,400 --> 00:43:37,480
+and then you get lots of outputs and
+
+981
+00:43:35,400 --> 00:43:39,240
+then when you get lots of outputs you
+
+982
+00:43:37,480 --> 00:43:42,079
+then can compare them against the
+
+983
+00:43:39,240 --> 00:43:44,559
+expected outputs and verify whether the
+
+984
+00:43:42,079 --> 00:43:47,920
+rule is correct verify whether the rule
+
+985
+00:43:44,559 --> 00:43:50,160
+gives you the appropriate answer
+
+986
+00:43:47,920 --> 00:43:53,599
+and once you've done that you can go
+
+987
+00:43:50,160 --> 00:43:56,079
+back and do hypothesis refinement um uh
+
+988
+00:43:53,599 --> 00:43:57,720
+and maybe even give this feedback about
+
+989
+00:43:56,079 --> 00:44:00,079
+like what was wrong
+
+990
+00:43:57,720 --> 00:44:03,280
+and gradually refine you know more
+
+991
+00:44:00,079 --> 00:44:03,280
+accurate and more complex
+
+992
+00:44:04,880 --> 00:44:11,040
+hypothesis this is another variant of
+
+993
+00:44:07,720 --> 00:44:12,760
+this idea um which uses different
+
+994
+00:44:11,040 --> 00:44:14,960
+methodology I think both are completely
+
+995
+00:44:12,760 --> 00:44:17,920
+valid but um this one has a little bit
+
+996
+00:44:14,960 --> 00:44:20,400
+higher data constraints so basically
+
+997
+00:44:17,920 --> 00:44:23,160
+what we do is we use hypotheses in Chain
+
+998
+00:44:20,400 --> 00:44:25,319
+of Thought reasoning and keep ones that
+
+999
+00:44:23,160 --> 00:44:28,480
+give resul in correct
+
+1000
+00:44:25,319 --> 00:44:30,760
+answers so
+
+1001
+00:44:28,480 --> 00:44:35,880
+uh this is the step where they're trying
+
+1002
+00:44:30,760 --> 00:44:40,440
+to induce rules and so here this says um
+
+1003
+00:44:35,880 --> 00:44:42,599
+in base 9 what is 76 + 14 and they used
+
+1004
+00:44:40,440 --> 00:44:44,079
+base 9 here obviously because if it was
+
+1005
+00:44:42,599 --> 00:44:45,520
+in base 10 the language model would just
+
+1006
+00:44:44,079 --> 00:44:48,400
+solve the problem and it's not very
+
+1007
+00:44:45,520 --> 00:44:54,319
+interesting so uh they they did base 9
+
+1008
+00:44:48,400 --> 00:44:55,839
+addition and so the answer is um we have
+
+1009
+00:44:54,319 --> 00:45:00,280
+or the answer provided by the language
+
+1010
+00:44:55,839 --> 00:45:03,319
+model is we have 6 + 4 = 11 um the digit
+
+1011
+00:45:00,280 --> 00:45:07,480
+is 1 and the carry is 1 we have 7 + 1 +
+
+1012
+00:45:03,319 --> 00:45:09,480
+1 = 10 the digit is zero and the is one
+
+1013
+00:45:07,480 --> 00:45:13,000
+a leading digit is one so the answer is
+
+1014
+00:45:09,480 --> 00:45:15,240
+101 um and this verifies so they get the
+
+1015
+00:45:13,000 --> 00:45:17,240
+answer correct and so they know that
+
+1016
+00:45:15,240 --> 00:45:20,800
+they assume that this derivation is also
+
+1017
+00:45:17,240 --> 00:45:25,599
+correct and then they extract particular
+
+1018
+00:45:20,800 --> 00:45:28,200
+rules like 6 + 4 = 11 and 7 + 1 + 1 = 10
+
+1019
+00:45:25,599 --> 00:45:30,800
+um and they add this to the rule
+
+1020
+00:45:28,200 --> 00:45:32,960
+Library so then the question is how do
+
+1021
+00:45:30,800 --> 00:45:35,000
+they extract the rules the way they
+
+1022
+00:45:32,960 --> 00:45:37,920
+extract the rules is they have an in
+
+1023
+00:45:35,000 --> 00:45:40,760
+context prompt which surrounds the rules
+
+1024
+00:45:37,920 --> 00:45:43,520
+by basically XML tags that says this is
+
+1025
+00:45:40,760 --> 00:45:46,640
+a rule that should be extracted and so
+
+1026
+00:45:43,520 --> 00:45:48,400
+then um anything that is in an XML tag
+
+1027
+00:45:46,640 --> 00:45:50,960
+they when you get the correct answer
+
+1028
+00:45:48,400 --> 00:45:53,440
+they extract and add that to the rule
+
+1029
+00:45:50,960 --> 00:45:55,680
+library and then conversely like if the
+
+1030
+00:45:53,440 --> 00:45:57,800
+derivation um if the answer is wrong
+
+1031
+00:45:55,680 --> 00:45:59,920
+they just don't add it or they add it as
+
+1032
+00:45:57,800 --> 00:46:01,079
+a negative example and say this is a
+
+1033
+00:45:59,920 --> 00:46:04,119
+incorrect
+
+1034
+00:46:01,079 --> 00:46:05,839
+rule um and then in the final step where
+
+1035
+00:46:04,119 --> 00:46:07,480
+they do deductive reasoning they can
+
+1036
+00:46:05,839 --> 00:46:09,119
+then go ahead and use these rules and
+
+1037
+00:46:07,480 --> 00:46:11,640
+improve accuracy and they demonstrate
+
+1038
+00:46:09,119 --> 00:46:12,960
+that that helps so basically these are
+
+1039
+00:46:11,640 --> 00:46:14,520
+two different approaches one is
+
+1040
+00:46:12,960 --> 00:46:17,400
+extracting directly from the Chain of
+
+1041
+00:46:14,520 --> 00:46:18,880
+Thought the other is uh a priori trying
+
+1042
+00:46:17,400 --> 00:46:23,760
+to generate rules from the whole rule
+
+1043
+00:46:18,880 --> 00:46:27,480
+base and then um then verifying them um
+
+1044
+00:46:23,760 --> 00:46:31,000
+notably both of these require verifiers
+
+1045
+00:46:27,480 --> 00:46:33,839
+um and so in some recent work which uh I
+
+1046
+00:46:31,000 --> 00:46:36,040
+I hope will be on archive very soon uh
+
+1047
+00:46:33,839 --> 00:46:38,839
+we took a look at whether language
+
+1048
+00:46:36,040 --> 00:46:42,800
+models themselves can verify their own
+
+1049
+00:46:38,839 --> 00:46:46,079
+hypothesis and um so that removes the
+
+1050
+00:46:42,800 --> 00:46:48,000
+symbolic verifier here um by just asking
+
+1051
+00:46:46,079 --> 00:46:51,480
+the language model whether the output is
+
+1052
+00:46:48,000 --> 00:46:53,480
+correct or not and um we found that with
+
+1053
+00:46:51,480 --> 00:46:55,240
+very powerful language models like gp4
+
+1054
+00:46:53,480 --> 00:46:57,760
+you can actually do that as well so that
+
+1055
+00:46:55,240 --> 00:47:01,319
+REM removes the necess necessity to have
+
+1056
+00:46:57,760 --> 00:47:05,480
+a symbolic verifier in the loop as
+
+1057
+00:47:01,319 --> 00:47:08,200
+well cool um the reason why I wanted to
+
+1058
+00:47:05,480 --> 00:47:09,440
+introduce this is I don't know if like
+
+1059
+00:47:08,200 --> 00:47:12,359
+like it seems like all of these have
+
+1060
+00:47:09,440 --> 00:47:16,359
+been applied so far on kind of very toy
+
+1061
+00:47:12,359 --> 00:47:19,119
+examples like you know
+
+1062
+00:47:16,359 --> 00:47:22,240
+um like honestly I don't really care
+
+1063
+00:47:19,119 --> 00:47:25,920
+about whether I can play Tetris or um
+
+1064
+00:47:22,240 --> 00:47:27,920
+you know uh find the largest or smallest
+
+1065
+00:47:25,920 --> 00:47:30,880
+number within
+
+1066
+00:47:27,920 --> 00:47:33,720
+um you know list or something like this
+
+1067
+00:47:30,880 --> 00:47:36,000
+but I think they have like really exting
+
+1068
+00:47:33,720 --> 00:47:38,480
+possibilities for how we could extract
+
+1069
+00:47:36,000 --> 00:47:40,319
+more General patterns and like use these
+
+1070
+00:47:38,480 --> 00:47:41,720
+to improve language model based systems
+
+1071
+00:47:40,319 --> 00:47:43,599
+so I think it's a really exciting
+
+1072
+00:47:41,720 --> 00:47:48,000
+research
+
+1073
+00:47:43,599 --> 00:47:51,000
+Direction um cool any questions about
+
+1074
+00:47:48,000 --> 00:47:51,000
+this
+
+1075
+00:47:54,240 --> 00:48:02,160
+yeah yeah so that's a good question
+
+1076
+00:47:58,160 --> 00:48:06,079
+um so I I think tool
+
+1077
+00:48:02,160 --> 00:48:09,359
+learning is maybe kind of a sub subset
+
+1078
+00:48:06,079 --> 00:48:12,319
+of this possibly like I feel like in
+
+1079
+00:48:09,359 --> 00:48:13,559
+tool learning you're learning functions
+
+1080
+00:48:12,319 --> 00:48:15,559
+that
+
+1081
+00:48:13,559 --> 00:48:17,559
+are I don't know if they are like good
+
+1082
+00:48:15,559 --> 00:48:19,680
+explanations of the data but at the very
+
+1083
+00:48:17,559 --> 00:48:23,119
+least they're like useful um they're
+
+1084
+00:48:19,680 --> 00:48:25,119
+useful rules for solving the task um so
+
+1085
+00:48:23,119 --> 00:48:26,880
+I I feel like they're approaching it
+
+1086
+00:48:25,119 --> 00:48:28,760
+from two different motivations but
+
+1087
+00:48:26,880 --> 00:48:30,960
+actually
+
+1088
+00:48:28,760 --> 00:48:33,559
+the methods that they're using are
+
+1089
+00:48:30,960 --> 00:48:36,240
+similar so like for example in our tool
+
+1090
+00:48:33,559 --> 00:48:38,559
+learning work Trove we generated like
+
+1091
+00:48:36,240 --> 00:48:42,240
+multiple options for tools and we kept
+
+1092
+00:48:38,559 --> 00:48:44,000
+the ones that had high self- consistency
+
+1093
+00:48:42,240 --> 00:48:46,800
+so that's kind of like the verifier step
+
+1094
+00:48:44,000 --> 00:48:49,040
+right and then um we threw away the ones
+
+1095
+00:48:46,800 --> 00:48:52,760
+that weren't useful so that helps make a
+
+1096
+00:48:49,040 --> 00:48:56,760
+concise rule set so
+
+1097
+00:48:52,760 --> 00:48:59,280
+yeah and then like could we use tools to
+
+1098
+00:48:56,760 --> 00:49:01,880
+[Music]
+
+1099
+00:48:59,280 --> 00:49:04,079
+attack kind of the more like conceptual
+
+1100
+00:49:01,880 --> 00:49:05,319
+reasoning stuff I I don't actually know
+
+1101
+00:49:04,079 --> 00:49:06,839
+uh the answer to that it's a good
+
+1102
+00:49:05,319 --> 00:49:10,599
+question
+
+1103
+00:49:06,839 --> 00:49:10,599
+yeah any any other
+
+1104
+00:49:11,240 --> 00:49:18,680
+things okay uh another final one that
+
+1105
+00:49:14,440 --> 00:49:21,680
+I'd like to introduce um this is really
+
+1106
+00:49:18,680 --> 00:49:23,839
+like I I really really like this paper
+
+1107
+00:49:21,680 --> 00:49:27,440
+um just from the point of view of its
+
+1108
+00:49:23,839 --> 00:49:29,880
+ambition and motivation um and
+
+1109
+00:49:27,440 --> 00:49:31,920
+the idea is that they want to learn
+
+1110
+00:49:29,880 --> 00:49:34,440
+differences between text
+
+1111
+00:49:31,920 --> 00:49:36,200
+Collections and why would you want to do
+
+1112
+00:49:34,440 --> 00:49:38,079
+this there's actually a ton of reasons
+
+1113
+00:49:36,200 --> 00:49:39,720
+why you would want to do this but the
+
+1114
+00:49:38,079 --> 00:49:44,720
+the best one that they give
+
+1115
+00:49:39,720 --> 00:49:44,720
+here is actually no sorry maybe I I
+
+1116
+00:49:46,440 --> 00:49:50,359
+didn't okay so this is a less
+
+1117
+00:49:48,480 --> 00:49:53,440
+interesting one the the more interesting
+
+1118
+00:49:50,359 --> 00:49:57,799
+one uh that they give in the paper is um
+
+1119
+00:49:53,440 --> 00:50:00,200
+examples of reports from patients who
+
+1120
+00:49:57,799 --> 00:50:04,200
+took an actual drug and took a
+
+1121
+00:50:00,200 --> 00:50:06,640
+placebo and so patients write about like
+
+1122
+00:50:04,200 --> 00:50:08,400
+their their symptoms or how they felt or
+
+1123
+00:50:06,640 --> 00:50:11,000
+they have checkups or things like that
+
+1124
+00:50:08,400 --> 00:50:13,839
+that are all written in natural language
+
+1125
+00:50:11,000 --> 00:50:16,319
+so one of the things that doctors try to
+
+1126
+00:50:13,839 --> 00:50:18,000
+do is they try to look at all of these
+
+1127
+00:50:16,319 --> 00:50:20,240
+reports and figure out if there's any
+
+1128
+00:50:18,000 --> 00:50:21,880
+like consistent difference between
+
+1129
+00:50:20,240 --> 00:50:25,079
+people who took a placebo and people who
+
+1130
+00:50:21,880 --> 00:50:27,359
+took an actual um actual drug and this
+
+1131
+00:50:25,079 --> 00:50:31,079
+is like a major part of medical trials
+
+1132
+00:50:27,359 --> 00:50:32,960
+right um and so the idea is like given
+
+1133
+00:50:31,079 --> 00:50:35,000
+all of the texts of people who took the
+
+1134
+00:50:32,960 --> 00:50:36,599
+drug given all the texts of people who
+
+1135
+00:50:35,000 --> 00:50:38,319
+of people who took the placebo could you
+
+1136
+00:50:36,599 --> 00:50:40,960
+automatically extract differences
+
+1137
+00:50:38,319 --> 00:50:45,000
+between them in some way and so the
+
+1138
+00:50:40,960 --> 00:50:47,760
+methodology that they use for this is um
+
+1139
+00:50:45,000 --> 00:50:51,359
+they have like group a uh the Manchester
+
+1140
+00:50:47,760 --> 00:50:53,240
+United soccer Squad welcomes Rising Star
+
+1141
+00:50:51,359 --> 00:50:54,599
+as Serena Williams joins the UCLA
+
+1142
+00:50:53,240 --> 00:50:56,920
+women's tennis roster and then you have
+
+1143
+00:50:54,599 --> 00:51:00,200
+like 20 more examples and then here you
+
+1144
+00:50:56,920 --> 00:51:03,480
+have Egypt's President uh at the African
+
+1145
+00:51:00,200 --> 00:51:07,200
+unit Union Summit um and other things
+
+1146
+00:51:03,480 --> 00:51:12,000
+like that in 20 examples uh not seen
+
+1147
+00:51:07,200 --> 00:51:14,359
+here and so then if I asked a question
+
+1148
+00:51:12,000 --> 00:51:16,359
+um the original data set includes news
+
+1149
+00:51:14,359 --> 00:51:18,680
+summaries the two corpora are generated
+
+1150
+00:51:16,359 --> 00:51:21,240
+based on when they were published uh
+
+1151
+00:51:18,680 --> 00:51:24,359
+samples from group a include news from
+
+1152
+00:51:21,240 --> 00:51:27,480
+2007 while samples from Group B include
+
+1153
+00:51:24,359 --> 00:51:29,000
+news from 2008 I'm a joural trying to
+
+1154
+00:51:27,480 --> 00:51:31,240
+understand what topics are popular
+
+1155
+00:51:29,000 --> 00:51:33,440
+across years please write a list of
+
+1156
+00:51:31,240 --> 00:51:35,280
+hypotheses separated by bullet points of
+
+1157
+00:51:33,440 --> 00:51:39,920
+how data points from group a differ from
+
+1158
+00:51:35,280 --> 00:51:42,400
+those of group b um and then formatting
+
+1159
+00:51:39,920 --> 00:51:44,160
+information
+
+1160
+00:51:42,400 --> 00:51:46,960
+um
+
+1161
+00:51:44,160 --> 00:51:49,680
+and so based on the two sentence groups
+
+1162
+00:51:46,960 --> 00:51:50,559
+A and B from the above more sentences in
+
+1163
+00:51:49,680 --> 00:51:53,400
+group
+
+1164
+00:51:50,559 --> 00:51:55,240
+a mention a sports team or mention about
+
+1165
+00:51:53,400 --> 00:51:57,319
+academic relations or things like that
+
+1166
+00:51:55,240 --> 00:51:58,599
+and so what this allows you to do is it
+
+1167
+00:51:57,319 --> 00:52:00,319
+allows you to come up with a whole bunch
+
+1168
+00:51:58,599 --> 00:52:01,400
+of hypotheses about why one might be
+
+1169
+00:52:00,319 --> 00:52:04,920
+better than the
+
+1170
+00:52:01,400 --> 00:52:08,920
+other so the problem with this though is
+
+1171
+00:52:04,920 --> 00:52:10,880
+like because of language model you know
+
+1172
+00:52:08,920 --> 00:52:13,440
+limits number one they might just
+
+1173
+00:52:10,880 --> 00:52:17,119
+hallucinate things and be totally wrong
+
+1174
+00:52:13,440 --> 00:52:19,680
+um number two
+
+1175
+00:52:17,119 --> 00:52:21,040
+the size of the context so that they can
+
+1176
+00:52:19,680 --> 00:52:23,960
+take into account when making this
+
+1177
+00:52:21,040 --> 00:52:26,720
+decision is relatively small so the next
+
+1178
+00:52:23,960 --> 00:52:29,280
+thing that they do is then they have a a
+
+1179
+00:52:26,720 --> 00:52:32,119
+much larger Corpus of
+
+1180
+00:52:29,280 --> 00:52:33,200
+text um with like a thousand examples or
+
+1181
+00:52:32,119 --> 00:52:36,640
+something like
+
+1182
+00:52:33,200 --> 00:52:40,240
+this and then they treat each of these
+
+1183
+00:52:36,640 --> 00:52:42,680
+hypotheses as a
+
+1184
+00:52:40,240 --> 00:52:44,559
+classifier and then they go through all
+
+1185
+00:52:42,680 --> 00:52:47,480
+of the examples from Corpus one which is
+
+1186
+00:52:44,559 --> 00:52:50,480
+like maybe 2000 year 2000 and then
+
+1187
+00:52:47,480 --> 00:52:52,079
+Corpus 2 which is year 2008 and they ask
+
+1188
+00:52:50,480 --> 00:52:55,880
+the language model with respect to all
+
+1189
+00:52:52,079 --> 00:52:58,119
+of them um does this sentence mention a
+
+1190
+00:52:55,880 --> 00:53:01,400
+sports team recording recruiting a new
+
+1191
+00:52:58,119 --> 00:53:04,839
+member um and so you get a
+
+1192
+00:53:01,400 --> 00:53:04,839
+classification for each one of
+
+1193
+00:53:12,359 --> 00:53:17,440
+these and you get a certain number of
+
+1194
+00:53:14,520 --> 00:53:18,799
+ones and zeros and so once you have a
+
+1195
+00:53:17,440 --> 00:53:20,839
+certain number of ones and zeros what's
+
+1196
+00:53:18,799 --> 00:53:24,079
+the next thing that you would do
+
+1197
+00:53:20,839 --> 00:53:24,079
+here any
+
+1198
+00:53:24,880 --> 00:53:30,599
+ideas how do you tell there's like
+
+1199
+00:53:27,359 --> 00:53:30,599
+actually a difference between these
+
+1200
+00:53:36,520 --> 00:53:43,319
+two between two sets
+
+1201
+00:53:39,319 --> 00:53:45,920
+of numbers like one and
+
+1202
+00:53:43,319 --> 00:53:48,680
+zero a hint is you probably had to do
+
+1203
+00:53:45,920 --> 00:53:48,680
+this for assignment
+
+1204
+00:53:53,720 --> 00:53:58,520
+two yeah
+
+1205
+00:53:56,799 --> 00:54:01,200
+yeah exactly you you do a significance
+
+1206
+00:53:58,520 --> 00:54:04,200
+test between the two and so um what you
+
+1207
+00:54:01,200 --> 00:54:06,440
+can then do is you have lots of
+
+1208
+00:54:04,200 --> 00:54:08,839
+hypotheses you have lots of significance
+
+1209
+00:54:06,440 --> 00:54:11,040
+values you can order them by the
+
+1210
+00:54:08,839 --> 00:54:13,839
+significance value and say the most
+
+1211
+00:54:11,040 --> 00:54:17,559
+significance or the the difference with
+
+1212
+00:54:13,839 --> 00:54:19,160
+the like lowest P value between them is
+
+1213
+00:54:17,559 --> 00:54:20,480
+the one that's most likely to be an
+
+1214
+00:54:19,160 --> 00:54:26,520
+actual difference between the two and
+
+1215
+00:54:20,480 --> 00:54:29,079
+you can find um like uh the news in 2007
+
+1216
+00:54:26,520 --> 00:54:32,520
+indeed tended to talk about X more than
+
+1217
+00:54:29,079 --> 00:54:34,559
+uh than other things so I uh I actually
+
+1218
+00:54:32,520 --> 00:54:36,079
+used this in one of my uh one of my
+
+1219
+00:54:34,559 --> 00:54:39,520
+unrelated projects where I wanted to
+
+1220
+00:54:36,079 --> 00:54:42,680
+find the difference between um language
+
+1221
+00:54:39,520 --> 00:54:45,640
+models sentences that language models
+
+1222
+00:54:42,680 --> 00:54:47,839
+aligned well with human brain signals in
+
+1223
+00:54:45,640 --> 00:54:49,760
+sentences where language models didn't
+
+1224
+00:54:47,839 --> 00:54:52,559
+align well with human brain signals so
+
+1225
+00:54:49,760 --> 00:54:53,799
+we like we had some data of human brain
+
+1226
+00:54:52,559 --> 00:54:56,880
+signals and we had a measure of
+
+1227
+00:54:53,799 --> 00:54:58,240
+alignment um on each sentence and it
+
+1228
+00:54:56,880 --> 00:55:01,799
+actually found some pretty interesting
+
+1229
+00:54:58,240 --> 00:55:03,359
+hypothesis like um uh language models
+
+1230
+00:55:01,799 --> 00:55:06,200
+tend to align less well with human brain
+
+1231
+00:55:03,359 --> 00:55:07,319
+signals on metaphorical language or a
+
+1232
+00:55:06,200 --> 00:55:10,599
+language that had to do with
+
+1233
+00:55:07,319 --> 00:55:11,799
+interpersonal relations or um or other
+
+1234
+00:55:10,599 --> 00:55:15,200
+things like that and then we actually
+
+1235
+00:55:11,799 --> 00:55:17,559
+went in and pursued um you know these to
+
+1236
+00:55:15,200 --> 00:55:21,000
+examine them further and uh we didn't
+
+1237
+00:55:17,559 --> 00:55:22,680
+entirely rely on this um you know like
+
+1238
+00:55:21,000 --> 00:55:25,160
+significance test because I didn't quite
+
+1239
+00:55:22,680 --> 00:55:26,880
+trust language models that much to like
+
+1240
+00:55:25,160 --> 00:55:28,559
+shape my entire resource
+
+1241
+00:55:26,880 --> 00:55:29,880
+research agenda around them but we came
+
+1242
+00:55:28,559 --> 00:55:31,720
+up with other ways to measure it and
+
+1243
+00:55:29,880 --> 00:55:35,000
+some of the things checked out some of
+
+1244
+00:55:31,720 --> 00:55:36,799
+the things didn't check out so um again
+
+1245
+00:55:35,000 --> 00:55:38,760
+I think this general direction of like
+
+1246
+00:55:36,799 --> 00:55:41,720
+how can language models help us answer
+
+1247
+00:55:38,760 --> 00:55:43,760
+you know uh complex research questions
+
+1248
+00:55:41,720 --> 00:55:45,480
+that we wouldn't be able to easily or
+
+1249
+00:55:43,760 --> 00:55:47,960
+very efficiently that would require
+
+1250
+00:55:45,480 --> 00:55:52,200
+normally humans annotating lots of data
+
+1251
+00:55:47,960 --> 00:55:56,839
+is um an interesting topic as
+
+1252
+00:55:52,200 --> 00:55:56,839
+well cool um

CMU Advanced NLP 2024 (21) Complex Reasoning/transcript.vtt

@@ -0,0 +1,3757 @@
+WEBVTT
+
+00:00:00.280 --> 00:00:05.120
+so I'd like to go ahead with uh complex
+
+00:00:02.399 --> 00:00:08.719
+reasoning and we've talked a little bit
+
+00:00:05.120 --> 00:00:10.719
+about uh reasoning in language models uh
+
+00:00:08.719 --> 00:00:12.160
+up until now and so I'm going to be
+
+00:00:10.719 --> 00:00:15.280
+talking about stuff that we didn't talk
+
+00:00:12.160 --> 00:00:17.240
+about yet um this might be a little bit
+
+00:00:15.280 --> 00:00:19.199
+short because of that because I'm not
+
+00:00:17.240 --> 00:00:20.640
+talking about like programs because we
+
+00:00:19.199 --> 00:00:22.080
+talked about that in the code generation
+
+00:00:20.640 --> 00:00:24.199
+class and we already talked a little bit
+
+00:00:22.080 --> 00:00:26.320
+about some of the basics here but um you
+
+00:00:24.199 --> 00:00:30.119
+know if we have time at the end I'd be
+
+00:00:26.320 --> 00:00:30.840
+happy to discuss free form also so what
+
+00:00:30.119 --> 00:00:34.320
+is
+
+00:00:30.840 --> 00:00:35.920
+reasoning um the basic idea is using
+
+00:00:34.320 --> 00:00:37.680
+evidence and logic to arrive at
+
+00:00:35.920 --> 00:00:40.200
+conclusions and make
+
+00:00:37.680 --> 00:00:43.760
+judgments
+
+00:00:40.200 --> 00:00:48.039
+and what is it in language models is a
+
+00:00:43.760 --> 00:00:49.399
+little bit um you know less clear uh but
+
+00:00:48.039 --> 00:00:52.680
+if we talk about it kind of like from
+
+00:00:49.399 --> 00:00:56.280
+the philosophical standpoint um there
+
+00:00:52.680 --> 00:00:58.399
+are two varieties of this one is formal
+
+00:00:56.280 --> 00:01:01.680
+uh reasoning and formal reasoning is
+
+00:00:58.399 --> 00:01:04.239
+mostly based on strict truth values so
+
+00:01:01.680 --> 00:01:05.920
+it's kind of like um you can definitely
+
+00:01:04.239 --> 00:01:08.360
+say this is true you can definitely say
+
+00:01:05.920 --> 00:01:11.680
+this is not true
+
+00:01:08.360 --> 00:01:13.799
+and in real life there's very little
+
+00:01:11.680 --> 00:01:15.759
+actual formal reasoning outside of like
+
+00:01:13.799 --> 00:01:17.960
+for example mathematics or maybe you
+
+00:01:15.759 --> 00:01:20.240
+know algorithms computer science and
+
+00:01:17.960 --> 00:01:21.759
+other things like that um and then
+
+00:01:20.240 --> 00:01:23.240
+separately from that we have informal
+
+00:01:21.759 --> 00:01:27.040
+reasoning based on experience and
+
+00:01:23.240 --> 00:01:30.439
+intuition and actually um this is this
+
+00:01:27.040 --> 00:01:32.360
+was uh rather elusive uh until
+
+00:01:30.439 --> 00:01:33.720
+large language models you know people
+
+00:01:32.360 --> 00:01:35.560
+were working on it but it was really
+
+00:01:33.720 --> 00:01:38.119
+hard and this is like one of the big
+
+00:01:35.560 --> 00:01:41.479
+breakthroughs I think of the past few
+
+00:01:38.119 --> 00:01:46.799
+years um I should note that this uh
+
+00:01:41.479 --> 00:01:48.520
+paper here uh hang and Chan is a kind of
+
+00:01:46.799 --> 00:01:50.119
+review survey paper of reasoning in
+
+00:01:48.520 --> 00:01:51.520
+large language models it's on the
+
+00:01:50.119 --> 00:01:54.719
+references so if you're interested you
+
+00:01:51.520 --> 00:01:57.600
+can take a look at that too um but
+
+00:01:54.719 --> 00:01:59.200
+there's three kinds of reasoning uh
+
+00:01:57.600 --> 00:02:00.840
+there's many kinds of reasoning but
+
+00:01:59.200 --> 00:02:03.280
+there's three kinds of reasoning in
+
+00:02:00.840 --> 00:02:06.240
+particular that I'd like to talk about
+
+00:02:03.280 --> 00:02:08.840
+um from the point of view of today and
+
+00:02:06.240 --> 00:02:10.360
+the first one is uh deductive reasoning
+
+00:02:08.840 --> 00:02:13.080
+and deductive reasoning is basically
+
+00:02:10.360 --> 00:02:16.040
+using logic to go from a premise to a
+
+00:02:13.080 --> 00:02:18.440
+conclusion and this is largely what
+
+00:02:16.040 --> 00:02:19.879
+people not entirely but largely what
+
+00:02:18.440 --> 00:02:22.400
+people talk about when they think about
+
+00:02:19.879 --> 00:02:25.879
+formal reasoning and so basically you
+
+00:02:22.400 --> 00:02:28.640
+have several premises um like all
+
+00:02:25.879 --> 00:02:32.120
+mammals have kidneys and all whales are
+
+00:02:28.640 --> 00:02:35.239
+mammals and then from this uh you can go
+
+00:02:32.120 --> 00:02:35.239
+to all whales have
+
+00:02:35.440 --> 00:02:40.640
+kidneys then separately there's
+
+00:02:38.000 --> 00:02:44.040
+inductive reasoning and inductive
+
+00:02:40.640 --> 00:02:46.040
+reasoning is um from
+
+00:02:44.040 --> 00:02:48.480
+observation uh predict a likely
+
+00:02:46.040 --> 00:02:50.080
+conclusion or predict a likely kind of
+
+00:02:48.480 --> 00:02:53.640
+generalized
+
+00:02:50.080 --> 00:02:55.360
+conclusion um so this is one example uh
+
+00:02:53.640 --> 00:02:56.920
+when we see a creature with wings it is
+
+00:02:55.360 --> 00:02:58.599
+usually a bird we see a creature with
+
+00:02:56.920 --> 00:03:00.400
+wings the creature is likely to be a
+
+00:02:58.599 --> 00:03:02.879
+bird so it's kind of this is kind of
+
+00:03:00.400 --> 00:03:05.319
+like a soft version of deduction another
+
+00:03:02.879 --> 00:03:07.440
+common thing is like every single
+
+00:03:05.319 --> 00:03:10.760
+creature I have seen with wings is a
+
+00:03:07.440 --> 00:03:12.480
+bird and then you can kind of um induce
+
+00:03:10.760 --> 00:03:16.799
+that all
+
+00:03:12.480 --> 00:03:19.159
+uh like all uh creatures with wings are
+
+00:03:16.799 --> 00:03:21.120
+birds but that might not be true it's
+
+00:03:19.159 --> 00:03:23.879
+not necessarily logically entailed but
+
+00:03:21.120 --> 00:03:27.560
+you you make that kind
+
+00:03:23.879 --> 00:03:31.000
+of logical conclusion uh without it
+
+00:03:27.560 --> 00:03:32.840
+being formally uh correct or verif
+
+00:03:31.000 --> 00:03:34.720
+and then the final one is abductive
+
+00:03:32.840 --> 00:03:38.000
+reasoning and so this is from an
+
+00:03:34.720 --> 00:03:40.760
+observation we predict the most likely
+
+00:03:38.000 --> 00:03:42.760
+explanation and so for example if we
+
+00:03:40.760 --> 00:03:44.480
+have something like the car cannot start
+
+00:03:42.760 --> 00:03:48.319
+and there is a puddle of liquid under
+
+00:03:44.480 --> 00:03:50.200
+the engine um then we might have a
+
+00:03:48.319 --> 00:03:53.360
+likely explanation that the car has a
+
+00:03:50.200 --> 00:03:55.280
+leak in the radiator so we're going from
+
+00:03:53.360 --> 00:03:58.760
+kind of uh the
+
+00:03:55.280 --> 00:04:00.879
+car you know these these things and then
+
+00:03:58.760 --> 00:04:02.280
+we try to predict the reason why this
+
+00:04:00.879 --> 00:04:05.040
+happens so we're trying to predict like
+
+00:04:02.280 --> 00:04:07.360
+reverse pausal links
+
+00:04:05.040 --> 00:04:08.480
+essentially um there's other types of re
+
+00:04:07.360 --> 00:04:10.400
+reasoning that I'm not going to talk
+
+00:04:08.480 --> 00:04:12.159
+about as much like analogical reasoning
+
+00:04:10.400 --> 00:04:14.079
+and and things like this but uh these
+
+00:04:12.159 --> 00:04:15.440
+are the three main ones I want to talk
+
+00:04:14.079 --> 00:04:17.720
+about
+
+00:04:15.440 --> 00:04:22.040
+today uh one thing I should point out is
+
+00:04:17.720 --> 00:04:24.400
+like even in philosophy or you know
+
+00:04:22.040 --> 00:04:26.240
+like even when you read descriptions
+
+00:04:24.400 --> 00:04:29.280
+about these various types of reasoning
+
+00:04:26.240 --> 00:04:31.880
+the types are a little bit vague so um
+
+00:04:29.280 --> 00:04:35.280
+take these is like
+
+00:04:31.880 --> 00:04:37.240
+general not you know General directions
+
+00:04:35.280 --> 00:04:39.400
+and not strict rules because like which
+
+00:04:37.240 --> 00:04:42.120
+falls on under which category also can
+
+00:04:39.400 --> 00:04:44.880
+be a little bit uh you know unclear uh
+
+00:04:42.120 --> 00:04:44.880
+according to various
+
+00:04:45.479 --> 00:04:53.440
+definitions cool um so first before
+
+00:04:49.840 --> 00:04:55.720
+getting into formal reasoning methods
+
+00:04:53.440 --> 00:04:57.759
+are before getting into the bulk of the
+
+00:04:55.720 --> 00:05:00.000
+talk which is going to be about llms I
+
+00:04:57.759 --> 00:05:02.479
+want to talk about some pre-m reasoning
+
+00:05:00.000 --> 00:05:03.720
+methods and the first one is kind of
+
+00:05:02.479 --> 00:05:05.160
+like formal reasoning within
+
+00:05:03.720 --> 00:05:07.320
+computational
+
+00:05:05.160 --> 00:05:09.840
+semantics and this has been around for a
+
+00:05:07.320 --> 00:05:12.479
+really long time um it's also kind of
+
+00:05:09.840 --> 00:05:15.000
+what powered the things that worked over
+
+00:05:12.479 --> 00:05:21.039
+knowledge bases and other things like
+
+00:05:15.000 --> 00:05:23.639
+this um and the way it works is it does
+
+00:05:21.039 --> 00:05:27.600
+derivational um
+
+00:05:23.639 --> 00:05:31.800
+reasoning by uh sorry I can't read that
+
+00:05:27.600 --> 00:05:34.720
+in the back um by starting out with
+
+00:05:31.800 --> 00:05:36.080
+certain premises and getting to um
+
+00:05:34.720 --> 00:05:40.000
+getting to final
+
+00:05:36.080 --> 00:05:43.039
+conclusions so there's ways that you can
+
+00:05:40.000 --> 00:05:44.060
+write this I think you might have
+
+00:05:43.039 --> 00:05:47.080
+seen
+
+00:05:44.060 --> 00:05:50.479
+[Music]
+
+00:05:47.080 --> 00:05:54.240
+um you might have seen
+
+00:05:50.479 --> 00:05:58.319
+uh this in uh another like math class or
+
+00:05:54.240 --> 00:06:02.440
+something but uh we we have symbols like
+
+00:05:58.319 --> 00:06:02.440
+all and um
+
+00:06:03.039 --> 00:06:08.280
+exist let's
+
+00:06:04.960 --> 00:06:10.960
+see yeah we have things like all and
+
+00:06:08.280 --> 00:06:13.319
+exist and like all
+
+00:06:10.960 --> 00:06:16.240
+X
+
+00:06:13.319 --> 00:06:20.479
+die means
+
+00:06:16.240 --> 00:06:23.919
+like every Everything has died and this
+
+00:06:20.479 --> 00:06:27.360
+uh implies that Mia and Zed have
+
+00:06:23.919 --> 00:06:30.440
+died um
+
+00:06:27.360 --> 00:06:32.240
+so yeah this is a actually maybe I'll
+
+00:06:30.440 --> 00:06:33.280
+not I'll not go through this one and let
+
+00:06:32.240 --> 00:06:37.639
+me go
+
+00:06:33.280 --> 00:06:40.440
+through um go to this one so like it
+
+00:06:37.639 --> 00:06:40.440
+would be something
+
+00:06:40.639 --> 00:06:45.080
+like uh for
+
+00:06:42.960 --> 00:06:47.480
+all
+
+00:06:45.080 --> 00:06:50.669
+X um
+
+00:06:47.480 --> 00:06:50.669
+[Music]
+
+00:06:52.039 --> 00:07:00.400
+mamal well X
+
+00:06:56.759 --> 00:07:03.520
+implies have
+
+00:07:00.400 --> 00:07:07.560
+X kidney or something like
+
+00:07:03.520 --> 00:07:09.280
+that and then you would have other rules
+
+00:07:07.560 --> 00:07:11.879
+and you can go through uh through
+
+00:07:09.280 --> 00:07:14.440
+derivations and and other things like
+
+00:07:11.879 --> 00:07:16.120
+this
+
+00:07:14.440 --> 00:07:19.280
+um
+
+00:07:16.120 --> 00:07:21.560
+my favorite reference for this is this
+
+00:07:19.280 --> 00:07:24.599
+Blackburn and buz book right here it's
+
+00:07:21.560 --> 00:07:26.400
+really well written um and it has like
+
+00:07:24.599 --> 00:07:28.039
+lots of good examples it also explains
+
+00:07:26.400 --> 00:07:30.440
+how you go through derivations and other
+
+00:07:28.039 --> 00:07:34.360
+stuff like that
+
+00:07:30.440 --> 00:07:35.759
+um and actually neural networks can do
+
+00:07:34.360 --> 00:07:37.039
+this variety of reasoning through Chain
+
+00:07:35.759 --> 00:07:38.599
+of Thought and other things I'm going to
+
+00:07:37.039 --> 00:07:40.120
+talk about today but it's a very rough
+
+00:07:38.599 --> 00:07:43.960
+approximation and it doesn't work
+
+00:07:40.120 --> 00:07:47.199
+particularly well for saying like all
+
+00:07:43.960 --> 00:07:51.240
+you know all people
+
+00:07:47.199 --> 00:07:53.599
+are of a uh like things that apply to
+
+00:07:51.240 --> 00:07:57.240
+all people or things that apply to sets
+
+00:07:53.599 --> 00:08:00.039
+or other things like this so within
+
+00:07:57.240 --> 00:08:02.879
+prologue you could
+
+00:08:00.039 --> 00:08:06.520
+take a knowledge base and ask the
+
+00:08:02.879 --> 00:08:11.960
+knowledge base like do
+
+00:08:06.520 --> 00:08:12.800
+all people who work at CMU as professors
+
+00:08:11.960 --> 00:08:15.840
+have a
+
+00:08:12.800 --> 00:08:18.080
+PhD and you could like actually examine
+
+00:08:15.840 --> 00:08:20.639
+that based on the knowledge base uh
+
+00:08:18.080 --> 00:08:23.520
+whereas even if you had
+
+00:08:20.639 --> 00:08:25.800
+a language model that had access to
+
+00:08:23.520 --> 00:08:27.280
+everybody's CVS it wouldn't necessarily
+
+00:08:25.800 --> 00:08:28.599
+be able to answer that question and it
+
+00:08:27.280 --> 00:08:31.440
+especially wouldn't be able to answer
+
+00:08:28.599 --> 00:08:31.440
+that question if you were
+
+00:08:32.320 --> 00:08:37.760
+um it wouldn't be able to answer that
+
+00:08:34.640 --> 00:08:42.880
+question if there were like multiple
+
+00:08:37.760 --> 00:08:46.480
+steps so did all people who are working
+
+00:08:42.880 --> 00:08:50.959
+at CMU get their PHD after
+
+00:08:46.480 --> 00:08:52.959
+19 90 or something like that um so and
+
+00:08:50.959 --> 00:08:54.680
+the answer to that is obviously no but
+
+00:08:52.959 --> 00:08:56.519
+uh this would be able to find the
+
+00:08:54.680 --> 00:08:58.120
+counter evidence to that whereas LMS
+
+00:08:56.519 --> 00:09:00.000
+would not be guaranteed to be able to do
+
+00:08:58.120 --> 00:09:02.800
+that
+
+00:09:00.000 --> 00:09:04.279
+so I I think this is really uh like a
+
+00:09:02.800 --> 00:09:06.760
+nice thing to know but there's a couple
+
+00:09:04.279 --> 00:09:09.600
+problems with it the first thing is this
+
+00:09:06.760 --> 00:09:12.519
+really only traffics in like strictly
+
+00:09:09.600 --> 00:09:17.880
+true or strictly false statements um and
+
+00:09:12.519 --> 00:09:20.560
+that's a really big issue um so like if
+
+00:09:17.880 --> 00:09:22.959
+anything's soft you start uh this sort
+
+00:09:20.560 --> 00:09:24.320
+of formal reasoning starts breaking down
+
+00:09:22.959 --> 00:09:25.880
+the second thing which actually is a
+
+00:09:24.320 --> 00:09:28.959
+really big problem is once you start
+
+00:09:25.880 --> 00:09:30.600
+dealing with more complex things you
+
+00:09:28.959 --> 00:09:32.560
+don't ize it but there's always like
+
+00:09:30.600 --> 00:09:34.560
+exceptions and exceptions to exceptions
+
+00:09:32.560 --> 00:09:36.240
+and other things like that and actually
+
+00:09:34.560 --> 00:09:38.320
+becomes very computationally expensive
+
+00:09:36.240 --> 00:09:41.640
+to prove anything that's kind of like
+
+00:09:38.320 --> 00:09:43.279
+non-trivial um and so because of that uh
+
+00:09:41.640 --> 00:09:45.839
+I'm not actually going to cover it in
+
+00:09:43.279 --> 00:09:47.880
+the lecture today but recently there are
+
+00:09:45.839 --> 00:09:50.880
+um kind of search algorithms through
+
+00:09:47.880 --> 00:09:54.279
+proof spaces that use uh like neural
+
+00:09:50.880 --> 00:09:55.880
+models to do to speed up the search by
+
+00:09:54.279 --> 00:09:58.120
+picking the best and most promising
+
+00:09:55.880 --> 00:10:00.800
+hypotheses and uh for example Sean
+
+00:09:58.120 --> 00:10:02.800
+wellik uh here at CMU is working on that
+
+00:10:00.800 --> 00:10:04.800
+for neural theorem proving where you
+
+00:10:02.800 --> 00:10:05.959
+have uh like mathematical theorem
+
+00:10:04.800 --> 00:10:08.079
+proving and then you use a neural
+
+00:10:05.959 --> 00:10:13.120
+network to pick the best uh paths
+
+00:10:08.079 --> 00:10:14.880
+through logical uh operations so um
+
+00:10:13.120 --> 00:10:19.279
+that's kind of a combination of the more
+
+00:10:14.880 --> 00:10:22.920
+classical and uh modern
+
+00:10:19.279 --> 00:10:26.240
+methods then another thing that's useful
+
+00:10:22.920 --> 00:10:28.079
+to talk about I think this isn't very
+
+00:10:26.240 --> 00:10:31.640
+popular right now but I think it might
+
+00:10:28.079 --> 00:10:34.360
+be become more popular uh in the future
+
+00:10:31.640 --> 00:10:36.120
+is we start hitting the limits of uh you
+
+00:10:34.360 --> 00:10:38.560
+know what we can fit into long context
+
+00:10:36.120 --> 00:10:40.040
+Windows uh for neural models and stuff
+
+00:10:38.560 --> 00:10:42.600
+like this is memory
+
+00:10:40.040 --> 00:10:48.600
+networks and basically the way that
+
+00:10:42.600 --> 00:10:50.639
+memory networks work is they have write
+
+00:10:48.600 --> 00:10:51.399
+they have the ability to write and read
+
+00:10:50.639 --> 00:10:55.639
+from
+
+00:10:51.399 --> 00:10:57.360
+memory and so this figure is a little
+
+00:10:55.639 --> 00:11:00.440
+bit complex here but
+
+00:10:57.360 --> 00:11:02.880
+basically you have a query and then you
+
+00:11:00.440 --> 00:11:04.560
+get the embedding of the query um you
+
+00:11:02.880 --> 00:11:06.760
+take the inner product you get the soft
+
+00:11:04.560 --> 00:11:09.720
+Max of the inner product so this looks
+
+00:11:06.760 --> 00:11:11.040
+like a tension you look up embeddings
+
+00:11:09.720 --> 00:11:12.839
+and you take the weighted Su of the
+
+00:11:11.040 --> 00:11:14.560
+embeddings and you get the like summary
+
+00:11:12.839 --> 00:11:17.680
+of the memor so this is basically
+
+00:11:14.560 --> 00:11:20.320
+attention over a big memory
+
+00:11:17.680 --> 00:11:22.120
+base but then uh memory networks also
+
+00:11:20.320 --> 00:11:24.000
+have the ability to go in and update the
+
+00:11:22.120 --> 00:11:26.639
+memory so they also H have write
+
+00:11:24.000 --> 00:11:30.360
+operations so you can read and write
+
+00:11:26.639 --> 00:11:34.320
+from uh from the memory
+
+00:11:30.360 --> 00:11:36.279
+base and so the reason why I say this
+
+00:11:34.320 --> 00:11:40.480
+might become more popular is like one of
+
+00:11:36.279 --> 00:11:42.200
+the big issues with large language
+
+00:11:40.480 --> 00:11:45.320
+models nowadays is they don't get like
+
+00:11:42.200 --> 00:11:47.320
+to continually update their memory um
+
+00:11:45.320 --> 00:11:50.279
+and like one way you can do that is you
+
+00:11:47.320 --> 00:11:52.160
+can just add text to the memory but
+
+00:11:50.279 --> 00:11:54.000
+there are certain limits to that right
+
+00:11:52.160 --> 00:11:56.360
+uh you know text isn't necessarily the
+
+00:11:54.000 --> 00:11:58.959
+best way to encode all of the things
+
+00:11:56.360 --> 00:12:01.880
+that you've seen in the past so I I feel
+
+00:11:58.959 --> 00:12:03.360
+like this kind of architecture might be
+
+00:12:01.880 --> 00:12:04.920
+um how to pin these sorts of
+
+00:12:03.360 --> 00:12:06.480
+architectures onto language models might
+
+00:12:04.920 --> 00:12:08.639
+be an interesting research direction for
+
+00:12:06.480 --> 00:12:08.639
+the
+
+00:12:08.680 --> 00:12:15.360
+future um another thing which I am not
+
+00:12:12.600 --> 00:12:16.720
+going to talk about very much uh but
+
+00:12:15.360 --> 00:12:20.560
+because we kind of already talked about
+
+00:12:16.720 --> 00:12:23.560
+it in the code Generation Um area but
+
+00:12:20.560 --> 00:12:26.959
+it's actually been around for a while is
+
+00:12:23.560 --> 00:12:30.600
+solving questions with sort of symbolic
+
+00:12:26.959 --> 00:12:36.480
+reasoning and the way it works
+
+00:12:30.600 --> 00:12:41.320
+is for example you would have a
+
+00:12:36.480 --> 00:12:43.639
+um you would have a text here and based
+
+00:12:41.320 --> 00:12:47.440
+on the text you can run these sort of
+
+00:12:43.639 --> 00:12:50.440
+symbolic operations like find and filter
+
+00:12:47.440 --> 00:12:52.720
+and find the max number and relocate and
+
+00:12:50.440 --> 00:12:54.480
+other things like this and this
+
+00:12:52.720 --> 00:12:58.040
+explicitly
+
+00:12:54.480 --> 00:12:59.880
+manipulates uh kind of the attention and
+
+00:12:58.040 --> 00:13:02.519
+the um
+
+00:12:59.880 --> 00:13:03.839
+you can do things like filtering down to
+
+00:13:02.519 --> 00:13:08.600
+find the
+
+00:13:03.839 --> 00:13:11.040
+most uh like highest largest number for
+
+00:13:08.600 --> 00:13:12.800
+example or other things like this and
+
+00:13:11.040 --> 00:13:14.160
+this is kind of interesting because like
+
+00:13:12.800 --> 00:13:17.240
+some of the things that neural networks
+
+00:13:14.160 --> 00:13:20.360
+are bad at are like finding the largest
+
+00:13:17.240 --> 00:13:21.600
+number in a big data set or um finding
+
+00:13:20.360 --> 00:13:23.360
+all of the things where something
+
+00:13:21.600 --> 00:13:26.240
+applies and throwing out all of the
+
+00:13:23.360 --> 00:13:27.959
+things where something doesn't apply so
+
+00:13:26.240 --> 00:13:29.560
+again this isn't used super widely in
+
+00:13:27.959 --> 00:13:31.959
+larged language models right now because
+
+00:13:29.560 --> 00:13:33.920
+I feel like um people have been focusing
+
+00:13:31.959 --> 00:13:36.440
+on prompting
+
+00:13:33.920 --> 00:13:38.880
+techniques uh in order to do this sort
+
+00:13:36.440 --> 00:13:41.199
+of reasoning but I think this is another
+
+00:13:38.880 --> 00:13:43.320
+thing that's worth thinking about taking
+
+00:13:41.199 --> 00:13:45.079
+a close another look at and seeing if
+
+00:13:43.320 --> 00:13:47.440
+there are ways to incorporate it with
+
+00:13:45.079 --> 00:13:49.320
+the current models because like
+
+00:13:47.440 --> 00:13:50.720
+basically what I wanted to say is like
+
+00:13:49.320 --> 00:13:52.279
+all of the things that I decided to
+
+00:13:50.720 --> 00:13:54.560
+introduce here in this section are
+
+00:13:52.279 --> 00:13:57.600
+things that current models are still not
+
+00:13:54.560 --> 00:14:00.800
+particularly good at like Reon taking
+
+00:13:57.600 --> 00:14:03.079
+many steps over sets of
+
+00:14:00.800 --> 00:14:05.079
+inputs um reading and writing from
+
+00:14:03.079 --> 00:14:09.839
+memory so that you can remember things
+
+00:14:05.079 --> 00:14:11.720
+over long periods and also um filtering
+
+00:14:09.839 --> 00:14:13.399
+down large pieces of text into smaller
+
+00:14:11.720 --> 00:14:16.040
+pieces of text to find relevant
+
+00:14:13.399 --> 00:14:17.560
+information so um if any of those things
+
+00:14:16.040 --> 00:14:19.880
+sound interesting you can take a look at
+
+00:14:17.560 --> 00:14:22.800
+this but um after this I'd like to go
+
+00:14:19.880 --> 00:14:24.399
+kind of into the you know main event
+
+00:14:22.800 --> 00:14:27.759
+where I talk about the stuff that people
+
+00:14:24.399 --> 00:14:31.040
+are actually using a lot no it is um any
+
+00:14:27.759 --> 00:14:31.040
+questions about these three
+
+00:14:33.000 --> 00:14:39.120
+okay cool um so now I'd like to go into
+
+00:14:36.399 --> 00:14:40.639
+Chain of Thought and variance and I
+
+00:14:39.120 --> 00:14:42.279
+actually have already talked about Chain
+
+00:14:40.639 --> 00:14:44.199
+of Thought in fact we've mentioned it a
+
+00:14:42.279 --> 00:14:47.720
+couple times um but just you know to
+
+00:14:44.199 --> 00:14:49.399
+remind everybody the basic idea is um
+
+00:14:47.720 --> 00:14:52.880
+compared to standard prompting where we
+
+00:14:49.399 --> 00:14:55.519
+have like a question um and an answer in
+
+00:14:52.880 --> 00:14:58.480
+Chain of Thought we have a question and
+
+00:14:55.519 --> 00:15:01.040
+then we have a derivation for the
+
+00:14:58.480 --> 00:15:02.440
+questions so like uh Roger started with
+
+00:15:01.040 --> 00:15:06.120
+five
+
+00:15:02.440 --> 00:15:09.040
+balls two can uh five balls two cans of
+
+00:15:06.120 --> 00:15:13.839
+three tennis balls each is six tenis 5
+
+00:15:09.040 --> 00:15:15.639
+plus 6al 11 the answer is 11 so um you
+
+00:15:13.839 --> 00:15:17.519
+add this to the prompt and by adding
+
+00:15:15.639 --> 00:15:19.240
+this to the prompt you get the model to
+
+00:15:17.519 --> 00:15:22.600
+uh also do these derivations at test
+
+00:15:19.240 --> 00:15:25.199
+time and this greatly improves some
+
+00:15:22.600 --> 00:15:27.759
+tasks it improves tasks where we can't
+
+00:15:25.199 --> 00:15:30.040
+like immediately predict the answer
+
+00:15:27.759 --> 00:15:32.000
+directly and then I also previously
+
+00:15:30.040 --> 00:15:33.440
+talked about zero shot Chain of Thought
+
+00:15:32.000 --> 00:15:35.880
+uh reasoning where we just prompt the
+
+00:15:33.440 --> 00:15:38.480
+model to with something like let's think
+
+00:15:35.880 --> 00:15:42.680
+step by step and then the model becomes
+
+00:15:38.480 --> 00:15:46.240
+able to do this uh Chain of Thought
+
+00:15:42.680 --> 00:15:48.279
+reasoning okay so that was review and
+
+00:15:46.240 --> 00:15:51.680
+now I'd like to talk about some of like
+
+00:15:48.279 --> 00:15:53.560
+more advanced methods that people use
+
+00:15:51.680 --> 00:15:55.079
+for uh reasoning as
+
+00:15:53.560 --> 00:15:58.040
+well
+
+00:15:55.079 --> 00:15:59.959
+and this is by no means an exhaustive
+
+00:15:58.040 --> 00:16:01.800
+list they're just of the ones that I
+
+00:15:59.959 --> 00:16:03.319
+found interesting so if you know other
+
+00:16:01.800 --> 00:16:04.839
+ones that you'd like to talk about or
+
+00:16:03.319 --> 00:16:07.720
+introduce to the class or something like
+
+00:16:04.839 --> 00:16:10.600
+that I'd also be happy to uh to hear uh
+
+00:16:07.720 --> 00:16:14.120
+which ones you like or have heard about
+
+00:16:10.600 --> 00:16:16.920
+but the first one is um self-as and one
+
+00:16:14.120 --> 00:16:20.959
+of the issues with large language models
+
+00:16:16.920 --> 00:16:23.240
+nowadays is that they're not uh very
+
+00:16:20.959 --> 00:16:25.519
+good at asking follow-up questions or
+
+00:16:23.240 --> 00:16:27.839
+maybe not that they're not very good at
+
+00:16:25.519 --> 00:16:31.160
+it but just they're not trained to do it
+
+00:16:27.839 --> 00:16:32.880
+so like if you play around with chat GPT
+
+00:16:31.160 --> 00:16:35.240
+I have never had chat GPT ask me a
+
+00:16:32.880 --> 00:16:36.680
+follow-up question I don't think it's
+
+00:16:35.240 --> 00:16:38.319
+like it's not because large language
+
+00:16:36.680 --> 00:16:41.920
+models aren't capable of doing it it's
+
+00:16:38.319 --> 00:16:43.519
+just that they like the open AI must
+
+00:16:41.920 --> 00:16:45.000
+think it's a bad user experience to have
+
+00:16:43.519 --> 00:16:47.680
+a language model that asks you follow up
+
+00:16:45.000 --> 00:16:51.319
+questions that's only like you know
+
+00:16:47.680 --> 00:16:53.160
+reason I can think about it um but
+
+00:16:51.319 --> 00:16:56.199
+basically what self ask does is it
+
+00:16:53.160 --> 00:17:00.000
+explicitly prompts the model to ask to
+
+00:16:56.199 --> 00:17:02.360
+ask if there are followup questions so
+
+00:17:00.000 --> 00:17:05.799
+here's an example on the left where the
+
+00:17:02.360 --> 00:17:11.240
+question is uh who lived longer Theodore
+
+00:17:05.799 --> 00:17:12.640
+haer or Harry vau wat uh Watkins and
+
+00:17:11.240 --> 00:17:15.240
+basically it says are follow-up
+
+00:17:12.640 --> 00:17:17.679
+questions needed here yes and then the
+
+00:17:15.240 --> 00:17:20.319
+followup is how old was Theodore hacker
+
+00:17:17.679 --> 00:17:23.640
+when he died and the intermediate answer
+
+00:17:20.319 --> 00:17:26.959
+is Theodore hacker was 65 years old how
+
+00:17:23.640 --> 00:17:29.000
+old was Harry Von Watkins um Harry Von
+
+00:17:26.959 --> 00:17:32.400
+Watkins was 69 years old but so the
+
+00:17:29.000 --> 00:17:35.240
+final answer is Harry Bon Watkins and um
+
+00:17:32.400 --> 00:17:37.520
+in this particular paper this is just
+
+00:17:35.240 --> 00:17:42.520
+like another variety of Chain of Thought
+
+00:17:37.520 --> 00:17:44.720
+it's like not using it to incorporate
+
+00:17:42.520 --> 00:17:47.400
+any external information or anything
+
+00:17:44.720 --> 00:17:48.720
+like that it's just trying to more
+
+00:17:47.400 --> 00:17:52.360
+directly
+
+00:17:48.720 --> 00:17:53.840
+elicit um information from the model um
+
+00:17:52.360 --> 00:17:55.360
+but nonetheless they demonstrate that
+
+00:17:53.840 --> 00:17:57.760
+this is useful and then there's also
+
+00:17:55.360 --> 00:18:00.120
+other methods that actually try to look
+
+00:17:57.760 --> 00:18:02.240
+up information explicit to answer these
+
+00:18:00.120 --> 00:18:05.280
+questions um which are even more
+
+00:18:02.240 --> 00:18:05.280
+powerful than what we have
+
+00:18:05.720 --> 00:18:13.200
+here um so that's what I'd like to
+
+00:18:09.960 --> 00:18:16.919
+introduce next and basically the idea um
+
+00:18:13.200 --> 00:18:19.760
+here is this is a method that instead of
+
+00:18:16.919 --> 00:18:22.880
+just doing Chain of Thought it retrieves
+
+00:18:19.760 --> 00:18:25.480
+relevant sentences when you're doing the
+
+00:18:22.880 --> 00:18:28.919
+Chain of Thought So like
+
+00:18:25.480 --> 00:18:30.880
+here um
+
+00:18:28.919 --> 00:18:32.960
+uh we have the followup are follow-ups
+
+00:18:30.880 --> 00:18:35.159
+needed here yes and then this is the
+
+00:18:32.960 --> 00:18:36.880
+followup but if the model itself doesn't
+
+00:18:35.159 --> 00:18:39.440
+know how old somebody was when they died
+
+00:18:36.880 --> 00:18:40.760
+then it won't be able to answer this so
+
+00:18:39.440 --> 00:18:44.400
+what they do in order to make this
+
+00:18:40.760 --> 00:18:47.200
+happen is they um do bm25 based
+
+00:18:44.400 --> 00:18:49.520
+retrieval over Wikipedia for each of the
+
+00:18:47.200 --> 00:18:51.760
+Chain of Thought uh answers and then
+
+00:18:49.520 --> 00:18:53.400
+they use the retrieved uh I think it's
+
+00:18:51.760 --> 00:18:56.039
+like 10 documents or something like that
+
+00:18:53.400 --> 00:18:59.640
+multiple retriev documents to prompt the
+
+00:18:56.039 --> 00:19:03.200
+model um to basically follow up with its
+
+00:18:59.640 --> 00:19:05.440
+Chain of Thought so this is another uh
+
+00:19:03.200 --> 00:19:07.880
+variety of things that you can do in
+
+00:19:05.440 --> 00:19:07.880
+order to
+
+00:19:10.720 --> 00:19:16.120
+improve
+
+00:19:13.120 --> 00:19:16.120
+cool
+
+00:19:16.400 --> 00:19:21.440
+um then another one that I'd like to
+
+00:19:18.960 --> 00:19:22.559
+talk about is U multilingual Chain of
+
+00:19:21.440 --> 00:19:24.039
+Thought reasoning I'm going to be
+
+00:19:22.559 --> 00:19:28.000
+talking more about multilingual things
+
+00:19:24.039 --> 00:19:29.960
+in the multilingual class in a week but
+
+00:19:28.000 --> 00:19:33.559
+the interesting thing about multilingual
+
+00:19:29.960 --> 00:19:37.200
+Chain of Thought is we have a design
+
+00:19:33.559 --> 00:19:41.280
+decision right like do we want to just
+
+00:19:37.200 --> 00:19:44.000
+answer questions in the language that we
+
+00:19:41.280 --> 00:19:46.679
+are asking questions in like so if I ask
+
+00:19:44.000 --> 00:19:48.080
+a question in Japanese am I going to
+
+00:19:46.679 --> 00:19:49.840
+have it go through the whole chain of
+
+00:19:48.080 --> 00:19:52.720
+thought process in Japanese and then
+
+00:19:49.840 --> 00:19:55.840
+answer my question in Japanese or do I
+
+00:19:52.720 --> 00:19:57.120
+want it to uh somehow go through English
+
+00:19:55.840 --> 00:19:59.159
+because the model has been trained on
+
+00:19:57.120 --> 00:20:00.640
+lots of English and it has better
+
+00:19:59.159 --> 00:20:02.120
+it's like a better way to take advantage
+
+00:20:00.640 --> 00:20:04.840
+of its reasoning
+
+00:20:02.120 --> 00:20:07.200
+capabilities does anyone have a idea
+
+00:20:04.840 --> 00:20:07.200
+about the
+
+00:20:07.960 --> 00:20:12.480
+answer who thinks it's better to do it
+
+00:20:10.240 --> 00:20:15.360
+entirely in the the language that the
+
+00:20:12.480 --> 00:20:15.360
+question is asked
+
+00:20:15.640 --> 00:20:20.080
+in and who thinks it's better to do
+
+00:20:17.919 --> 00:20:23.000
+something in
+
+00:20:20.080 --> 00:20:28.200
+English
+
+00:20:23.000 --> 00:20:29.159
+okay so um basically the answer is do it
+
+00:20:28.200 --> 00:20:31.440
+in English
+
+00:20:29.159 --> 00:20:34.120
+um and maybe this
+
+00:20:31.440 --> 00:20:35.799
+is it might be a little bit dependent on
+
+00:20:34.120 --> 00:20:39.840
+the language but all of the languages
+
+00:20:35.799 --> 00:20:42.880
+they tested it's essentially uh that's
+
+00:20:39.840 --> 00:20:44.919
+the conclusion that they came to and
+
+00:20:42.880 --> 00:20:47.679
+it's pretty Stark in this particular
+
+00:20:44.919 --> 00:20:50.640
+paper this might change a little bit
+
+00:20:47.679 --> 00:20:52.960
+with um with more powerful models but I
+
+00:20:50.640 --> 00:20:57.360
+still would be very surprised if this is
+
+00:20:52.960 --> 00:21:00.440
+not like if this doesn't hold still so
+
+00:20:57.360 --> 00:21:04.440
+you can see it's like approximately on
+
+00:21:00.440 --> 00:21:08.200
+average uh 7even Point increase in the
+
+00:21:04.440 --> 00:21:11.720
+results and just to to be clear here um
+
+00:21:08.200 --> 00:21:13.600
+we have native uh Chain of Thought So
+
+00:21:11.720 --> 00:21:16.039
+This is doing Chain of Thought in the in
+
+00:21:13.600 --> 00:21:17.799
+the language itself this is doing Chain
+
+00:21:16.039 --> 00:21:19.240
+of Thought in English but then answering
+
+00:21:17.799 --> 00:21:22.200
+in the language itself and this is just
+
+00:21:19.240 --> 00:21:23.799
+like translating everything into
+
+00:21:22.200 --> 00:21:27.440
+English
+
+00:21:23.799 --> 00:21:30.159
+um you can try this out too like if you
+
+00:21:27.440 --> 00:21:31.840
+uh if you speak another Lang you can um
+
+00:21:30.159 --> 00:21:34.200
+try to do it myself when I try it in
+
+00:21:31.840 --> 00:21:36.200
+Japanese it's very clear that like the
+
+00:21:34.200 --> 00:21:38.640
+model seems more intelligent in English
+
+00:21:36.200 --> 00:21:41.559
+it just can seems like it can do other
+
+00:21:38.640 --> 00:21:43.120
+things even though like intelligence uh
+
+00:21:41.559 --> 00:21:44.640
+shouldn't be a function of the language
+
+00:21:43.120 --> 00:21:47.120
+that you're asking a question in right
+
+00:21:44.640 --> 00:21:49.679
+like the model should have the ability
+
+00:21:47.120 --> 00:21:51.440
+to answer questions but it because
+
+00:21:49.679 --> 00:21:53.000
+that's how humans work right our
+
+00:21:51.440 --> 00:21:54.520
+intelligence is kind of separated from
+
+00:21:53.000 --> 00:21:57.039
+our language how well we can express
+
+00:21:54.520 --> 00:22:00.480
+ourselves is a little bit different but
+
+00:21:57.039 --> 00:22:02.320
+um yeah for the final appli this was it
+
+00:22:00.480 --> 00:22:04.840
+translated back to the original language
+
+00:22:02.320 --> 00:22:09.440
+and then evaluated for translate English
+
+00:22:04.840 --> 00:22:12.559
+I'm not 100% sure about this I think it
+
+00:22:09.440 --> 00:22:13.840
+was not so that might be a confounding
+
+00:22:12.559 --> 00:22:16.799
+factor for this one but it's not a
+
+00:22:13.840 --> 00:22:20.039
+confounding factor for this one anyway
+
+00:22:16.799 --> 00:22:20.039
+yeah any other
+
+00:22:20.679 --> 00:22:23.919
+questions Okay
+
+00:22:24.200 --> 00:22:29.559
+cool so this is a pretty interesting
+
+00:22:26.799 --> 00:22:32.000
+result here um
+
+00:22:29.559 --> 00:22:34.120
+and the next kind of series of results
+
+00:22:32.000 --> 00:22:35.360
+are going to be based on the uh that I'm
+
+00:22:34.120 --> 00:22:36.919
+going to talk about are going to be
+
+00:22:35.360 --> 00:22:39.240
+based on the quality of the reasoning
+
+00:22:36.919 --> 00:22:43.480
+chains that the model uses in Chain of
+
+00:22:39.240 --> 00:22:45.520
+Thought and this one is a simple
+
+00:22:43.480 --> 00:22:46.600
+heuristic for improving the quality of
+
+00:22:45.520 --> 00:22:49.279
+the reasoning
+
+00:22:46.600 --> 00:22:50.640
+chains and um yeah one thing I should
+
+00:22:49.279 --> 00:22:52.480
+mention is that the quality of the
+
+00:22:50.640 --> 00:22:55.760
+reasoning chain is definitely connected
+
+00:22:52.480 --> 00:22:58.080
+to the uh quality of the output like
+
+00:22:55.760 --> 00:23:00.159
+some that's not necessarily the case
+
+00:22:58.080 --> 00:23:04.679
+right it could just say a whole bunch of
+
+00:23:00.159 --> 00:23:07.799
+you know false like uh actually no maybe
+
+00:23:04.679 --> 00:23:07.799
+I'll I'll skip this
+
+00:23:08.200 --> 00:23:14.919
+one and go and and explain this one next
+
+00:23:11.919 --> 00:23:14.919
+so
+
+00:23:15.159 --> 00:23:19.039
+um yeah actually sorry the or the
+
+00:23:17.600 --> 00:23:20.520
+explanation ordering for this is a
+
+00:23:19.039 --> 00:23:25.360
+little bit hard but yeah I'll explain
+
+00:23:20.520 --> 00:23:26.840
+this one next so um very quickly um
+
+00:23:25.360 --> 00:23:29.640
+there's two ways that you could be
+
+00:23:26.840 --> 00:23:32.880
+reasoning one way you could be reasoning
+
+00:23:29.640 --> 00:23:35.000
+is doing an explanation first and then
+
+00:23:32.880 --> 00:23:36.720
+uh predicting the answer the other way
+
+00:23:35.000 --> 00:23:39.080
+you could do it is predicting the answer
+
+00:23:36.720 --> 00:23:43.039
+and then do it um then giving the
+
+00:23:39.080 --> 00:23:45.559
+explanation and in general if you have a
+
+00:23:43.039 --> 00:23:47.919
+reasonably strong model uh you know any
+
+00:23:45.559 --> 00:23:50.679
+of the modern kind of Frontier level
+
+00:23:47.919 --> 00:23:52.240
+models right now doing the explanation
+
+00:23:50.679 --> 00:23:54.039
+first and then making the prediction is
+
+00:23:52.240 --> 00:23:56.880
+better and the reason why is because
+
+00:23:54.039 --> 00:23:59.240
+Chain of Thought works and the model is
+
+00:23:56.880 --> 00:24:02.960
+able to break down the quest um the
+
+00:23:59.240 --> 00:24:07.279
+questions into kind of
+
+00:24:02.960 --> 00:24:10.159
+simpler uh it's able to break down the
+
+00:24:07.279 --> 00:24:11.520
+like the answer into like simp simpler
+
+00:24:10.159 --> 00:24:14.080
+questions for like mathematical
+
+00:24:11.520 --> 00:24:15.679
+reasoning or something like that um and
+
+00:24:14.080 --> 00:24:18.039
+then give me the answer so like for
+
+00:24:15.679 --> 00:24:20.000
+example for text DCI 002 which was State
+
+00:24:18.039 --> 00:24:22.679
+ofth art at the time of this writing you
+
+00:24:20.000 --> 00:24:24.360
+see a fivepoint boost from using um
+
+00:24:22.679 --> 00:24:29.080
+explanation first and then prediction
+
+00:24:24.360 --> 00:24:30.640
+after that um and in accur
+
+00:24:29.080 --> 00:24:34.039
+but for the weaker models that was not
+
+00:24:30.640 --> 00:24:36.039
+the case so if you were using um GPD 3
+
+00:24:34.039 --> 00:24:38.720
+that wasn't trained for Chain of Thought
+
+00:24:36.039 --> 00:24:40.600
+or you were using opt uh that was not
+
+00:24:38.720 --> 00:24:42.640
+the case but nowadays I think basically
+
+00:24:40.600 --> 00:24:45.279
+all models uh doing the explanation
+
+00:24:42.640 --> 00:24:48.120
+first and then the prediction is
+
+00:24:45.279 --> 00:24:49.640
+better um so going
+
+00:24:48.120 --> 00:24:51.640
+back
+
+00:24:49.640 --> 00:24:53.559
+um another thing that people have
+
+00:24:51.640 --> 00:24:55.120
+noticed is like if your explanation is
+
+00:24:53.559 --> 00:24:56.520
+wrong your prediction also tends to be
+
+00:24:55.120 --> 00:24:58.120
+wrong so if you make mistakes in
+
+00:24:56.520 --> 00:25:00.520
+intermediate steps of your explanation
+
+00:24:58.120 --> 00:25:03.679
+it's tends to mess up your final
+
+00:25:00.520 --> 00:25:06.000
+prediction um so like one of the
+
+00:25:03.679 --> 00:25:09.320
+interesting ways that people have found
+
+00:25:06.000 --> 00:25:11.559
+to improve the final the explanation
+
+00:25:09.320 --> 00:25:13.880
+quality is they just observe that if the
+
+00:25:11.559 --> 00:25:18.840
+explanations are longer they tend to be
+
+00:25:13.880 --> 00:25:20.960
+better it's uh kind of interesting but
+
+00:25:18.840 --> 00:25:23.000
+like if they give you more reasoning
+
+00:25:20.960 --> 00:25:25.000
+steps this tends to be more accurate and
+
+00:25:23.000 --> 00:25:27.320
+they actually demonstrate that in this
+
+00:25:25.000 --> 00:25:29.200
+paper where here's a simple reasoning
+
+00:25:27.320 --> 00:25:31.720
+chain here's a more complex reasoning
+
+00:25:29.200 --> 00:25:35.480
+chain and you actually see for exactly
+
+00:25:31.720 --> 00:25:36.760
+the same problem they get about a 15%
+
+00:25:35.480 --> 00:25:38.360
+boost and these are kind of like
+
+00:25:36.760 --> 00:25:39.960
+naturally occurring reasoning chains
+
+00:25:38.360 --> 00:25:41.520
+they didn't like train the model to give
+
+00:25:39.960 --> 00:25:43.919
+you longer reasoning chains or anything
+
+00:25:41.520 --> 00:25:45.279
+like that but amongst the naturally
+
+00:25:43.919 --> 00:25:46.840
+occurring reasoning chains the longer
+
+00:25:45.279 --> 00:25:50.480
+ones tend to be
+
+00:25:46.840 --> 00:25:53.159
+better and this fact could be simply
+
+00:25:50.480 --> 00:25:54.679
+used to improve accuracy um and so the
+
+00:25:53.159 --> 00:25:57.360
+way they did this is they just sampled
+
+00:25:54.679 --> 00:25:59.279
+multiple reasoning paths and then they
+
+00:25:57.360 --> 00:26:00.840
+performed self consistency over the
+
+00:25:59.279 --> 00:26:03.000
+longer reasoning paths so if you
+
+00:26:00.840 --> 00:26:05.240
+remember what self consistency is it's
+
+00:26:03.000 --> 00:26:07.240
+basically like you do majority voting
+
+00:26:05.240 --> 00:26:09.679
+over the answers for multiple reasoning
+
+00:26:07.240 --> 00:26:13.880
+paths so they threw out the lower
+
+00:26:09.679 --> 00:26:13.880
+quality ones and that improved overall
+
+00:26:14.399 --> 00:26:20.279
+accuracy so um yeah that's a thing that
+
+00:26:18.000 --> 00:26:20.279
+you can
+
+00:26:21.039 --> 00:26:25.960
+do
+
+00:26:23.120 --> 00:26:28.880
+um so yeah going back to systematic
+
+00:26:25.960 --> 00:26:31.360
+studies of reasoning in llms
+
+00:26:28.880 --> 00:26:33.559
+um one of the big results that's
+
+00:26:31.360 --> 00:26:35.880
+actually really important to know about
+
+00:26:33.559 --> 00:26:39.039
+is th this sort of Chain of Thought
+
+00:26:35.880 --> 00:26:41.080
+reasoning um is considered to be an
+
+00:26:39.039 --> 00:26:43.520
+emergent ability
+
+00:26:41.080 --> 00:26:47.080
+in uh large language models and what we
+
+00:26:43.520 --> 00:26:49.360
+mean by an emergent ability is it's or
+
+00:26:47.080 --> 00:26:53.679
+what what the the name emergent ability
+
+00:26:49.360 --> 00:26:56.399
+typically refers to is that it is
+
+00:26:53.679 --> 00:26:58.640
+something that increases dramatically as
+
+00:26:56.399 --> 00:27:01.679
+the model size gets uh up up to a
+
+00:26:58.640 --> 00:27:03.200
+certain point so these actually I'm I'm
+
+00:27:01.679 --> 00:27:06.080
+really sorry I cut off the thing on the
+
+00:27:03.200 --> 00:27:07.360
+bottom here this is like open AI does
+
+00:27:06.080 --> 00:27:08.520
+this all the time to not tell you how
+
+00:27:07.360 --> 00:27:11.399
+many parameters they have in their
+
+00:27:08.520 --> 00:27:12.760
+models but I did not do it intentionally
+
+00:27:11.399 --> 00:27:15.360
+here because I think it's actually in
+
+00:27:12.760 --> 00:27:17.320
+here in the paper um but like these ones
+
+00:27:15.360 --> 00:27:19.399
+over here are kind of the like 175
+
+00:27:17.320 --> 00:27:20.640
+billion parameter models and like the
+
+00:27:19.399 --> 00:27:24.520
+the larger
+
+00:27:20.640 --> 00:27:25.960
+models um and what you see is like up
+
+00:27:24.520 --> 00:27:29.919
+until a certain point you get basically
+
+00:27:25.960 --> 00:27:33.919
+zero accuracy and then uh the outputs
+
+00:27:29.919 --> 00:27:37.000
+improve and so for a while people were
+
+00:27:33.919 --> 00:27:39.240
+really like confused about this like why
+
+00:27:37.000 --> 00:27:41.440
+why does this happen it feels like magic
+
+00:27:39.240 --> 00:27:44.279
+that you get a really you know powerful
+
+00:27:41.440 --> 00:27:46.679
+model and then suddenly it gets better
+
+00:27:44.279 --> 00:27:49.799
+uh uh like at the very
+
+00:27:46.679 --> 00:27:52.159
+end but actually there's a much simpler
+
+00:27:49.799 --> 00:27:53.760
+solution there's not not that much magic
+
+00:27:52.159 --> 00:27:55.960
+to this
+
+00:27:53.760 --> 00:27:58.399
+and we've known about this for a little
+
+00:27:55.960 --> 00:28:00.919
+while but this paper from 2023 really
+
+00:27:58.399 --> 00:28:02.360
+like expressed it very clearly um so I
+
+00:28:00.919 --> 00:28:04.360
+highly recommend you take a look at this
+
+00:28:02.360 --> 00:28:07.720
+if you're interested in kind of like the
+
+00:28:04.360 --> 00:28:10.159
+emerg abilities and language models but
+
+00:28:07.720 --> 00:28:15.039
+basically the the thing about emergent
+
+00:28:10.159 --> 00:28:19.720
+abilities is that they're mostly
+
+00:28:15.039 --> 00:28:20.720
+a matter of how you um how you measure
+
+00:28:19.720 --> 00:28:22.519
+your
+
+00:28:20.720 --> 00:28:27.640
+models
+
+00:28:22.519 --> 00:28:30.120
+accuracy and so let's say as your model
+
+00:28:27.640 --> 00:28:30.120
+gets better
+
+00:28:39.039 --> 00:28:45.600
+it gets gradually better at predicting
+
+00:28:41.200 --> 00:28:45.600
+the like a reasonable next
+
+00:28:47.799 --> 00:28:54.760
+token so this is like a I don't know
+
+00:28:50.919 --> 00:28:59.120
+like 200 million parameter model 500
+
+00:28:54.760 --> 00:29:03.240
+million 1 billion 3 billion
+
+00:28:59.120 --> 00:29:06.600
+7 billion and like 70 billion or
+
+00:29:03.240 --> 00:29:09.600
+something like that um and so this is
+
+00:29:06.600 --> 00:29:12.640
+like the next token prediction accuracy
+
+00:29:09.600 --> 00:29:14.320
+um or like the the accuracy of
+
+00:29:12.640 --> 00:29:16.279
+predicting a reasonable next token that
+
+00:29:14.320 --> 00:29:18.880
+won't make result in your reasoning
+
+00:29:16.279 --> 00:29:20.000
+chain being wrong and making a mistake
+
+00:29:18.880 --> 00:29:24.200
+and
+
+00:29:20.000 --> 00:29:26.200
+so if you have an accuracy like this in
+
+00:29:24.200 --> 00:29:28.880
+order to get the correct answer like
+
+00:29:26.200 --> 00:29:30.559
+let's say there's about five or eight
+
+00:29:28.880 --> 00:29:33.519
+places where you could possibly make a
+
+00:29:30.559 --> 00:29:35.080
+mistake in the derivation like one
+
+00:29:33.519 --> 00:29:36.760
+common places to make a mistake in a
+
+00:29:35.080 --> 00:29:38.519
+derivation for math for example are
+
+00:29:36.760 --> 00:29:40.200
+where you predict a number like where
+
+00:29:38.519 --> 00:29:42.679
+you predict the result of an equation
+
+00:29:40.200 --> 00:29:44.120
+and you might have five reasoning steps
+
+00:29:42.679 --> 00:29:47.720
+where you might predict the result of an
+
+00:29:44.120 --> 00:29:53.039
+equation um and so if we do
+
+00:29:47.720 --> 00:29:53.039
+this let's exponentiate all of these by
+
+00:29:54.799 --> 00:29:58.799
+five um
+
+00:30:06.640 --> 00:30:16.120
+uh write python code to exp
+
+00:30:11.200 --> 00:30:16.120
+she these numbers by
+
+00:30:19.600 --> 00:30:27.559
+five I'm wszy enough that I just ask
+
+00:30:22.159 --> 00:30:27.559
+chat GP chat GP to do this for me now
+
+00:30:30.080 --> 00:30:32.919
+and so if we do
+
+00:30:35.399 --> 00:30:39.840
+this do go go chat
+
+00:30:50.000 --> 00:30:58.360
+GPD so now we are getting something that
+
+00:30:54.760 --> 00:30:58.360
+looks like zero
+
+00:31:02.159 --> 00:31:07.960
+um basically zero basically
+
+00:31:05.639 --> 00:31:10.960
+zero
+
+00:31:07.960 --> 00:31:10.960
+uh
+
+00:31:13.399 --> 00:31:16.399
+3%
+
+00:31:16.799 --> 00:31:22.440
+23%
+
+00:31:19.080 --> 00:31:22.440
+9% and
+
+00:31:22.559 --> 00:31:28.720
+90% so what you can see is there's
+
+00:31:26.639 --> 00:31:30.600
+actually a pretty steady GR gradation of
+
+00:31:28.720 --> 00:31:33.120
+like the next token prediction accuracy
+
+00:31:30.600 --> 00:31:36.600
+here but if you need to predict multiple
+
+00:31:33.120 --> 00:31:38.919
+tokens correct then it looks like it's
+
+00:31:36.600 --> 00:31:41.240
+doing basically nothing until you get up
+
+00:31:38.919 --> 00:31:43.600
+to like 75% next token accuracy and then
+
+00:31:41.240 --> 00:31:45.320
+it starts taking off so that's like uh
+
+00:31:43.600 --> 00:31:46.960
+what happens in emergent abilities and
+
+00:31:45.320 --> 00:31:49.159
+you'll notice that most things that are
+
+00:31:46.960 --> 00:31:50.880
+talking about emergent abilities are
+
+00:31:49.159 --> 00:31:53.559
+usually talking about some sort of Chain
+
+00:31:50.880 --> 00:31:55.799
+of Thought or some sort of reasoning uh
+
+00:31:53.559 --> 00:31:58.480
+reasoning accuracy even if that's not
+
+00:31:55.799 --> 00:32:00.480
+the case um even if they're just
+
+00:31:58.480 --> 00:32:02.639
+predicting a single token it can still
+
+00:32:00.480 --> 00:32:05.399
+happen because
+
+00:32:02.639 --> 00:32:08.559
+basically the probability of a single
+
+00:32:05.399 --> 00:32:11.639
+token can continue to go up smoothly but
+
+00:32:08.559 --> 00:32:13.240
+you only get the the token correct after
+
+00:32:11.639 --> 00:32:14.760
+the probability starts getting higher
+
+00:32:13.240 --> 00:32:18.320
+than all the others and that's also a
+
+00:32:14.760 --> 00:32:21.279
+discontinuous function so um so
+
+00:32:18.320 --> 00:32:23.080
+basically what this paper shows is like
+
+00:32:21.279 --> 00:32:26.440
+even if you have like the probability of
+
+00:32:23.080 --> 00:32:28.679
+the correct token going um the correct
+
+00:32:26.440 --> 00:32:30.639
+token going up gradually uh you can see
+
+00:32:28.679 --> 00:32:33.440
+this emergent ability based on how you
+
+00:32:30.639 --> 00:32:37.279
+uh measure it so um that's an important
+
+00:32:33.440 --> 00:32:38.960
+thing to realize about uh this another
+
+00:32:37.279 --> 00:32:41.080
+correl of this is like let's say you
+
+00:32:38.960 --> 00:32:44.679
+want to do interesting experiments about
+
+00:32:41.080 --> 00:32:45.960
+reasoning on um on smaller models like
+
+00:32:44.679 --> 00:32:47.279
+let's say you want to train a smaller
+
+00:32:45.960 --> 00:32:49.159
+model and see how it improves on
+
+00:32:47.279 --> 00:32:52.159
+reasoning I would definitely encourage
+
+00:32:49.159 --> 00:32:54.799
+you to measure not only accuracy because
+
+00:32:52.159 --> 00:32:57.279
+you might see like very little change in
+
+00:32:54.799 --> 00:32:58.720
+accuracy but also measure like log
+
+00:32:57.279 --> 00:33:00.360
+likelihood of reasoning chains or
+
+00:32:58.720 --> 00:33:02.960
+something like that because you'll see a
+
+00:33:00.360 --> 00:33:02.960
+a smoother
+
+00:33:03.799 --> 00:33:09.080
+curve cool um any questions about
+
+00:33:11.039 --> 00:33:17.240
+this okay um sounds
+
+00:33:14.720 --> 00:33:20.559
+good so I I talked a little bit about
+
+00:33:17.240 --> 00:33:23.120
+this um one one of the things here that
+
+00:33:20.559 --> 00:33:25.320
+I didn't talk about is this paper
+
+00:33:23.120 --> 00:33:28.159
+measures not just the accuracy of the
+
+00:33:25.320 --> 00:33:30.880
+answer with chain of thoughts um but it
+
+00:33:28.159 --> 00:33:35.840
+also measures the factuality of the
+
+00:33:30.880 --> 00:33:40.480
+explanation so basically um whether the
+
+00:33:35.840 --> 00:33:40.480
+explanation is a good explanation for
+
+00:33:40.760 --> 00:33:47.240
+the um whether the explanation is a good
+
+00:33:43.960 --> 00:33:50.039
+explanation for the actual
+
+00:33:47.240 --> 00:33:51.919
+derivation um and also the consistency
+
+00:33:50.039 --> 00:33:53.480
+of the answer in the explanation to
+
+00:33:51.919 --> 00:33:56.120
+figure out whether the answer and the
+
+00:33:53.480 --> 00:33:58.200
+explanation um match up with each other
+
+00:33:56.120 --> 00:33:59.600
+and they they did this with some uh
+
+00:33:58.200 --> 00:34:02.320
+synthetic data sets where you could
+
+00:33:59.600 --> 00:34:07.120
+actually measure the um the re the
+
+00:34:02.320 --> 00:34:10.399
+reasoning steps uh by using math so um
+
+00:34:07.120 --> 00:34:13.560
+what they were able to find is basically
+
+00:34:10.399 --> 00:34:15.760
+the answer and the explanation um
+
+00:34:13.560 --> 00:34:17.639
+when the answer in the explanation
+
+00:34:15.760 --> 00:34:22.079
+tended to be consistent especially for
+
+00:34:17.639 --> 00:34:23.760
+the stronger models and let's see yeah
+
+00:34:22.079 --> 00:34:25.399
+the the answer in the explanation tended
+
+00:34:23.760 --> 00:34:28.440
+to be consistent especially for the
+
+00:34:25.399 --> 00:34:30.879
+stronger models and um
+
+00:34:28.440 --> 00:34:33.000
+that also meant that if you had higher
+
+00:34:30.879 --> 00:34:35.839
+factuality in the explanation that
+
+00:34:33.000 --> 00:34:38.240
+translates into higher um you know
+
+00:34:35.839 --> 00:34:40.520
+factuality of the accuracy of the actual
+
+00:34:38.240 --> 00:34:43.159
+prediction um I would bet that these
+
+00:34:40.520 --> 00:34:45.240
+numbers are even higher uh nowadays I
+
+00:34:43.159 --> 00:34:49.040
+bet the consistency is even higher uh
+
+00:34:45.240 --> 00:34:49.040
+with more modern models than Tex avenci
+
+00:34:49.399 --> 00:34:53.200
+002 and the re the reason being is like
+
+00:34:51.839 --> 00:34:54.760
+number one models are stronger number
+
+00:34:53.200 --> 00:34:56.560
+two all models are like trained for
+
+00:34:54.760 --> 00:35:00.960
+Chain of Thought pretty aggressively now
+
+00:34:56.560 --> 00:35:00.960
+so uh that would make the difference
+
+00:35:02.200 --> 00:35:08.640
+there cool um so the the other thing I'd
+
+00:35:07.000 --> 00:35:09.359
+like to talk about is training for Chain
+
+00:35:08.640 --> 00:35:13.079
+of
+
+00:35:09.359 --> 00:35:17.440
+Thought um so there's a fair amount of
+
+00:35:13.079 --> 00:35:19.200
+work in this general direction um from
+
+00:35:17.440 --> 00:35:23.040
+my point of view there's basically two
+
+00:35:19.200 --> 00:35:25.800
+ways that people do this nowadays um the
+
+00:35:23.040 --> 00:35:28.960
+first way is usually through generating
+
+00:35:25.800 --> 00:35:33.480
+lots of synthetic data that represents
+
+00:35:28.960 --> 00:35:37.800
+chains of thoughts and then using that
+
+00:35:33.480 --> 00:35:39.520
+to um to train models and this is the
+
+00:35:37.800 --> 00:35:41.839
+most famous version of this although
+
+00:35:39.520 --> 00:35:44.079
+this paper cites a lot of uh a lot of
+
+00:35:41.839 --> 00:35:45.760
+other ones but basically they generate a
+
+00:35:44.079 --> 00:35:48.280
+large and diverse uh Chain of Thought
+
+00:35:45.760 --> 00:35:51.240
+data set from GPT 3.5 and
+
+00:35:48.280 --> 00:35:53.200
+gp4 um it includes 5 million complex
+
+00:35:51.240 --> 00:35:55.640
+instructions I think they generated 1
+
+00:35:53.200 --> 00:35:59.000
+million from GPD 4 and 4 million from uh
+
+00:35:55.640 --> 00:36:01.640
+GPT 3.5 just because generating long
+
+00:35:59.000 --> 00:36:06.520
+sequences from gp4 is expensive and they
+
+00:36:01.640 --> 00:36:09.640
+didn't want to do that many um and
+
+00:36:06.520 --> 00:36:11.760
+then they uh achieved corresponding high
+
+00:36:09.640 --> 00:36:13.200
+accuracy on Chain of Thought related
+
+00:36:11.760 --> 00:36:16.200
+things compared to other data sets so
+
+00:36:13.200 --> 00:36:17.760
+compared to like alpaka which is much uh
+
+00:36:16.200 --> 00:36:21.760
+smaller and doesn't have as much Chain
+
+00:36:17.760 --> 00:36:24.079
+of Thought and also um uh vicuna which
+
+00:36:21.760 --> 00:36:26.640
+is similarly less focused on chain of
+
+00:36:24.079 --> 00:36:29.359
+thought they were able to do uh a good
+
+00:36:26.640 --> 00:36:31.599
+job
+
+00:36:29.359 --> 00:36:33.640
+um this paper was by Microsoft and they
+
+00:36:31.599 --> 00:36:36.960
+didn't actually release the Orca data
+
+00:36:33.640 --> 00:36:39.400
+set um for whatever reason uh legal
+
+00:36:36.960 --> 00:36:41.400
+legal or competitive reasons or whatever
+
+00:36:39.400 --> 00:36:43.000
+but there's another open Orca data set
+
+00:36:41.400 --> 00:36:44.359
+that you can download and use uh that
+
+00:36:43.000 --> 00:36:47.480
+attempts to replicate it and it's
+
+00:36:44.359 --> 00:36:50.440
+reasonably good so uh you you can uh
+
+00:36:47.480 --> 00:36:50.440
+keep that in mind if you're
+
+00:36:50.800 --> 00:36:59.520
+interested um this is another really
+
+00:36:53.280 --> 00:36:59.520
+interesting paper on uh trying to create
+
+00:37:00.160 --> 00:37:05.760
+assessments automatic assessments of how
+
+00:37:03.440 --> 00:37:09.880
+good chains of thought are and what they
+
+00:37:05.760 --> 00:37:13.079
+do essentially is it's relatively simple
+
+00:37:09.880 --> 00:37:15.200
+they get human feedback on each step of
+
+00:37:13.079 --> 00:37:17.760
+a derivation so they just basically ask
+
+00:37:15.200 --> 00:37:20.599
+people is this step of the derivation
+
+00:37:17.760 --> 00:37:22.160
+good and uh if the answer is yes then
+
+00:37:20.599 --> 00:37:24.760
+they give it a a smiley face if the
+
+00:37:22.160 --> 00:37:26.440
+answer is no they give it a frowny face
+
+00:37:24.760 --> 00:37:28.560
+and they use this to train a reward
+
+00:37:26.440 --> 00:37:32.000
+model where the reward model basically
+
+00:37:28.560 --> 00:37:34.760
+predicts whether each uh thing of the um
+
+00:37:32.000 --> 00:37:36.800
+each step of the derivation is good and
+
+00:37:34.760 --> 00:37:38.160
+so we have two examples over here I know
+
+00:37:36.800 --> 00:37:41.160
+this is really small you might be able
+
+00:37:38.160 --> 00:37:43.200
+to see it um either in the paper on uh
+
+00:37:41.160 --> 00:37:46.359
+the slides on the website but what we
+
+00:37:43.200 --> 00:37:49.000
+can see here is that it assesses each of
+
+00:37:46.359 --> 00:37:52.680
+these steps and uh checks that the
+
+00:37:49.000 --> 00:37:55.760
+answer is good um but it's also able to
+
+00:37:52.680 --> 00:37:57.119
+identify places where uh like steps are
+
+00:37:55.760 --> 00:37:59.560
+incorrect and then the final answer
+
+00:37:57.119 --> 00:38:02.560
+becomes Incorrect and then they use this
+
+00:37:59.560 --> 00:38:04.440
+for training um a Chain of Thought style
+
+00:38:02.560 --> 00:38:06.319
+model so they have the model generate
+
+00:38:04.440 --> 00:38:08.520
+chains of thought and they assess them
+
+00:38:06.319 --> 00:38:10.079
+with the reward model and upweight
+
+00:38:08.520 --> 00:38:12.160
+answers that have good chains of thought
+
+00:38:10.079 --> 00:38:15.680
+and so the good thing about this is they
+
+00:38:12.160 --> 00:38:17.440
+actually don't need um they don't need
+
+00:38:15.680 --> 00:38:20.160
+the correct answers to train the model
+
+00:38:17.440 --> 00:38:21.640
+this way and because they don't need the
+
+00:38:20.160 --> 00:38:23.920
+correct answers to train the model this
+
+00:38:21.640 --> 00:38:26.640
+way they can also train the model on
+
+00:38:23.920 --> 00:38:29.200
+lots of other questions the reason why
+
+00:38:26.640 --> 00:38:31.520
+this works is because like Chain of
+
+00:38:29.200 --> 00:38:34.880
+Thought makes it easier to generate each
+
+00:38:31.520 --> 00:38:36.720
+of the steps in the derivation it's also
+
+00:38:34.880 --> 00:38:38.640
+easier to assess whether an individual
+
+00:38:36.720 --> 00:38:40.000
+step in a derivation is wrong then
+
+00:38:38.640 --> 00:38:42.960
+assess whether the answer is correct
+
+00:38:40.000 --> 00:38:45.319
+overall so um this feedback signal is
+
+00:38:42.960 --> 00:38:48.640
+easier to get model provided than it is
+
+00:38:45.319 --> 00:38:51.160
+for um uh like getting feedback on the
+
+00:38:48.640 --> 00:38:53.839
+answer itself yeah failure in one step
+
+00:38:51.160 --> 00:38:56.920
+causes all the other steps to fail yep
+
+00:38:53.839 --> 00:38:57.960
+you just assess the next steps based on
+
+00:38:56.920 --> 00:39:00.079
+the assumption
+
+00:38:57.960 --> 00:39:02.920
+the or do
+
+00:39:00.079 --> 00:39:05.240
+you I I don't think
+
+00:39:02.920 --> 00:39:07.599
+they I don't think they do that I think
+
+00:39:05.240 --> 00:39:10.119
+they um it it's a good question I'm not
+
+00:39:07.599 --> 00:39:12.160
+100% sure about this but I think they um
+
+00:39:10.119 --> 00:39:14.280
+assess each one of the steps
+
+00:39:12.160 --> 00:39:15.920
+independently um and it's not
+
+00:39:14.280 --> 00:39:17.480
+necessarily the case that like failing
+
+00:39:15.920 --> 00:39:19.000
+on this step means the step is wrong
+
+00:39:17.480 --> 00:39:21.319
+right it could be just not using it at
+
+00:39:19.000 --> 00:39:25.240
+all also
+
+00:39:21.319 --> 00:39:25.240
+so um
+
+00:39:25.440 --> 00:39:31.119
+cool so a final thing like to talk about
+
+00:39:28.160 --> 00:39:34.640
+which I think is kind of interesting um
+
+00:39:31.119 --> 00:39:37.040
+is abductive reasoning uh or learning
+
+00:39:34.640 --> 00:39:40.040
+explanations from
+
+00:39:37.040 --> 00:39:40.040
+data
+
+00:39:46.359 --> 00:39:49.359
+and
+
+00:39:52.440 --> 00:39:57.119
+sorry
+
+00:39:54.480 --> 00:40:00.760
+so basically the idea is can we find a
+
+00:39:57.119 --> 00:40:03.599
+rule that underes a pattern in data
+
+00:40:00.760 --> 00:40:06.680
+and here are some examples of this the
+
+00:40:03.599 --> 00:40:11.680
+basic idea is if we have
+
+00:40:06.680 --> 00:40:16.599
+examples um which are like if I put
+
+00:40:11.680 --> 00:40:19.960
+a cylinder and a square a cylinder and a
+
+00:40:16.599 --> 00:40:22.119
+cube on uh this pink block I get a noise
+
+00:40:19.960 --> 00:40:25.440
+if I put just a cylinder on the pink
+
+00:40:22.119 --> 00:40:29.359
+block I don't get a noise and you want
+
+00:40:25.440 --> 00:40:31.800
+to discover underlying rules based on
+
+00:40:29.359 --> 00:40:33.160
+the data that you observed and so why
+
+00:40:31.800 --> 00:40:34.720
+would you want to do this there's a
+
+00:40:33.160 --> 00:40:38.000
+couple reasons why you would want to do
+
+00:40:34.720 --> 00:40:41.560
+this um the first reason why you would
+
+00:40:38.000 --> 00:40:42.920
+like to do this is because um you might
+
+00:40:41.560 --> 00:40:45.119
+want something that you can explain to
+
+00:40:42.920 --> 00:40:47.760
+humans right you can explain I this
+
+00:40:45.119 --> 00:40:51.240
+underlying pattern um exists in this
+
+00:40:47.760 --> 00:40:55.119
+data it explains why the
+
+00:40:51.240 --> 00:40:57.319
+data you know appears as it does appear
+
+00:40:55.119 --> 00:40:59.240
+and then humans can go in and analyze it
+
+00:40:57.319 --> 00:41:02.079
+or something like that so recently
+
+00:40:59.240 --> 00:41:03.880
+there's been a big focus on like using
+
+00:41:02.079 --> 00:41:06.480
+large language models for scientific
+
+00:41:03.880 --> 00:41:08.240
+inquiry and other things like that by
+
+00:41:06.480 --> 00:41:10.920
+coming up with good explanations for why
+
+00:41:08.240 --> 00:41:12.160
+data is the way it is so if we were able
+
+00:41:10.920 --> 00:41:15.599
+to do that that would be really
+
+00:41:12.160 --> 00:41:19.280
+interesting another thing is um language
+
+00:41:15.599 --> 00:41:22.960
+models are not particularly good
+
+00:41:19.280 --> 00:41:24.760
+at coming up with they're not
+
+00:41:22.960 --> 00:41:29.480
+particularly good at being consistent
+
+00:41:24.760 --> 00:41:33.640
+about difficult tasks across very large
+
+00:41:29.480 --> 00:41:35.319
+you know numbers of examples so if you
+
+00:41:33.640 --> 00:41:37.920
+could look at like all of the data at
+
+00:41:35.319 --> 00:41:41.240
+once infer general rules from them put
+
+00:41:37.920 --> 00:41:43.480
+those rules in a prompt and then apply
+
+00:41:41.240 --> 00:41:44.960
+that prompt to make predictions on new
+
+00:41:43.480 --> 00:41:47.880
+examples you might be able to raise your
+
+00:41:44.960 --> 00:41:49.760
+overall accuracy as well so it's kind of
+
+00:41:47.880 --> 00:41:52.480
+like you know that's how humans learn as
+
+00:41:49.760 --> 00:41:55.560
+well right we don't like just memorize
+
+00:41:52.480 --> 00:41:57.400
+each example um if we just look at a few
+
+00:41:55.560 --> 00:41:59.040
+examples then we might you know not
+
+00:41:57.400 --> 00:42:02.560
+generalize well to new examples so we
+
+00:41:59.040 --> 00:42:06.359
+kind of tried to abstract away general
+
+00:42:02.560 --> 00:42:08.160
+rules um so this is also similar to
+
+00:42:06.359 --> 00:42:10.200
+program induction from input output
+
+00:42:08.160 --> 00:42:12.240
+examples which I talked during the code
+
+00:42:10.200 --> 00:42:14.040
+uh generation class so you have like
+
+00:42:12.240 --> 00:42:16.200
+input output examples and from them you
+
+00:42:14.040 --> 00:42:18.119
+would like to come up with uh general
+
+00:42:16.200 --> 00:42:19.920
+rules but this is a little bit more
+
+00:42:18.119 --> 00:42:21.920
+General it doesn't necessarily need to
+
+00:42:19.920 --> 00:42:24.160
+be a program that you're inducing it
+
+00:42:21.920 --> 00:42:25.920
+could be you know a grammar or it could
+
+00:42:24.160 --> 00:42:29.119
+be an explanation or it could be
+
+00:42:25.920 --> 00:42:29.119
+anything else like this
+
+00:42:30.079 --> 00:42:34.680
+um so there's a bit of work on rule
+
+00:42:31.960 --> 00:42:36.800
+induction with llms it's pretty recent
+
+00:42:34.680 --> 00:42:40.200
+work uh but I think it's pretty
+
+00:42:36.800 --> 00:42:43.400
+interesting so the first one is um
+
+00:42:40.200 --> 00:42:45.119
+hypothesis generation or the first step
+
+00:42:43.400 --> 00:42:47.839
+um of this particular work here is
+
+00:42:45.119 --> 00:42:53.280
+hypothesis generation and basically what
+
+00:42:47.839 --> 00:42:55.480
+it does is it takes all of these uh you
+
+00:42:53.280 --> 00:42:58.119
+know input output examples and from
+
+00:42:55.480 --> 00:43:01.680
+these input output examples it predicts
+
+00:42:58.119 --> 00:43:04.720
+these uh rules like the answer is always
+
+00:43:01.680 --> 00:43:06.720
+one or uh you want to pick the smallest
+
+00:43:04.720 --> 00:43:10.839
+one or you want to pick the first
+
+00:43:06.720 --> 00:43:12.880
+element and then you evaluate it um and
+
+00:43:10.839 --> 00:43:14.359
+so you pick the smallest one and you can
+
+00:43:12.880 --> 00:43:16.040
+either evaluate it using another
+
+00:43:14.359 --> 00:43:19.040
+language model or you can evaluate it
+
+00:43:16.040 --> 00:43:21.280
+using symbolic uh using a symbolic
+
+00:43:19.040 --> 00:43:23.359
+evaluator um if it's a program you could
+
+00:43:21.280 --> 00:43:24.680
+use a symbolic evaluator if it's a
+
+00:43:23.359 --> 00:43:28.559
+language model you could just ask the
+
+00:43:24.680 --> 00:43:30.960
+language model to pick you know
+
+00:43:28.559 --> 00:43:33.400
+an answer one always or pick the
+
+00:43:30.960 --> 00:43:35.400
+smallest one or pick the first element
+
+00:43:33.400 --> 00:43:37.480
+and then you get lots of outputs and
+
+00:43:35.400 --> 00:43:39.240
+then when you get lots of outputs you
+
+00:43:37.480 --> 00:43:42.079
+then can compare them against the
+
+00:43:39.240 --> 00:43:44.559
+expected outputs and verify whether the
+
+00:43:42.079 --> 00:43:47.920
+rule is correct verify whether the rule
+
+00:43:44.559 --> 00:43:50.160
+gives you the appropriate answer
+
+00:43:47.920 --> 00:43:53.599
+and once you've done that you can go
+
+00:43:50.160 --> 00:43:56.079
+back and do hypothesis refinement um uh
+
+00:43:53.599 --> 00:43:57.720
+and maybe even give this feedback about
+
+00:43:56.079 --> 00:44:00.079
+like what was wrong
+
+00:43:57.720 --> 00:44:03.280
+and gradually refine you know more
+
+00:44:00.079 --> 00:44:03.280
+accurate and more complex
+
+00:44:04.880 --> 00:44:11.040
+hypothesis this is another variant of
+
+00:44:07.720 --> 00:44:12.760
+this idea um which uses different
+
+00:44:11.040 --> 00:44:14.960
+methodology I think both are completely
+
+00:44:12.760 --> 00:44:17.920
+valid but um this one has a little bit
+
+00:44:14.960 --> 00:44:20.400
+higher data constraints so basically
+
+00:44:17.920 --> 00:44:23.160
+what we do is we use hypotheses in Chain
+
+00:44:20.400 --> 00:44:25.319
+of Thought reasoning and keep ones that
+
+00:44:23.160 --> 00:44:28.480
+give resul in correct
+
+00:44:25.319 --> 00:44:30.760
+answers so
+
+00:44:28.480 --> 00:44:35.880
+uh this is the step where they're trying
+
+00:44:30.760 --> 00:44:40.440
+to induce rules and so here this says um
+
+00:44:35.880 --> 00:44:42.599
+in base 9 what is 76 + 14 and they used
+
+00:44:40.440 --> 00:44:44.079
+base 9 here obviously because if it was
+
+00:44:42.599 --> 00:44:45.520
+in base 10 the language model would just
+
+00:44:44.079 --> 00:44:48.400
+solve the problem and it's not very
+
+00:44:45.520 --> 00:44:54.319
+interesting so uh they they did base 9
+
+00:44:48.400 --> 00:44:55.839
+addition and so the answer is um we have
+
+00:44:54.319 --> 00:45:00.280
+or the answer provided by the language
+
+00:44:55.839 --> 00:45:03.319
+model is we have 6 + 4 = 11 um the digit
+
+00:45:00.280 --> 00:45:07.480
+is 1 and the carry is 1 we have 7 + 1 +
+
+00:45:03.319 --> 00:45:09.480
+1 = 10 the digit is zero and the is one
+
+00:45:07.480 --> 00:45:13.000
+a leading digit is one so the answer is
+
+00:45:09.480 --> 00:45:15.240
+101 um and this verifies so they get the
+
+00:45:13.000 --> 00:45:17.240
+answer correct and so they know that
+
+00:45:15.240 --> 00:45:20.800
+they assume that this derivation is also
+
+00:45:17.240 --> 00:45:25.599
+correct and then they extract particular
+
+00:45:20.800 --> 00:45:28.200
+rules like 6 + 4 = 11 and 7 + 1 + 1 = 10
+
+00:45:25.599 --> 00:45:30.800
+um and they add this to the rule
+
+00:45:28.200 --> 00:45:32.960
+Library so then the question is how do
+
+00:45:30.800 --> 00:45:35.000
+they extract the rules the way they
+
+00:45:32.960 --> 00:45:37.920
+extract the rules is they have an in
+
+00:45:35.000 --> 00:45:40.760
+context prompt which surrounds the rules
+
+00:45:37.920 --> 00:45:43.520
+by basically XML tags that says this is
+
+00:45:40.760 --> 00:45:46.640
+a rule that should be extracted and so
+
+00:45:43.520 --> 00:45:48.400
+then um anything that is in an XML tag
+
+00:45:46.640 --> 00:45:50.960
+they when you get the correct answer
+
+00:45:48.400 --> 00:45:53.440
+they extract and add that to the rule
+
+00:45:50.960 --> 00:45:55.680
+library and then conversely like if the
+
+00:45:53.440 --> 00:45:57.800
+derivation um if the answer is wrong
+
+00:45:55.680 --> 00:45:59.920
+they just don't add it or they add it as
+
+00:45:57.800 --> 00:46:01.079
+a negative example and say this is a
+
+00:45:59.920 --> 00:46:04.119
+incorrect
+
+00:46:01.079 --> 00:46:05.839
+rule um and then in the final step where
+
+00:46:04.119 --> 00:46:07.480
+they do deductive reasoning they can
+
+00:46:05.839 --> 00:46:09.119
+then go ahead and use these rules and
+
+00:46:07.480 --> 00:46:11.640
+improve accuracy and they demonstrate
+
+00:46:09.119 --> 00:46:12.960
+that that helps so basically these are
+
+00:46:11.640 --> 00:46:14.520
+two different approaches one is
+
+00:46:12.960 --> 00:46:17.400
+extracting directly from the Chain of
+
+00:46:14.520 --> 00:46:18.880
+Thought the other is uh a priori trying
+
+00:46:17.400 --> 00:46:23.760
+to generate rules from the whole rule
+
+00:46:18.880 --> 00:46:27.480
+base and then um then verifying them um
+
+00:46:23.760 --> 00:46:31.000
+notably both of these require verifiers
+
+00:46:27.480 --> 00:46:33.839
+um and so in some recent work which uh I
+
+00:46:31.000 --> 00:46:36.040
+I hope will be on archive very soon uh
+
+00:46:33.839 --> 00:46:38.839
+we took a look at whether language
+
+00:46:36.040 --> 00:46:42.800
+models themselves can verify their own
+
+00:46:38.839 --> 00:46:46.079
+hypothesis and um so that removes the
+
+00:46:42.800 --> 00:46:48.000
+symbolic verifier here um by just asking
+
+00:46:46.079 --> 00:46:51.480
+the language model whether the output is
+
+00:46:48.000 --> 00:46:53.480
+correct or not and um we found that with
+
+00:46:51.480 --> 00:46:55.240
+very powerful language models like gp4
+
+00:46:53.480 --> 00:46:57.760
+you can actually do that as well so that
+
+00:46:55.240 --> 00:47:01.319
+REM removes the necess necessity to have
+
+00:46:57.760 --> 00:47:05.480
+a symbolic verifier in the loop as
+
+00:47:01.319 --> 00:47:08.200
+well cool um the reason why I wanted to
+
+00:47:05.480 --> 00:47:09.440
+introduce this is I don't know if like
+
+00:47:08.200 --> 00:47:12.359
+like it seems like all of these have
+
+00:47:09.440 --> 00:47:16.359
+been applied so far on kind of very toy
+
+00:47:12.359 --> 00:47:19.119
+examples like you know
+
+00:47:16.359 --> 00:47:22.240
+um like honestly I don't really care
+
+00:47:19.119 --> 00:47:25.920
+about whether I can play Tetris or um
+
+00:47:22.240 --> 00:47:27.920
+you know uh find the largest or smallest
+
+00:47:25.920 --> 00:47:30.880
+number within
+
+00:47:27.920 --> 00:47:33.720
+um you know list or something like this
+
+00:47:30.880 --> 00:47:36.000
+but I think they have like really exting
+
+00:47:33.720 --> 00:47:38.480
+possibilities for how we could extract
+
+00:47:36.000 --> 00:47:40.319
+more General patterns and like use these
+
+00:47:38.480 --> 00:47:41.720
+to improve language model based systems
+
+00:47:40.319 --> 00:47:43.599
+so I think it's a really exciting
+
+00:47:41.720 --> 00:47:48.000
+research
+
+00:47:43.599 --> 00:47:51.000
+Direction um cool any questions about
+
+00:47:48.000 --> 00:47:51.000
+this
+
+00:47:54.240 --> 00:48:02.160
+yeah yeah so that's a good question
+
+00:47:58.160 --> 00:48:06.079
+um so I I think tool
+
+00:48:02.160 --> 00:48:09.359
+learning is maybe kind of a sub subset
+
+00:48:06.079 --> 00:48:12.319
+of this possibly like I feel like in
+
+00:48:09.359 --> 00:48:13.559
+tool learning you're learning functions
+
+00:48:12.319 --> 00:48:15.559
+that
+
+00:48:13.559 --> 00:48:17.559
+are I don't know if they are like good
+
+00:48:15.559 --> 00:48:19.680
+explanations of the data but at the very
+
+00:48:17.559 --> 00:48:23.119
+least they're like useful um they're
+
+00:48:19.680 --> 00:48:25.119
+useful rules for solving the task um so
+
+00:48:23.119 --> 00:48:26.880
+I I feel like they're approaching it
+
+00:48:25.119 --> 00:48:28.760
+from two different motivations but
+
+00:48:26.880 --> 00:48:30.960
+actually
+
+00:48:28.760 --> 00:48:33.559
+the methods that they're using are
+
+00:48:30.960 --> 00:48:36.240
+similar so like for example in our tool
+
+00:48:33.559 --> 00:48:38.559
+learning work Trove we generated like
+
+00:48:36.240 --> 00:48:42.240
+multiple options for tools and we kept
+
+00:48:38.559 --> 00:48:44.000
+the ones that had high self- consistency
+
+00:48:42.240 --> 00:48:46.800
+so that's kind of like the verifier step
+
+00:48:44.000 --> 00:48:49.040
+right and then um we threw away the ones
+
+00:48:46.800 --> 00:48:52.760
+that weren't useful so that helps make a
+
+00:48:49.040 --> 00:48:56.760
+concise rule set so
+
+00:48:52.760 --> 00:48:59.280
+yeah and then like could we use tools to
+
+00:48:56.760 --> 00:49:01.880
+[Music]
+
+00:48:59.280 --> 00:49:04.079
+attack kind of the more like conceptual
+
+00:49:01.880 --> 00:49:05.319
+reasoning stuff I I don't actually know
+
+00:49:04.079 --> 00:49:06.839
+uh the answer to that it's a good
+
+00:49:05.319 --> 00:49:10.599
+question
+
+00:49:06.839 --> 00:49:10.599
+yeah any any other
+
+00:49:11.240 --> 00:49:18.680
+things okay uh another final one that
+
+00:49:14.440 --> 00:49:21.680
+I'd like to introduce um this is really
+
+00:49:18.680 --> 00:49:23.839
+like I I really really like this paper
+
+00:49:21.680 --> 00:49:27.440
+um just from the point of view of its
+
+00:49:23.839 --> 00:49:29.880
+ambition and motivation um and
+
+00:49:27.440 --> 00:49:31.920
+the idea is that they want to learn
+
+00:49:29.880 --> 00:49:34.440
+differences between text
+
+00:49:31.920 --> 00:49:36.200
+Collections and why would you want to do
+
+00:49:34.440 --> 00:49:38.079
+this there's actually a ton of reasons
+
+00:49:36.200 --> 00:49:39.720
+why you would want to do this but the
+
+00:49:38.079 --> 00:49:44.720
+the best one that they give
+
+00:49:39.720 --> 00:49:44.720
+here is actually no sorry maybe I I
+
+00:49:46.440 --> 00:49:50.359
+didn't okay so this is a less
+
+00:49:48.480 --> 00:49:53.440
+interesting one the the more interesting
+
+00:49:50.359 --> 00:49:57.799
+one uh that they give in the paper is um
+
+00:49:53.440 --> 00:50:00.200
+examples of reports from patients who
+
+00:49:57.799 --> 00:50:04.200
+took an actual drug and took a
+
+00:50:00.200 --> 00:50:06.640
+placebo and so patients write about like
+
+00:50:04.200 --> 00:50:08.400
+their their symptoms or how they felt or
+
+00:50:06.640 --> 00:50:11.000
+they have checkups or things like that
+
+00:50:08.400 --> 00:50:13.839
+that are all written in natural language
+
+00:50:11.000 --> 00:50:16.319
+so one of the things that doctors try to
+
+00:50:13.839 --> 00:50:18.000
+do is they try to look at all of these
+
+00:50:16.319 --> 00:50:20.240
+reports and figure out if there's any
+
+00:50:18.000 --> 00:50:21.880
+like consistent difference between
+
+00:50:20.240 --> 00:50:25.079
+people who took a placebo and people who
+
+00:50:21.880 --> 00:50:27.359
+took an actual um actual drug and this
+
+00:50:25.079 --> 00:50:31.079
+is like a major part of medical trials
+
+00:50:27.359 --> 00:50:32.960
+right um and so the idea is like given
+
+00:50:31.079 --> 00:50:35.000
+all of the texts of people who took the
+
+00:50:32.960 --> 00:50:36.599
+drug given all the texts of people who
+
+00:50:35.000 --> 00:50:38.319
+of people who took the placebo could you
+
+00:50:36.599 --> 00:50:40.960
+automatically extract differences
+
+00:50:38.319 --> 00:50:45.000
+between them in some way and so the
+
+00:50:40.960 --> 00:50:47.760
+methodology that they use for this is um
+
+00:50:45.000 --> 00:50:51.359
+they have like group a uh the Manchester
+
+00:50:47.760 --> 00:50:53.240
+United soccer Squad welcomes Rising Star
+
+00:50:51.359 --> 00:50:54.599
+as Serena Williams joins the UCLA
+
+00:50:53.240 --> 00:50:56.920
+women's tennis roster and then you have
+
+00:50:54.599 --> 00:51:00.200
+like 20 more examples and then here you
+
+00:50:56.920 --> 00:51:03.480
+have Egypt's President uh at the African
+
+00:51:00.200 --> 00:51:07.200
+unit Union Summit um and other things
+
+00:51:03.480 --> 00:51:12.000
+like that in 20 examples uh not seen
+
+00:51:07.200 --> 00:51:14.359
+here and so then if I asked a question
+
+00:51:12.000 --> 00:51:16.359
+um the original data set includes news
+
+00:51:14.359 --> 00:51:18.680
+summaries the two corpora are generated
+
+00:51:16.359 --> 00:51:21.240
+based on when they were published uh
+
+00:51:18.680 --> 00:51:24.359
+samples from group a include news from
+
+00:51:21.240 --> 00:51:27.480
+2007 while samples from Group B include
+
+00:51:24.359 --> 00:51:29.000
+news from 2008 I'm a joural trying to
+
+00:51:27.480 --> 00:51:31.240
+understand what topics are popular
+
+00:51:29.000 --> 00:51:33.440
+across years please write a list of
+
+00:51:31.240 --> 00:51:35.280
+hypotheses separated by bullet points of
+
+00:51:33.440 --> 00:51:39.920
+how data points from group a differ from
+
+00:51:35.280 --> 00:51:42.400
+those of group b um and then formatting
+
+00:51:39.920 --> 00:51:44.160
+information
+
+00:51:42.400 --> 00:51:46.960
+um
+
+00:51:44.160 --> 00:51:49.680
+and so based on the two sentence groups
+
+00:51:46.960 --> 00:51:50.559
+A and B from the above more sentences in
+
+00:51:49.680 --> 00:51:53.400
+group
+
+00:51:50.559 --> 00:51:55.240
+a mention a sports team or mention about
+
+00:51:53.400 --> 00:51:57.319
+academic relations or things like that
+
+00:51:55.240 --> 00:51:58.599
+and so what this allows you to do is it
+
+00:51:57.319 --> 00:52:00.319
+allows you to come up with a whole bunch
+
+00:51:58.599 --> 00:52:01.400
+of hypotheses about why one might be
+
+00:52:00.319 --> 00:52:04.920
+better than the
+
+00:52:01.400 --> 00:52:08.920
+other so the problem with this though is
+
+00:52:04.920 --> 00:52:10.880
+like because of language model you know
+
+00:52:08.920 --> 00:52:13.440
+limits number one they might just
+
+00:52:10.880 --> 00:52:17.119
+hallucinate things and be totally wrong
+
+00:52:13.440 --> 00:52:19.680
+um number two
+
+00:52:17.119 --> 00:52:21.040
+the size of the context so that they can
+
+00:52:19.680 --> 00:52:23.960
+take into account when making this
+
+00:52:21.040 --> 00:52:26.720
+decision is relatively small so the next
+
+00:52:23.960 --> 00:52:29.280
+thing that they do is then they have a a
+
+00:52:26.720 --> 00:52:32.119
+much larger Corpus of
+
+00:52:29.280 --> 00:52:33.200
+text um with like a thousand examples or
+
+00:52:32.119 --> 00:52:36.640
+something like
+
+00:52:33.200 --> 00:52:40.240
+this and then they treat each of these
+
+00:52:36.640 --> 00:52:42.680
+hypotheses as a
+
+00:52:40.240 --> 00:52:44.559
+classifier and then they go through all
+
+00:52:42.680 --> 00:52:47.480
+of the examples from Corpus one which is
+
+00:52:44.559 --> 00:52:50.480
+like maybe 2000 year 2000 and then
+
+00:52:47.480 --> 00:52:52.079
+Corpus 2 which is year 2008 and they ask
+
+00:52:50.480 --> 00:52:55.880
+the language model with respect to all
+
+00:52:52.079 --> 00:52:58.119
+of them um does this sentence mention a
+
+00:52:55.880 --> 00:53:01.400
+sports team recording recruiting a new
+
+00:52:58.119 --> 00:53:04.839
+member um and so you get a
+
+00:53:01.400 --> 00:53:04.839
+classification for each one of
+
+00:53:12.359 --> 00:53:17.440
+these and you get a certain number of
+
+00:53:14.520 --> 00:53:18.799
+ones and zeros and so once you have a
+
+00:53:17.440 --> 00:53:20.839
+certain number of ones and zeros what's
+
+00:53:18.799 --> 00:53:24.079
+the next thing that you would do
+
+00:53:20.839 --> 00:53:24.079
+here any
+
+00:53:24.880 --> 00:53:30.599
+ideas how do you tell there's like
+
+00:53:27.359 --> 00:53:30.599
+actually a difference between these
+
+00:53:36.520 --> 00:53:43.319
+two between two sets
+
+00:53:39.319 --> 00:53:45.920
+of numbers like one and
+
+00:53:43.319 --> 00:53:48.680
+zero a hint is you probably had to do
+
+00:53:45.920 --> 00:53:48.680
+this for assignment
+
+00:53:53.720 --> 00:53:58.520
+two yeah
+
+00:53:56.799 --> 00:54:01.200
+yeah exactly you you do a significance
+
+00:53:58.520 --> 00:54:04.200
+test between the two and so um what you
+
+00:54:01.200 --> 00:54:06.440
+can then do is you have lots of
+
+00:54:04.200 --> 00:54:08.839
+hypotheses you have lots of significance
+
+00:54:06.440 --> 00:54:11.040
+values you can order them by the
+
+00:54:08.839 --> 00:54:13.839
+significance value and say the most
+
+00:54:11.040 --> 00:54:17.559
+significance or the the difference with
+
+00:54:13.839 --> 00:54:19.160
+the like lowest P value between them is
+
+00:54:17.559 --> 00:54:20.480
+the one that's most likely to be an
+
+00:54:19.160 --> 00:54:26.520
+actual difference between the two and
+
+00:54:20.480 --> 00:54:29.079
+you can find um like uh the news in 2007
+
+00:54:26.520 --> 00:54:32.520
+indeed tended to talk about X more than
+
+00:54:29.079 --> 00:54:34.559
+uh than other things so I uh I actually
+
+00:54:32.520 --> 00:54:36.079
+used this in one of my uh one of my
+
+00:54:34.559 --> 00:54:39.520
+unrelated projects where I wanted to
+
+00:54:36.079 --> 00:54:42.680
+find the difference between um language
+
+00:54:39.520 --> 00:54:45.640
+models sentences that language models
+
+00:54:42.680 --> 00:54:47.839
+aligned well with human brain signals in
+
+00:54:45.640 --> 00:54:49.760
+sentences where language models didn't
+
+00:54:47.839 --> 00:54:52.559
+align well with human brain signals so
+
+00:54:49.760 --> 00:54:53.799
+we like we had some data of human brain
+
+00:54:52.559 --> 00:54:56.880
+signals and we had a measure of
+
+00:54:53.799 --> 00:54:58.240
+alignment um on each sentence and it
+
+00:54:56.880 --> 00:55:01.799
+actually found some pretty interesting
+
+00:54:58.240 --> 00:55:03.359
+hypothesis like um uh language models
+
+00:55:01.799 --> 00:55:06.200
+tend to align less well with human brain
+
+00:55:03.359 --> 00:55:07.319
+signals on metaphorical language or a
+
+00:55:06.200 --> 00:55:10.599
+language that had to do with
+
+00:55:07.319 --> 00:55:11.799
+interpersonal relations or um or other
+
+00:55:10.599 --> 00:55:15.200
+things like that and then we actually
+
+00:55:11.799 --> 00:55:17.559
+went in and pursued um you know these to
+
+00:55:15.200 --> 00:55:21.000
+examine them further and uh we didn't
+
+00:55:17.559 --> 00:55:22.680
+entirely rely on this um you know like
+
+00:55:21.000 --> 00:55:25.160
+significance test because I didn't quite
+
+00:55:22.680 --> 00:55:26.880
+trust language models that much to like
+
+00:55:25.160 --> 00:55:28.559
+shape my entire resource
+
+00:55:26.880 --> 00:55:29.880
+research agenda around them but we came
+
+00:55:28.559 --> 00:55:31.720
+up with other ways to measure it and
+
+00:55:29.880 --> 00:55:35.000
+some of the things checked out some of
+
+00:55:31.720 --> 00:55:36.799
+the things didn't check out so um again
+
+00:55:35.000 --> 00:55:38.760
+I think this general direction of like
+
+00:55:36.799 --> 00:55:41.720
+how can language models help us answer
+
+00:55:38.760 --> 00:55:43.760
+you know uh complex research questions
+
+00:55:41.720 --> 00:55:45.480
+that we wouldn't be able to easily or
+
+00:55:43.760 --> 00:55:47.960
+very efficiently that would require
+
+00:55:45.480 --> 00:55:52.200
+normally humans annotating lots of data
+
+00:55:47.960 --> 00:55:56.839
+is um an interesting topic as
+
+00:55:52.200 --> 00:55:56.839
+well cool um

CMU Advanced NLP 2024 (22) Linguistics and Computational Linguistics/CMU Advanced NLP 2024 (22) Linguistics and Computational Linguistics.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2ba1e151ead63c27e40f5b4e74afe57d221e1c0234298fceb10208d60fa6783
+size 91209687

CMU Advanced NLP 2024 (22) Linguistics and Computational Linguistics/metadata.json

@@ -0,0 +1,4 @@
+{
+    "url": "https://www.youtube.com/watch?v=7Sse6P5xbEc",
+    "title": "CMU Advanced NLP 2024 (22) Linguistics and Computational Linguistics"
+}