text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
Voluntary Transparency for Content Takedowns
Virtually since their emergence in the early 2000s, platform companies have had to
weigh legal requests for content takedowns from individuals and governments
around the world (Goldsmith and Wu 2006). As Daphne Keller and Paddy
Leerssen explain in Chapter 10 in this volume, ... | Social_Media_and_Democracy |
The Ninth Circuit – adopting a rationale parallel to that of the Seventh
Circuit – held that Roommates.com did not receive immunity from CDA 230
since it played the role of a “information content provider” (Quist 2012). By
designing a website registration process that included questions around
categories like gender an... | Social_Media_and_Democracy |
The a16z Investment Thesis on AI in Bio + Health | Andreessen Horowitz
https://a16z.com/2023/06/21/ai-bio-health-thesis/
4/9 | The a16z Investment Thesis on AI in Bio + Health _ Andreessen Horowitz |
Co., New York, 1976. ISBN 0-7167-0463-3.
[104] Zachary Kenton, Tom Everitt, Laura Weidinger, Iason Gabriel, Vladimir Mikulik, and Geoffrey Irving. Alignment
of language agents. arXiv preprint arXiv:2103.14659, 2021.
[105] Clifford Nass and Youngme Moon. Machines and mindlessness: Social responses to computers. Journ... | LaMDA- Language Models for Dialog Applications |
6. Conclusion
We take inspiration from a mechanism described in neu-
rophysiological research with the introduction of a priority
map module that combines temporal sequence alignment
enabled by high-level trajectory estimation and feature-
level localisation. Two new resources comprised of in-
domain samples and a tai... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
We are grateful for the many individuals and institutions that made this
volume possible. The John S. and James L. Knight Foundation provided critical
funding for this volume, as well as support for the labs of the two editors and
many of the chapter authors. Sam Gill from Knight also provided helpful
comments on sever... | Social_Media_and_Democracy |
the learnable weight matrix and bias vector, respectively, of
the linear layer.
3.3. Large Language Model Architecture
We adopt Llama-7B model as the LLM component of
Inspired by mPLUG-Owl [50], GPT4Video
GPT4Video.
employs a two-stage training strategy.
In the first phase,
we freeze the parameters of LLM and focus on... | GPT4Video |
3.2 Exploration Phase
Exploring by autonomous interactions. The Ex-
ploration Phase is central to our framework. Here,
the agent learns about the functionalities and fea-
tures of smartphone apps through trial and error. In
this phase, the agent is assigned a task and starts
interacting autonomously with the UI element... | AppAgents |
Starting Matlab. Matlab. The MathWorks, Natick, MA, 2012.
Pierre-Emmanuel Mazaré, Samuel Humeau, Martin Raison, and Antoine Bordes. Training millions of
personalized dialogue agents. In Proceedings of the 2018 Conference on Empirical Methods in Natural
Language Processing, pp. 2775–2779, Brussels, Belgium, 2018. Assoc... | Tool Learning with Foundation Models |
Let me experience thefestival inthis world...UserMulti-AgentOrdering dishes and cooking Taskplanning and solvingBand performingDiscussing decorationKitchenConcertCooperationOutdoorsActingwithtoolsAn Envisioned Agent Society collaboration, negotiation, or competition. Regardless of the mode of interaction, agents collec... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
stored in pretrained LLMs into the planning process.
With few exceptions, the parameters of the LLMs employed
in many of these works are employed as-is without further
training. In LID (Li et al., 2022), this constraint is relaxed
and LLM parameters are finetuned to produce a planning net-
work for generating high-level... | PaLM-E- An Embodied Multimodal Language Model |
ratio, providing optimal per-token loss efficiency [2]. We
accomplish this by grouping each task by sequence length,
and alternately sample one group at each iteration. Because
sequence lengths are fixed per task, we can optimally train
without any padding. Due to images requiring fewer to-
kens, we can include roughly 5... | VideoPoet |
Luyu Gao, Zhuyun Dai, Panupong Pasupat, Anthony Chen, Arun Tejasvi Chaganty, Yicheng Fan,
Vincent Zhao, Ni Lao, Hongrae Lee, Da-Cheng Juan, et al. Rarr: Researching and revising what
language models say, using language models. In Proceedings of the 61st Annual Meeting of the
Association for Computational Linguistics (V... | LARGELANGUAGEMODELSCANNOTSELF-CORRECT REASONINGYET |
Longer training with more tokens: PaLMChilla 62B was trained longer than PaLM
62B, with almost double the number of tokens but with only fractional increase in training
FLOP count; it performed slightly better on some zero-shot English NLP tasks like reasoning
[4]. Our studies comparing Flan-PaLM 62B and Flan-PaLMChill... | PersonalityTraitsinLargeLanguageModels |
64
Alexandra A. Siegel
Studying the network structure of users who produce online hate speech,
Magdy et al. (2016) find that they can predict the likelihood that Twitter users
tweet anti-Muslim messages after the 2015 Paris attacks with high levels of
precision and accuracy based on their Twitter networks, even if the... | Social_Media_and_Democracy |
156162
VOLUME 9, 2021
M. F. Mridha et al.: Comprehensive Review on Fake News Detection With Deep Learning
FIGURE 10. The BERT architecture taken from Devlin et al. [89].
the proposed model named exBAKE (BERT with extra
unlabeled news corpora) outperformed by a 0.137 F1-score.
Ding et al. [154] discovered that incl... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
Encoder-decoder. Cao et al. [19] extract fact descriptions from the source text and apply a dual-
attention seq-to-seq framework to force the summaries to be conditioned on both source documents
and the extracted fact descriptions. Li et al. [103] propose an entailment-aware encoder and
decoder with multi-task learning... | SurveyofHallucinationinNatural Language Generation |
information, 41
web crawlers, limitations in gathering
advertising data, 130, see also bots
Webster, James, 139
Weichart, Stephan, 204
Westwood, Sean J., 39
WhatsApp, 25–26, 328
WhoTargetsMe, 300
wikiedits bots, 95
Williams, Christine B., 128–129
Williams, Ev, 279
Winter, Fabian, 75
“wisdom of the crowds,” failure o... | Social_Media_and_Democracy |
image retrieval. In International Conference on Learning Representations, 2021.
P. Welinder, S. Branson, T. Mita, C. Wah, F. Schroff, S. Belongie, and P. Perona. Caltech-UCSD Birds 200.
Technical Report CNS-TR-2010-001, California Institute of Technology, 2010.
Guillaume Wenzek, Marie-Anne Lachaux, Alexis Conneau, Vi... | DINOv2- Learning Robust Visual Features without Supervision |
Aside from the work mentioned above, there are other studies that are based on the Tacotron
architecture. For example, Skerry-Ryan et al. [503] and Wang et al. [584] proposed Tacotron-based
models for prosody control. These models use a separate encoder to compute style information from
reference audio that is not prov... | AReviewofDeepLearningTechniquesforSpeechProcessing |
t≥1
DKL(q(xT|x0) (cid:107) p(xT )) +
DKL(q(xt−1|xt, x0) (cid:107) pθ(xt−1|xt)) − log pθ(x0|x1)
(22)
The following is an alternate version of L. It is not tractable to estimate, but it is useful for our
discussion in Section 4.3.
− log p(xT ) −
− log p(xT ) −
− log
p(xT )
q(xT ) −
t≥1
(cid:88)
(cid:88)
(cid... | Denoising Diffusion Probabilistic Models |
Hugo Touvron, Louis Martin, Kevin Stone, Peter Al-
bert, Amjad Almahairi, Yasmine Babaei, Nikolay
Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti
Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton
Ferrer, Moya Chen, Guillem Cucurull, David Esiobu,
Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller,
Cynthia Gao, Veda... | AComprehensiveSurveyofHallucinationMitigationTechniquesinLarge LanguageModels |
Misinformation is often defined in a way that allows for
its automatic detection. Dhar et al. (2016) describe misinfor-
mation as a rumor; pushing that definition further, Tsugawa
and Ohsaki (2017) identify misinformation with the concept
of “flaming” where falsehoods become viral when expressed
in negative terms; ... | Use of bot and content flags to limit the spread of misinformation among social networks: a behavior and attitude survey |
1.7. Additional Extrapolation Results
We show extrapolations as animations in our supplementary video. For each expression, we interpolate the individual
FLAME expression parameter from [-4, 4], and keep all other pose and expression parameters fixed as zero. We show the
smiling (1st), lip side movement (3rd), and eyeb... | I M Avatar- Implicit Morphable Head Avatars from Videos |
Practical Speedups. Finally, we study practical applications. As an interesting use-case, we focus
on the OPT-175B model: quantized to 3 bits, this model takes approximately 63GB of memory,
including the embeddings and the output layer, which are kept in full FP16 precision. Additionally,
storing the complete history o... | GPTQ |
degeneration. arXiv preprint arXiv:1904.09751, 2019.
Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Cosmos QA: machine reading
comprehension with contextual commonsense reasoning. In Proceedings of EMNLP, 2019.
Simon Hughes. Cut the bull. . . . detecting hallucinations in large language models. vect... | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
6
Figure 2: Left. The frontier of expected reward vs KL to the reference policy. DPO provides the highest expected
reward for all KL values, demonstrating the quality of the optimization. Right. TL;DR summarization win
rates vs. human-written summaries, using GPT-4 as evaluator. DPO exceeds PPO’s best-case performanc... | Direct Preference Optimization |
Assessment of expectations. We measured user expectations of performance and how they per-
sisted after the interaction. For overall performance expectations (judgments prior to interaction),
we used four questions: A seven-point Likert item (1: Strongly disagree, and, 7 Strongly agree), "I
think I will perform better ... | AI enhance sour performance |
“Portrait of Edmond Belamy” case to explore how anthropomorphization of an AI system influences the perception of
humans involved in the creation process. Stephensen [117] discusses the implication that the Belamy case has on the
philosophical understanding of creativity. Colton et al. [28] discuss how human understandi... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
1. It will become possible and financially feasible to build AI systems with the following
properties:
• Advanced capability: they outperform the best humans on some set of tasks which
when performed at advanced levels grant significant power in today’s world (tasks
like scientific research, business/military/political ... | Is Power-Seeking AI an Existential Risk? |
post with abike leaning against it.""Turn yourself so that you are going with the flow of traffic. There shouldbe a purple theater banner on your left. Go forward on this street until youcome to the first traffic light. Make a right at the light. You should see silvergates on your left. Go straight and when you come to... | APriorityMapforVision-and-LanguageNavigation withTrajectoryPlansandFeature-LocationCues |
In addition, we computed the average senti-
ment (Baccianella et al., 2010) of words co-
occurring with the gendered pronouns across each
dataset in Figure 13. Generally, we find no sig-
nificant sentiment bias towards men or women.
This, of course, does not mean that the dataset
is free of gender bias (as our co-occurre... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
5.1 SETTING THE NUMBER OF EXPERTS
One of the first questions is the number of experts to use. Fedus et al. (2021) presented the scaling-
properties of Switch Transformer which yielded monotonic pre-training benefits (on a step basis) on
13 | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
y = F(x; Θ).
(1)
trainable copyzero convolutionzero convolution+cControlNet(a) Before(b) Aerneural network blockxyxyc+neural network block (locked) In our setting, x and y are usually 2D feature maps, i.e., x ∈
Rh×w×c with {h, w, c} as the height, width, and number of
channels in the map, respectively (Figure 2a). | AddingConditionalControltoText-to-ImageDiffusionModels |
S2
l,j
∣=
∣wenc
(cid:96)1(wenc
l,⋅)= J∑
j=1
l,+− wenc
= wenc
l,−=
=(1− wenc
l,−)− wenc
l,−=
= 1− 2⋅ wenc
l,−=
= 1+ 2⋅∣wenc
l,−∣ .
(S7)
(S8)
(S9)
(S10)
This means that the (cid:96)1 penalty is equivalent to penalizing
(S6)
the absolute sum of the negative weights.
When all weights are non-negative, we get convex ... | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
Table 1. List of representative special tokens used in training
and inference.
When a modality is not included in a task, such as text
and audio for unconditioned video generation, then the cor-
responding input or output tokens together with the begin-
ning and end special tokens are omitted from the sequence
to redu... | VideoPoet |
4 Experiments
In this section, we will present our evaluation of
the multimodal agent framework through a combi-
nation of quantitative and qualitative experiments.
Our primary goal is to assess the agent’s perfor-
mance and its ability to operate a diverse set of
smartphone applications effectively.
4.1 Experimental... | AppAgents |
A.2 Patterns over Low-Resolution Images
In Fig. 10, we show an example in-context grasp detector which outputs target coordinates in a downsampled
image, given 6 in-context examples, as well as an example of a simple forward dynamics model predicting
spatial rearrangement of a red bowl into a green plate, given 9 in-c... | LargeLanguageModelsasGeneralPatternMachines |
6 Advanced Transfer Learning Techniques for Speech Processing
6.1 Domain Adaptation
6.1.1 Task Description
Domain adaptation is a field that deals with adapting a model trained on a labeled dataset from a
source domain to a target domain, where the source domain differs from the target domain. The
goal of domain adapta... | AReviewofDeepLearningTechniquesforSpeechProcessing |
In order to explore the applicability of convolutional neural
networks in understanding images beyond object detection
and classification, we aim to address image properties related
to the subjective and affective aspects of human perception.
We focus on three different levels of perceiving images:
the aesthetic evaluat... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
Insert the cork into one of the smaller holes
The cork should fit snugly but be able to move
As the steam builds up in the can, it
Because
Leave some room at the top of the can for the steam
Try experimenting with different
For
Table 12: Improvement over seed model in information seeking.
21
Prompt: What are so... | Self-AlignmentwithInstructionBacktranslation |
apetype.merge(A,B):youcanmergetwoshapesintoone.render(A):youcangetthemodelingdataofshapeA,andsaveitintothefile’data.json’HereisanexampleofhowtouseShapeEditor.DemonstrationExample:B=shape_2d.triangle(-1,-32,-23,-32,-20,0)A1=shape_3d.cylinder(shape_2d.triangle(-1,-32,-23,-32,-20,0),2,13,-16,[1/4*pi,1/2*pi,0])A1=transform(... | Tool Learning with Foundation Models |
(7)
while P ∈ RB×L is the original position in integer. B is
the batch size and N is the input text sequence length. Gs is
a hyperparameter of group size. It is the base of the FLOOR
operation. We denote the self-attention with this FLOOR
operations applied as ’grouped attention’.
③ Can LLMs work well without accurate ... | Self-Extend LLM |
to encompass the visual modality. Visual ChatGPT [20] fuses visual foundation models, such
as BLIP [21] and ControlNet [22], with LLMs. Visual Programming [23] and ViperGPT [20]
apply LLMs to visual objects by employing programming languages, parsing visual queries into
interpretable steps expressed as Python code. Mor... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
3
2
0
2
v
o
N
1
]
L
C
.
s
c
[
1
v
0
3
4
0
0
.
1
1
3
2
:
v
i
X
r
a
DISTIL-WHISPER: ROBUST KNOWLEDGE
DISTILLATION VIA LARGE-SCALE
PSEUDO LABELLING
Sanchit Gandhi, Patrick von Platen & Alexander M. Rush
Hugging Face
{sanchit, patrick, sasha}@huggingface.co
ABSTRACT | DISTIL-WHISPER |
Most current systems aren't even in the right ballpark. At a minimum, adequate
knowledge frameworks will require that we can represent and manipulate some
fraction of our knowledge in algebraic ways, by means of operations over variables; it is
likely that some (large) subset of that knowledge is encoded and maintai... | The Next Decade in AI- |
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike
Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining
approach, 2019.
Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization.
arXiv preprint
arXiv:1711.05101, 2017.
Ilya ... | LORA |
Familiarity backfire effects involve cases in which retractions increase, rather
than reduce, reliance on misinformation by making misinformation feel more
familiar. These effects are primarily studied in the context of repetition. In
particular, familiarity backfire effects are considered the product of the
illusory tru... | Social_Media_and_Democracy |
As the construction of these symbolic reasoning tasks is
well-defined, for each task we consider an in-domain test
set for which examples had the same number of steps as
the training/few-shot exemplars, as well as an out-of-domain (OOD) test set, for which evaluation
examples had more steps than those in the exemplars. ... | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
described more in detail in subsection 7.1.
7 DISCUSSION
In this section, we provide an overview of our approach, the necessary details for administering the SHAPE scale
as well as information on how to use it. In addition, we discuss the limitations of our approach and opportunities
for further developments.
In this... | Society’sAttitudesTowardsHumanAugmentation |
and is in 4/4 time signature. The melody is played on the right hand with the left hand playing chords. The piece is a classic ragtime style.A solo piano piece.Give me some recommendations for similar music.some similar music to “Solo Piano Ragtime” could be:1. “The Entertainer” by Scott Joplin2. “Maple Leaf Rag” by Sc... | Qwen-Audio |
2 BACKGROUND
Sparse expert models typically substitute a neural network layer with a set of experts, each having
unique weights (Jacobs et al., 1991; Jordan and Jacobs, 1994). Typically all the experts within a layer
are of the same type and shape (homogeneous), however, varied (heterogeneous) expert-types are
possibl... | ST-MOE- DESIGNING STABLE AND TRANSFERABLE SPARSE EXPERT MODELS |
Self-Declared Expertise. For level of expertise in using LLMs, with the five categories defined in [7], 243 participants
selected “Novice”, 181 “Advanced Beginner”, 52 “Competent”, 145 “Proficient”, and only six “Expert”. To gain an
understanding of expertise as a function of demographics, we looked at the effect of ge... | Adoptionand AppropriationofLLMs |
We recognize the potential risks of a model capable of generating speech in the style of arbitrary
people. In an effort to diminish these risks we show that a binary classification model is able to
consistently distinguish between real world speech and that which is generated from our model.
Inspired by [Kharitonov et ... | Voicebox-Text-GuidedMultilingual UniversalSpeechGenerationatScale |
(a) “a teddy bear sitting on a stone and wearing a scarf and wearing a baseball cap”
Figure 9: Visual results on the Animals set, which are inferred by our Instant3D for novel prompts. The results demonstrate
accurate text-3D alignment and satisfying multi-view consistency.
(b) “a panda sitting in a basket and wearin... | Instant3D |
Is there reason to believe the annotation judgments in this dataset may lose
Dataset Release and Maintenance
validity over time? If so, are there plans to update the dataset? Perceptions of toxic language will likely change over
time along with changing language or terminology and broader social views of acceptable lan... | PaLM 2 Technical Report |
[525] Jaesung Tae, Hyeongju Kim, and Taesu Kim. 2021. EdiTTS: Score-based Editing for Controllable Text-to-Speech.
CoRR abs/2110.02584 (2021). arXiv:2110.02584 https://arxiv.org/abs/2110.02584
[526] Ke Tan and DeLiang Wang. 2019. Learning complex spectral mapping with gated convolutional recurrent networks
for monaur... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Training setup. We train Transformer (Vaswani et al., 2017) decoder-only LMs with the standard
next-token language modeling loss. We conduct a controlled comparison by equalizing the amount
of compute, measured by the number of tokens processed during training. For The Pile, we train
each model for 200k steps; for the ... | DoReMi- Optimizing Data Mixtures Speeds Up Language Model Pretraining |
desirable for their ease-of-use, task-effectiveness,
parameter efficiency, and their ability to generate
fluent and plausible rationales. We expect models
of this kind to play an important role in continuing
research on explainable AI for these reasons.
We use the I→OR variant of T5 (Narang et al.,
2020). Because only... | Measuring Association Between Labels and Free-Text Rationales |
videos up to 25 minutes in length with video diffusion models, however the domain is restricted.
In this work, we introduce Imagen Video, a text-to-video generation system based on video diffusion
models (Ho et al., 2022b) that is capable of generating high definition videos with high frame fidelity,
strong temporal cons... | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
In evaluating the constitutionality of that law, a plurality of the Court noted
that “falsity alone may not suffice to bring the speech outside the First
Amendment” and rejected the argument that a government “interest in
truthful discourse alone [was] sufficient to sustain a ban on speech.”61 The
Court argued for battli... | Social_Media_and_Democracy |
Look at the sky, do you think it will rain tomorrow? Ifso, give the umbrella to me.EnvironmentPerceptionToolsCallingAPI …EmbodimentTextReasoningfromthe current weather conditionsand the weather reports on the internet, it is likely to rain tomorrow.Here is your umbrella.BrainKnowledgeMemoryStorageDecisionMakingPlanning... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
that renders density estimation relatively straightforward.
Of course, this does not escape the curse of dimensionality
so much as relocate it. The cost for this move is potentially
deep trees and/or many ARF training rounds, especially
when dependencies between covariates are strong or com-
plex. However, deep forests... | Adversarial Random Forests for Density Estimation and Generative Modeling |
Bordia, S. and Bowman, S. Identifying and reducing gender
bias in word-level language models. In Proceedings of
the 2019 Conference of the North American Chapter of
the Association for Computational Linguistics: Student
Research Workshop, pp. 7–15, 2019.
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,
Dhar... | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
new data, but merely to learn an independence-inducing
partition. Empirically, we find that this is often achieved in
just a single round even with the tolerance δ set to 0.
Formally, we seek a set of splits Θ such that, for all trees b,
j=1 p(xj|θ(cid:96)
leaves (cid:96), and samples x, we have p(x|θ(cid:96)
b).
Call t... | Adversarial Random Forests for Density Estimation and Generative Modeling |
Incorrect Answers
Correct Answers
41.6
43.6
52.2
GSM8K [12]
Accuracy
Data
Example 4.1: A Reasoning Path with Incorrect Answer
Question: Tonya is in a hamburger eating contest. Each hamburger is 4 ounces. Last year the winner ate 84
ounces. How many hamburgers does she have to eat to beat last year’s winner? (Groun... | METAMATH |
[34] Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch,
Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego
de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan
Clark, et al. Training compute-optimal large language mod-
els. arXiv preprint arXiv:2203.15556, 2022. 2, 3
14
[35] Wenyi Hong, Ming Ding, Wendi Zhe... | VideoPoet |
and experiments, authors of the AICAN system showed that people were very often unable to tell the difference between
AICAN-generated images and artworks produced by a human artist [40]. Besides the AICAN initiative, in order to
generate their digital artwork, many other developers and artists employed GANs with variou... | UNDERSTANDINGANDCREATINGARTWITHAI-REVIEWAND OUTLOOK |
32
Table 16: Examples of correct and incorrect chains of thought produced by LaMDA 137B on
StrategyQA. | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
One of the main benefits of structuring knowledge in the form of graphs instead of typical relational settings is the flex-
ibility towards the schema, that maintainers can define at a later stage, and change over time. This allows more flexibility
for data evolution, as well as capturing of incomplete knowledge [9]. Reas... | Knowledge graphs as tools for explainable machine learning: A survey |
the proofs help humans predict model de-
cisions correctly more often than using the
evidence directly.1 | ProoFVer- Natural Logic Theorem Proving for Fact Verification |
PRCA trains the adapter through a context extraction
The retriever’s out-
phase and a reward-driven phase.
put
is then optimized using a token-based autoregres-
sive strategy [Yang et al., 2023b]. The token filtering ap-
proach employs cross-attention scores to efficiently fil-
ter tokens, selecting only the highest-sc... | RAG forLargeLanguageModels-ASurvey |
and formatting. However, we show that SELF-
INSTRUCT stillbringsinadditionalgainswhencom-
bined with the SUPERNI training set, proving its
value as complementary data.
5.4 Experiment 2: Generalization to | SELF-INSTRUCT- Aligning Language Model with Self Generated Instructions |
Of course, research on the impact of technology, in general, and the Internet,
in particular, on democracy is not new. The early utopianism of the Internet
proffered a theory of “liberation technology” – a mode of unimpeded,
transnational communication that would disrupt authoritarian regimes and
promote freedom around... | Social_Media_and_Democracy |
conditional reasoning.
2345–2354, 2020.
Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empiri-
In International Conference on Learning Representations, 2018. URL https:
cal risk minimization.
//openreview.net/forum?id=r1Ddp1-Rb.
Sheng Zhang, Yanbo Xu, Naoto Usuyama, Jaspreet Bagga... | BiomedGPT |
Language models can explain neurons in language models
https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html
22/32 | Language models can explain neurons in language models |
cient for embodied reasoning tasks, as well as limitations of
a recent proposal for grounding language models through
affordances. To overcome these limitations, we proposed
PaLM-E, a single model that is able to control different
robots in simulation and in the real world, while at the same
time being quantitatively c... | PaLM-E- An Embodied Multimodal Language Model |
LLMs and Robotics. LLMs have been applied across several areas in robotics—such as decomposing
high-level task descriptions to mid-level plans [6, 7, 57, 58, 59, 60], robot code [13, 17, 14, 61], and plan-
ning domain definition languages [10]. These methods leverage semantic priors stored in LLMs to compose
plans or p... | LargeLanguageModelsasGeneralPatternMachines |
Impacts will be felt first where the truth is critical, news reporting, legal processes, and public
safety.158 There are examples already of outlets concerned that real images and videos are
19 Frontier AI – Capabilities and Risks
increasingly likely to be regarded as unreliable given they may have been AI generat... | Capabilities and risks from frontier AI |
For the two-operation experiment we drew 120 samples for each of the 29 formats,
which were divided equally between the train, dev and test sets.
Experiment 5 - generalization across the number of operations: | MRKL Systems |
7
2.3 Paradigm Shift
Figure 2: Tool categorization from the perspective of the user interface: (1) physical interaction-based tools,
(b) GUI-based tools, and (c) program-based tools. | Tool Learning with Foundation Models |
Dataset: The dataset name is "ozone-level-8hr". It contains 2 classes, 2534 instances, 73 features,
72 numeric features, 1 categorical features. The majority class size is 2374 and the minority class
size is 160.
Configuration 1: cost is small. gamma is small. kernel is radial.
Configuration 2: cost is very small. gamma ... | MLCopilot- Unleashing the Power of Large Language Models in Solving Machine Learning Tasks |
multiple GPUs and processes input mini-batches as smaller micro-batches. This approach allows for the efficient training of
significantly large models. GPipe also uses a strategy called rematerialization [43] to reduce memory usage by recalculating
activations during backward propagation instead of storing them. Howeve... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
errors in programs, which are valuable for bug fixing (Fig. 5, right);
(3) Self-verification for checking task success. Instead of manually coding success checkers
for each new task proposed by the automatic curriculum, we instantiate another GPT-4
agent for self-verification. By providing VOYAGER’s current state and ... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
def get_latest_tweet ( keyword ):
tweet = tweepy . Cursor ( api . search_tweets , q= keyword , lang =" en "
). items (1)
latest_tweet = ’’
for t in tweet :
latest_tweet = t. text
return latest_tweet
This function takes a keyword as input and returns the latest tweet containing the keyword as
a string. We can use th... | CAMEL- Communicative Agents for “Mind” Exploration of Large Scale Language Model Society |
u
e
i
n
a
n
y
a
r
b
i
t
r
a
r
y
…
…
T
h
e
r
e
l
a
t
i
o
n
t
o
c
o
n
t
r
a
c
t
i
b
l
e
c
o
n
t
r
a
c
t
s
a
n
d
c
o
m
p
l
e
x
m
e
c
h
a
n
i
s
m
s
T
h
e
t
e
r
m
c
o
n
t
r
a
c
t
i
b
l
e
c
o
n
t
r
a
c
t
s
a
s
r
e
f
e
r
r
e
d
t
o
i
n
p
r
e
v
i
o
u
s
l
i
t
e
r
a
t
u
r
e
d
i
f
f
e
r
... | Principal-agent VCG contracts - ScienceDirect |
5https://www.change.org/p/save-sydney-ai
41 | TheRiseandPotentialofLargeLanguageModel BasedAgents |
3.3 Finetuning the LLM on the Evolved Instructions
After all the evolutions are completed, we will merge the initial instruction dataset with all epochs
of evolved instruction data and randomly shuffle the data sample order to form the final fine-tuning
dataset. This processing can ensure that instructions of different d... | WizardLM- Empowering Large Language Models to Follow Complex Instructions |
fields ofelectric cars, space exploration, andrenewable energy. He is also knownfor his eccentric personality andoutspoken views on various topics. Figure 11: Image commenting | MiniGPT-4- Enhancing Vision-Language Understanding with Advanced Large Language Models |
Attitudes toward human augmentation play a crucial role in the adoption and socially acceptable development
of performance-enhancing technologies [82]; the lack of social acceptability of augmentation devices could affect
the self-perception of Augmentation Technologies (ATs) users and hinder the adoption of novel tech... | Society’sAttitudesTowardsHumanAugmentation |
230, see also content takedown
novelty, as main driver of misinformation, 22
NSA (National Security Agency), US, 296
Nyhan, B., 17, 18, 19, 20, 164, 169, 172, 180
Nyss, C., 100
Obama, Barack, 35
offline and online social ties, 39
offline consequences of online speech,
67–71, 241
offline vs. online information exposure... | Social_Media_and_Democracy |
4.2 Some chain-of-thought data is needed to maintain reasoning ability
We next ablate the effect of including just nine CoT datasets in instruction finetuning. We stratify evaluations
into held-out CoT benchmarks (MMLU, BBH, and MGSM) and held-out non-CoT benchmarks (MMLU, BBH,
and TyDiQA) and compute normalized averages... | Scaling Instruction-Finetuned Language Models |
[11] Xilun Chen, Kushal Lakhotia, Barlas O˘guz, Anchit Gupta, Patrick Lewis, Stan Peshterliev,
Yashar Mehdad, Sonal Gupta, and Wen-tau Yih. Salient phrase aware dense retrieval: Can a
dense retriever imitate a sparse one? arXiv preprint arXiv:2110.06918, 2021.
[12] Arman Cohan, Sergey Feldman, Iz Beltagy, Doug Downey,... | E5 |
sensitivity. Prompt template examples are presented in Table 11. Following Brown et al. (2020), we
classify tasks into two question types to get model predictions: 1) For multiple-choice questions,
we compare the per-token likelihood of each option to determine the model prediction; 2) For text
completion questions, we... | ADAPTINGLARGELANGUAGEMODELSVIA READINGCOMPREHENSION |
Y. Jia, M. Johnson, W. Macherey, R. J. Weiss, Y. Cao, C.-C. Chiu, N. Ari, S. Laurenzo, and Y. Wu.
Leveraging weakly supervised data to improve end-to-end speech-to-text translation. In Proc.
ICASSP, pages 7180–7184, 2019a.
Y. Jia, R. J. Weiss, F. Biadsy, W. Macherey, M. Johnson, Z. Chen, and Y. Wu. Direct speech-to-sp... | Translatotron3 |
[114], are capable of directly generating waveforms from text inputs. Compared to concatenative
synthesis 7 and statistical parametric synthesis, neural network-based speech synthesis offers
several advantages including superior voice quality, naturalness, intelligibility, and reduced reliance
on human preprocessing an... | AReviewofDeepLearningTechniquesforSpeechProcessing |
policy to produce responses assigned high reward without drifting excessively far from the original
model. While RLHF produces models with impressive conversational and coding abilities, the RLHF
pipeline is considerably more complex than supervised learning, involving training multiple LMs and
sampling from the LM pol... | Direct Preference Optimization |
the answer to Q1: Marathon is to race as hibernation is to what? And the second word is the answer to Q2: What is running but slower? A: The common phrase is:Input TextFlan-PaLM outputZero-shot reasoningsleep walk Figure 11: More qualitative examples of responses to challenging open-ended questions. | Scaling Instruction-Finetuned Language Models |
UL2 to future work.
For supervised finetuning, we generally adopt a learning rate in the range of {5 × 10−5, 1 × 10−5 1 × 10−4}
using the Adafactor optimizer. The general recipe is that we reset Adafactor optimizer states and/or adopt a
loss normalization based on the number of real target tokens. This is reminiscent of... | UL2- Unifying Language Learning Paradigms |
[17] Meng Cao, Yue Dong, Jiapeng Wu, and Jackie Chi Kit Cheung. 2020. Factual Error Correction for Abstractive
Summarization Models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing
(EMNLP). 6251–6258.
[18] Shuyang Cao and Lu Wang. 2021. CLIFF: Contrastive Learning for Improvin... | SurveyofHallucinationinNatural Language Generation |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.