text stringlengths 1 1k ⌀ | title stringclasses 230
values |
|---|---|
Social Media, Echo Chambers, and Political Polarization
51
Allport, G. W. (1954). The Nature of Prejudice. Reading, MA: Addison-Wesley.
Bail, C. A., Argyle, L. P., Brown, T. W. et al. (2018). Exposure to opposing views on
social media can increase political polarization. Proceedings of the National
Academy of Science... | Social_Media_and_Democracy |
39
020406080100% of Harmlessness Training Data0.500.550.600.650.70Mean Test AccMean Test Acc vs. % of Harmlessness Training Data1081091010Parameters020406080100% of Harmlessness Training Data0.750.800.850.900.951.00Normalized Mean Test AccNormalized mean accuracy vs. % of Harmlessness Data1081091010Parameters Figure 2... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
For some low-level intermediate tasks, which are not intended for regular users but rather for high level tasks, such
as named entity recognition (NER) and dependency parsing, there’s not enough result coming from LLMs, because
the most current evaluation of LLMs focuses on practical tasks. According to available evalu... | Harnessing the Power of LLMs in Practice- A Survey on ChatGPT and Beyond |
1
Fine-tuning is the prevalent paradigm for using
large pretrained language models (LMs) (Radford
et al., 2019; Devlin et al., 2019) to perform down-
stream tasks (e.g., summarization), but it requires
updating and storing all the parameters of the LM.
Consequently, to build and deploy NLP systems
that rely on large pr... | Prefix-Tuning |
2.3.2 Automatic metrics
We sampled a total of 500 source texts from each dataset and fed them through each of the evaluated models, using their
respective APIs as described above in Section 2.1. We applied the following automatic metrics to evaluate the results.
Faithfulness: Based on a proprietary classifier developed... | AI21 SUMMARIZE API- TECHNICAL EVALUATION |
(1) Gender Agreement and Coherency Evaluation
For each translation, you will be asked to evaluate if the gender(s) in the translated sentence are correct (and faithful to
the source sentence) based on the main person, entity or people referred to in those sentences. To evaluate the gender,
look for gender-specific words... | PaLM 2 Technical Report |
1 Indeed, though we attempt to provide a comprehensive review of the literature on misinformation
correction, the field is moving so fast that this review may soon be out of date.
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
166
Chloe Wittenberg & Adam J. Berinsky | Social_Media_and_Democracy |
6 Conclusion
In this paper, we present MozArt, a new multilin-
gual dataset of parallel cloze examples with anno-
tations from balanced demographics. This dataset
is, to the best of our knowledge, the first to enable
apples-to-apples comparison of group disparity of
multilingual PLMs across languages. The dataset
inclu... | Are Pretrained Multilingual Models Equally Fair Across Languages? |
4.2 Depth-ambiguity-aware Reconstruction Loss
To train our network, we sample 3D points in 3D space
around the human model, infer their occupancy proba-
bilities and construct a per-point reconstruction loss. The
traditional reconstruction loss is defined as the mean square
error between the predicted occupancy probabil... | PaMIR- Parametric Model-Conditioned Implicit Representation for Image-based Human Reconstruction |
c
h
.
c
o
m
/
i
n
d
u
s
t
r
y
-
a
n
a
l
y
s
i
s
/
w
e
a
l
t
h
-
m
a
n
a
g
e
m
e
n
t
-
s
o
f
t
w
a
r
e
-
m
a
r
k
e
t
W
e
a
l
t
h
s
e
r
v
i
c
e
s
:
h
t
t
p
s
:
/
/
w
w
w
.
| Product-Led AI _ Greylock |
above information and update the 3D scene based on the PIU
strategy. To avoid overfitting and geometric ambiguity during
view-by-view updating, we introduce support sets to provide
multi-view constraints for single-view training in NeRF. More-
over, we adopt depth and transmittance losses along with
the RGB loss to achi... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
[17] A. Lecoutre and B. Negrevergne, and F. Yger, ‘‘Recognizing art style
automatically in painting with deep learning,’’ in Proc. 9th Asian Conf.
Mach. Learn. (ACML), Seoul, South Korea, Nov. 2017, pp. 327–342.
[18] E. Cetinic, T. Lipic, and S. Grgic, ‘‘Fine-tuning convolutional neu-
ral networks for fine art classific... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
During data collection, we must decide which solutions to surface to data-
labelers. The most straightforward strategy is to uniformly surface solutions
produced by the generator. However, if we surface solutions that make obvious
errors, the human feedback we get is less valuable. We would prefer to surface
solutions ... | Let’s Verify Step by Step |
A.5 Data Annotation
We have relied on human annotators in order to collect annotations for the supervised fine-tuning stage and
human preferences to train the reward models. In this section, we provide details about the data annotation
process.
A.5.1 SFT Annotation Instructions
We have collected single-turn and multi-... | Llama2 |
but that there is no clear effect from Flan-PaLM model scale. Interestingly, Flan-T5-XXL seems to perform
differently than Flan-PaLM models at higher prompt toxicity, even producing TPC worse the human baselines
from Gehman et al. (2020). One possibility is that the input prompt toxicity has a more significant influence
on... | Scaling Instruction-Finetuned Language Models |
models on all tasks.
We scale UL2 up to a moderate scale setting of approximately 20B (19.5 to be exact) parameters and run
experiments across a very diverse suite of 50+ NLP tasks ranging from language generation (with automated
and human evaluation), language understanding, text classification, question answering, com... | UL2- Unifying Language Learning Paradigms |
information is usually applied in attention layers. Exam-
ples of such include a learnable attention logit bias as in | Self-Extend LLM |
[78] Sophie Jentzsch and Kristian Kersting. 2023. ChatGPT is fun, but it is not funny! Humor is still challenging Large
Language Models. arXiv preprint arXiv:2306.04563 (2023).
[79] Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Wayne Xin Zhao, and Ji-Rong Wen. 2023. Structgpt: A general
framework for large language... | ASurveyonEvaluationofLargeLanguageModels |
the same output. This process was run on each problem for a maximum of 10 CPU hours or 200
generated tests. Because of complex input formats, we failed to generate the full set of 200 tests for
6.3% of problems. | alphacode |
[25] H.-H. Lee and A. X. Chang, “Understanding pure clip guidance for
voxel grid nerf models,” arXiv preprint arXiv:2209.15172, 2022.
[26] A. Ramesh, M. Pavlov, G. Goh, S. Gray, C. Voss, A. Radford, M. Chen,
and I. Sutskever, “Zero-shot text-to-image generation,” in International
Conference on Machine Learning. PMLR,... | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
[63] Timo Stich, Christian Linz, Georgia Albuquerque, and Mar-
cus Magnor. View and time interpolation in image space. In
Computer Graphics Forum, volume 27, pages 1781–1787.
Wiley Online Library, 2008.
[64] Mohammed Suhail, Carlos Esteves, Leonid Sigal, and
In Proc.
Ameesh Makadia. Light field neural rendering.
Compu... | DynIBaR-NeuralDynamicImage-BasedRendering |
have, as well.”
If anything, AI’s impact will be even greater, says NEA partner Aaron Jacobson. While previous upheavals involved how and where
technology could be used, “AI is actually shifting who does the work,” he says. “That’s never happened before, so this disruption
will be faster, fiercer, and bigger than ever s... | 4 Trends for AI Startups and Generative AI Companies |
38
A.1 Hidden tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
A.2 Program judging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
A.3 Evaluation metrics
41
B.1 GitHub da... | alphacode |
4.2.1 Importance sampling
A current survey [37] notes that importance sampling (data pruning) significantly
influences models’ data efficiency during pre-training. Importance sampling means to
prioritize informative training instances, so it involves estimating per-sample impor-
tance. It is also called data pruning. A ma... | Beyond Efficiency |
Google I/O 2023: Making AI more helpful for everyone
READ ARTICLE
While PaLM 2 is highly capable, it really shines when fine-tuned on domain-specific knowledge. We recently
released Sec-PaLM, fine-tuned for security use cases. It uses AI to better detect malicious scripts, and it can help
security experts understand... | Google I_O 2023_ Making AI more helpful for everyone |
Interestingly, while GPT-2 and GPT-3 were not
trained on the Pile, there still appears to be a clear
scaling law without diminishing returns. We hy-
pothesize that this is due to the inherent generaliza-
tion capability of these models. We leave a more
10While the sizes of GPT-3 models on the OpenAI API have
not been ... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Fake news detection strategies concentrate on apply-
ing news content and social context features [98]. News
the meta-information
content
relevant
in news
validation, news content (linguistics and visual informa-
tion) is used as a feature [99], [100]. Textual features
comprise the writing style and emotion [101],
[102... | A_Comprehensive_Review_on_Fake_News_Detection_With_Deep_Learning |
7
Table 5: Distil-Whisper retains the WER performance of the Whisper model but with faster
inference speed. Average WER results over the four OOD short-form test sets and the four OOD
long-form test sets. Relative latency is the inference time relative to the large-v2 checkpoint. For
short-form evaluation, the batch ... | DISTIL-WHISPER |
Howard, J. and Ruder, S. Universal language model fine-tuning for text classification. In Proceedings of the 56th
Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 328–339, Melbourne,
Australia, July 2018. Association for Computational Linguistics. doi: 10.18653/v1/P18-1031. URL... | PaLM 2 Technical Report |
38M73M244M768M1549M1549MModel parameters0.02.55.07.510.012.515.017.520.0WER on 12 datasets (%)English Speech RecognitionAverageLarge V238M73M244M768M1549M1549MModel parameters020406080100WER on 67 languages (%)Multilingual Speech Recognition (Fleurs)AverageLarge V238M73M244M768M1549M1549MModel parameters01020304050BLEU... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
1. Introduction
The goal of this work is to study the integration and the role of knowledge graphs in the context of Explainable Machine
Learning. Explanations have been the subject of study in a variety of fields for a long time [1], but are experiencing a
new wave of popularity due to the recen... | Knowledge graphs as tools for explainable machine learning: A survey |
That's such a hard problem that (outside of the domain of scene comprehension,
discussed below) most people instead work on other things, and to a surprising extent
try to make do without cognitive models altogether.
§ | The Next Decade in AI- |
models scale better and for our largest experiments outper-
form their English-only counterparts demonstrating positive
transfer from other tasks. For our largest experiments, joint
models also slightly outperform English-only models even
when not adjusting for compute spent per task.
4.4. Text Normalization | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
18
Response Format: Please respond in the format of “Option id. Option content”, for example, “A. xxx”.
Let’s think outside the box. The satisfactory option is
<Response>
where the tag <Question> denotes the text question of Oogiri data.
Instruction Templates of Image&Text to Text. The instruction templates for Imag... | Let’sThinkOutsidetheBox |
return ("0.0.0.0", int(host))
if re.match(r'^(\\w+)://', host) is None:
host = "//" + host
o = parse.urlparse(host)
hostname = o.hostname or "0.0.0.0"
port = o.port or 0
return (hostname, port)
if container.name == name:
return True
return False
def create_evaluate_ops(task_prefix,
data_format,
input_paths,
pre... | CodeLlama2 |
This line of work continues to this day, where large-scale, publicly available knowledge graphs are embedded in more
scalable learning algorithms to visually explain the model behaviour (e.g. visualising hidden states of a neural network).
For example, [56] show how background knowledge from Con... | Knowledge graphs as tools for explainable machine learning: A survey |
These methods are part of a broader trend wherein
in their operations, as
LLMs employ active judgment
seen in model agents like AutoGPT, Toolformer, and
[Yang et al., 2023c, Schick et al., 2023,
Graph-Toolformer
Zhang, 2023]. Graph-Toolformer, for instance, divides its re-
trieval process into distinct steps where LL... | RAG forLargeLanguageModels-ASurvey |
game of Go without human knowledge. Nature, 550(7676), 354-359.
Smolensky, P., Lee, M., He, X., Yih, W.-t., Gao, J., & Deng, L. (2016). Basic Reasoning with Tensor Product
Representations. arXiv, cs.AI.
Spelke, E. (1994). Initial knowledge: six suggestions. Cognition, 50(1-3), 431-445.
Sun, R. (1996). Hybrid Con... | The Next Decade in AI- |
training with limited budget, and over the recipe described in (Izsak et al., 2021). For (Izsak et al.,
2021), the described recipe was originally designed for a full 8 GPU server blade, and squeezing the
BERT-large model therein onto the smaller GPUs in this experiment is resposnsible for most of the
performance degra... | CRAMMING-TRAININGALANGUAGEMODELONA SINGLEGPUINONEDAY |
[47] Ye Zhu, Yu Wu, Zhiwei Deng, Olga Russakovsky, and
Yan Yan. Boundary guided mixing trajectory for se-
mantic control with diffusion models. arXiv preprint
arXiv:2302.08357, 2023. 3
[29] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja
Giryes, and Daniel Cohen-Or. Texture: Text-guided
texturing of 3d shapes, 2023. 3... | A Neural Space-Time Representation for Text-to-Image Personalization |
where “<API>”, “</API>” and “→” are special
tokens.1 Some examples of linearized API calls
inserted into text sequences are shown in Figure 1.
Given a dataset C = {x1, . . . , x|C|} of plain
texts, we first convert this dataset into a dataset
C∗ augmented with API calls. This is done in three
steps, illustrated in Figur... | Toolformer |
Example 10. We reconsider Example 3. The transformation function f1 is M↑ but not M↓ while the transformation function
f3 is M↓ but not M↑. Define a label relation R1 = {a, b} × {c} and the transformation τ1 = (cid:3) f1, R1(cid:4) from G1 to G2. Then
τ1 has both property R↑ and R↓. It also has property C↓, but not ... | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
3.2 Model Selection
After parsing the list of tasks, HuggingGPT next needs to match the tasks and models, i.e., select the
appropriate model for each task in the task list. For this purpose, we first obtain the descriptions of
expert models from the Hugging Face Hub and then dynamically select models for the tasks thro... | HuggingGPT- Solving AI Tasks with ChatGPT and its Friends in Hugging Face |
Haiyang Xu, Ming Yan, Chenliang Li, Bin Bi, Songfang Huang, Wenming Xiao, and Fei Huang. E2e-vlp:
End-to-end vision-language pre-training enhanced by visual learning. In Proceedings of the 59th Annual
Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on
Natural Languag... | BiomedGPT |
Moreover, the interactions between animals and their hold-
ing items could be disordered, for example, the erroneous
orientations of the guitar and katana. In contrast, our In-
stant3D consistently produces sharp renderings and ensures
plausible interactions between entities.
Computation Costs. We compare our computati... | Instant3D |
Robust Speech Recognition via Large-Scale Weak Supervision
data sources and performed manual inspection of these data
sources sorting by a combination of both high error rate and
data source size in order to identify and remove low-quality
ones efficiently. This inspection showed a large amount of
only partially transc... | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
tendencies associated with belief
in misinformation,
in fact,
While political knowledge has been firmly established as a key moderator of
misinformation effects, via its relationship to directionally motivated reasoning,
the jury is still out regarding the role of political ideology and partisanship.
Although conserv... | Social_Media_and_Democracy |
c
a
l
c
u
l
a
t
e
d
a
s
,
w
h
e
r
e
i
s
a
n
o
n
l
i
n
e
a
r
a
c
t
i
v
a
t
i
o
n
f
u
n
c
t
i
o
n
(
s
p
e
c
i
f
i
c
a
l
l
y
G
E
L
U
f
o
r
G
P
T
-
2
)
.
T
h
e
n
e
u
r
o
n
a
c
t
i
v
a
t
i
o
n
i
s
t
h
e
n
u
s
e
d
t
o
u
p
d
a
t
e
t
h
e
r
e
s
i
d
u
a
l
s
t
r
e
a
m
b
y
a
d
d
i
n
... | Language models can explain neurons in language models |
Dialog metrics: Defining effective metrics for dialog models remains an open research topic. Our approach is
inspired by Adiwardana et al. [17], who argued for human-like metrics, such as sensibleness and specificity. Many
automated metrics for dialog models have been studied, including perplexity [16, 17], F1, Hits@1/N ... | LaMDA- Language Models for Dialog Applications |
63
Table 22: Relative frequency of signals within pre-training data
Dimension
Sexuality
gay
homosexuality
lgbt
lesbian
bisexuality
queer
heterosexuality
Race, religion, ethnicity and nationality
americans
japanese people
indian people
jewish people
muslim
asian people
Gender
man
woman
female
girl
boy
male
Grammatica... | PaLM 2 Technical Report |
. then ( response => response.json () )
. then ( data => {
// Set the image source
catImage.src = data [0]. url;
}) ;
}) ;
</script >
</body >
</html >
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
This code uses the The Cat API to get a random cat image URL. When the "Go!" button is clic... | PaLM 2 Technical Report |
Other approaches, such as probabilistic programming, that allow for explicitly-
represented symbolic constraints while at the same time striving to learn from subtle
statistical information, are worth serious consideration.
Taking a step back to stock, the vast majority of what human beings know about the
world is ... | The Next Decade in AI- |
Worse yet, people may come to believe in misinformation even more strongly
post-correction. In particular, retractions that run counter to individuals’ prior
attitudes may bolster beliefs in the original misinformation (what are known as
worldview backfire effects). These worldview backfire effects have their roots in
di... | Social_Media_and_Democracy |
The regulations in 11 CFR 110.11 lay out the disclaimer requirements for all
“public communications,” including certain requirements on size and
readability (disclaimers “must be presented in a clear and conspicuous
manner,” for example). Notably, some advertisements are exempt, such as
bumper stickers, pins, buttons, ... | Social_Media_and_Democracy |
Traditional methods required manual design of prompt templates and verbalizers, which often resulted in sensitive and
varying efficacy. However, recent advancements in prompt learning have led to the automation and optimization of prompt
construction. AutoPrompt [244] introduces a gradient-based approach to automate th... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
6
4 EXPERIMENTS
We report image classification results on MNIST (LeCun, 1998) and CIFAR10 (Krizhevsky & Hinton,
2009). For MNIST, the distilled images are trained with LENET (LeCun et al., 1998), which achieves
about 99% test accuracy if fully trained. For CIFAR10, we use a network architecture (Krizhevsky,
2012) tha... | DATASET DISTILLATION |
Continuously generated through the control diffusion model, we manually screen until we identify 30 unambiguous and
difficulty challenging white cloud images for each category. The difficulty is adjusted by the “controlnet scale,” a coefficient
used to control the intensity of mask control. A higher value implies a str... | Let’sThinkOutsidetheBox |
4948 (86% of the linked mentions). These are processed
with the BERT tokenizer using the lowercase vo-
cabulary, limited to 128 word-piece tokens.
In
addition to the Wikipedia links, we annotate each
sentence with unlinked mention spans using the
mention detector from Section 2.2. These are used
as additional signal fo... | Entities as Experts- Sparse Memory Access with Entity Supervision |
In conjunction with these components, GPT4Video in-
corporates a safety alignment method to ensure that the gen-
erated content adheres to safety standards, thus addressing
the critical issue of content appropriateness in video gener-
ation. This method section delineates the systematic frame-
work and methodologies em... | GPT4Video |
• We introduce Improved RVQGAN a high fidelity universal audio compression model, that
can compress 44.1 KHz audio into discrete codes at 8 kbps bitrate (~90x compression) with
minimal loss in quality and fewer artifacts. Our model outperforms state-of-the-art methods
by a large margin even at lower bitrates (higher co... | RVQGAN |
• Commonsense Reasoning - We use HellaSwag (Zellers et al., 2019), SocialIQA/SIQA (Sap et al., 2019),
PhysicalIQA/PIQA (Bisk et al., 2020), CosmosQA (Huang et al., 2019), AbductiveNLI (Bhagavatula
et al., 2019), CommonsenseQA (Talmor et al., 2018), CommonsenseQA2 (Talmor et al., 2021).
• Long Range Reasoning - We use ... | UL2- Unifying Language Learning Paradigms |
H. Chang, H. Zhang, J. Barber, A. Maschinot, J. Lezama, L. Jiang, M.-H. Yang, K. Murphy,
W. T. Freeman, M. Rubinstein, et al. Muse: Text-to-image generation via masked
generative transformers. arXiv preprint arXiv:2301.00704, 2023. 16
G. Chechik, V. Sharma, U. Shalit, and S. Bengio. Large scale online learning of imag... | A Cookbook of Self-Supervised Learning |
has 32 encoder and 32 decoder layers. This suggests that FA2 should always be incorporated for
Distil-Whisper when operating at higher batch sizes. Using a static key/value cache would result in
a more significant speed-up to the inference time of the decoder. We leave this as future works.
Table 25 reports the RTF on ... | DISTIL-WHISPER |
1.10 ± 0.05
3.21 ± 0.62
2.52 ± 0.44
3.93 ± 0.59
4.22 ± 0.43
then multiply a weight wi to each connection between Stable
Diffusion and ControlNet according to the resolution of each
block wi = 64/hi, where hi is the size of ith block, e.g.,
h1 = 8, h2 = 16, ..., h13 = 64. By reducing the CFG guid-
ance strength , we ca... | AddingConditionalControltoText-to-ImageDiffusionModels |
98
Samuel C. Woolley
understanding computational propaganda
Computational propaganda is specifically defined as “the assemblage of social
media platforms, autonomous agents, and big data tasked with the
manipulation of public opinion” (Woolley and Howard 2016a). Research
on computational propaganda spans the social sc... | Social_Media_and_Democracy |
68.3
71.0
66.3
76.9
70.1
73.0
76.0
77.0
69.2
72.8
76.7
80.2
51.4
52.0
51.6
56.6
57.2
56.4
58.6
60.2
58.6
57.0
58.2
60.2
76.4
79.9
74.1
83.6
76.1
79.2
82.8
84.2
77.2
80.7
83.3
85.3
70.2
76.5
70.0
79.2
72.8
74.8
80.0
78.9
75.2
77.3
79.4
80.2
48.5
48.9
47.2
50.1
48.9
50.4
50.4
52.3
48.3
50.3
50.9
50.7
MPT
Falcon
Ll... | Llama2 |
False refusals. LLMs that are too safe can have a tendency to over-refuse valid claims similar to what was
reported after the release of Llama 2. We specifically asked red teamers to test for this behavior. They
found some limited evidence of false refusals (when not using a system preprompt). False refusals could also... | CodeLlama2 |
F.2 ADDITIONAL EXPERIMENTS ON GPT-3
We present additional runs on GPT-3 with different adaptation methods in Table 15. The focus is on
identifying the trade-off between performance and the number of trainable parameters.
F.3 LOW-DATA REGIME | LORA |
Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Rad-
ford, A., Chen, M., and Sutskever, I. Zero-shot text-to-
image generation. In Meila, M. and Zhang, T. (eds.), Inter-
national Conference on Machine Learning (ICML), 2021.
Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen,
M. Hierarchical text-conditional ... | MusicLM |
As shown in Figure 2, naively prompting GPT-3.5-Turbo for answer augmentation leads to a clear
accuracy saturation. After accuracy saturation, increasing the AnsAug data only yields a limited
performance gain. For instance, using 80K answer augmentation data to train a LLaMA-2 7B model
leads to a 59.6% accuracy, adding... | METAMATH |
Reasoners. arXiv preprint arXiv:2305.19555 (2023).
[53] Aidan Gilson, Conrad W Safranek, Thomas Huang, Vimig Socrates, Ling Chi, Richard Andrew Taylor, David Chartash,
et al. 2023. How does CHATGPT perform on the United States Medical Licensing Examination? the implications of
large language models for medical educati... | ASurveyonEvaluationofLargeLanguageModels |
Facebook, Google, and Twitter have made efforts to bring more transparency to
political advertising and content policy. These efforts certainly demonstrate a
willingness to be more transparent about processes than a decade ago but,
according to many civil society organizations, still do not go far enough.
Efforts to me... | Social_Media_and_Democracy |
the activities of private and civic organizations rather than governments
themselves” (Fung 2013, p. 188), corporations that play outsize roles in
public life should also ideally be as transparent as possible. Nevertheless,
corporate transparency and accountability is generally only demanded once a
corporation appears ... | Social_Media_and_Democracy |
captions when training generative models. We also establish a reproducible baseline performance
profile for a suite of evals that measure prompt following.
This paper focuses on evaluating the improved prompt following of DALL-E 3 as a result of training
on highly descriptive generated captions. It does not cover train... | Improving Image Generation with Better Captions |
Paul Barham, Aakanksha Chowdhery, Jeff Dean, Sanjay Ghemawat, Steven Hand, Daniel Hurt, Michael
Isard, Hyeontaek Lim, Ruoming Pang, Sudip Roy, et al. Pathways: Asynchronous distributed dataflow
for ml. Proceedings of Machine Learning and Systems, 4:430–449, 2022.
James Bradbury, Roy Frostig, Peter Hawkins, Matthew Jam... | gemini_1_report |
𝑝(𝑥) =
𝑝(𝑥|𝑧)𝑝(𝑧)𝑑𝑧.
(26)
Probabilistic latent variable models provide a powerful way to learn a representation that captures
the underlying relationships between observed and unobserved variables, without requiring explicit
supervision or labels. These models involve unobserved latent variables that must b... | AReviewofDeepLearningTechniquesforSpeechProcessing |
1https://openai.com/
6
Technical Report
As for LLMs with 11-50B parameters, the
proposed MetaMath performs the best. Par-
ticularly, on both GSM8K and MATH,
MetaMath achieves higher accuracy than
SFT, RFT, and WizardMath by a large mar-
gin (+7%), demonstrating the effectiveness
of the MetaMath data in improving ma... | METAMATH |
In current debates over the Internet’s impact on global democracy, the prospect
of state regulation of social media has been proffered as a solution to problems
like fake news, hate speech, conspiracy-mongering, and similar ills. For
example, US Senator Mark Warner has proposed a bill that would enhance
privacy protect... | Social_Media_and_Democracy |
thropic researchers and our crowdworkers, as compared to recent similar work such as [Stiennon et al., 2020,
Ouyang et al., 2022].
As an important caveat, our crowdworker distribution was not held fixed throughout this work, and we expect
that crowdworker quality probably improved as the project went on. We mention this... | Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback |
Michael M. Franz is Professor of Government and Legal Studies at Bowdoin
College and Co-Director of the Wesleyan Media Project.
Francis Fukuyama is the Olivier Nomellini Senior Fellow at the Freeman
Spogli Institute for International Studies and the Mosbacher Director of the
Center on Democracy, Development, and the R... | Social_Media_and_Democracy |
To address this need, we introduce the Pile: a
825.18 GiB English text dataset designed for train-
ing large scale language models. The Pile is com-
posed of 22 diverse and high-quality datasets, in-
cluding both established natural language process-
ing datasets and several newly introduced ones.
In addition to its ut... | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
For associable generation, “USER-INPUTs” contains
“Task-specific Prompt” along with two optional conditions,
“Image” and “Condition”. For “Task-specific Prompt”, we
elaborately design several templates for different types of
Oogiri game. See the Appendix for details and there is an
image-2-text (I2T) Oogiri example in ... | Let’sThinkOutsidetheBox |
Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization, 2016.
Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhari-
wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal,
Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Chi... | LORA |
hours of audio or less, and then follows a roughly log-linear
improvement trend till 54,000 hours before also showing | RobustSpeechRecognitionviaLarge-ScaleWeakSupervision |
language models. CoRR, abs/2204.12000, 2022.
[539] Zhang, S., E. Dinan, J. Urbanek, et al. Personalizing dialogue agents: I have a dog, do you
have pets too? In I. Gurevych, Y. Miyao, eds., Proceedings of the 56th Annual Meeting of the
Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-... | TheRiseandPotentialofLargeLanguageModel BasedAgents |
The attention maps for aesthetic prediction tend to cover
larger regions of images, while sentiment and memorabil-
ity usually localize into few smaller peaks. Face and body
regions are commonly triggered for sentiment and memo-
rability. In addition to the attention maps shown in Fig. 2,
in the Supplemental files (Fig... | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
media usage, and Chapter 8 by Wittenberg and Berinsky on correcting
misinformation. | Social_Media_and_Democracy |
11
[9] DeepMind Interactive Agents Team, Josh Abramson, Arun Ahuja, Arthur Brussee, Federico
Carnevale, Mary Cassin, Felix Fischer, Petko Georgiev, Alex Goldin, Mansi Gupta, Tim
Harley, Felix Hill, Peter C Humphreys, Alden Hung, Jessica Landon, Timothy Lillicrap, Hamza
Merzic, Alistair Muldal, Adam Santoro, Guy Scull... | VOYAGER- An Open-Ended Embodied Agent with Large Language Models |
However, the outcome is more ambiguous when considering other features
common to web services. To the extent that disinformation campaigns operate
through advertising channels provided by the platforms, a claim might be made
that companies like Facebook and Google materially contribute to the illegality.
Cases applying... | Social_Media_and_Democracy |
[231] Ozan Sener and Silvio Savarese. 2017. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489 (2017).
[232] Burr Settles. 2009. Active learning literature survey. University of Wisconsin-Madison Department of Computer Sciences.
[233] Burr Settles. 2011. From theorie... | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
a unique andinformed viewpoint, and enjoys collaborating with a progressive, nimble anddecentralized approach to develop real-world solutions and positive user experiencesat every interaction.Objectives of this role:Develop specialized software for specific machine learning (ML) use cases that havebroad applications, s... | Machine Learning Engineer, Fast Optimized Inference - EMEA Remote - Hugging Face |
Generally speaking, we find PMC-LLaMA has more expertise in medical knowledge and
responds better to some cutting-edge professional medical concepts.
Case 1: COPD. According to World Health Organization (WHO1), “Chronic obstructive
pulmonary disease (COPD) is a common and treatable disease characterized by progressive
a... | PMC-LLaMA- Further Finetuning LLaMA on Medical Papers |
6.1 Take-off
Discussions of existential risk from misaligned AI often focus on the transition from some lower (but
still more advanced than today) level of frontier AI capability (call it “A”) to some much higher and
riskier level (call it “B”). Call this transition “take-off.”141
In particular, some of the literature... | Is Power-Seeking AI an Existential Risk? |
What, if any, risks did the task pose to annotators, and were they informed of the risks prior to engagement with the task?
The task required annotators to read text that may contain language that misgenders people references in translations.
Gender agreement is left intentionally open-ended and not defined in solely bi... | PaLM 2 Technical Report |
McKinsey subject matter experts, data scientists,
and analytics translators from QuantumBlack
worked closely with a team from Vistra comprised
of power generation and process experts as well
as front-line operators to understand how the plant
works, what data was available from the sensors
already in place, and w... | an-ai-power-play-fueling-the-next-wave-of-innovation-in-the-energy-sector-may-2022 |
The increase in capabilities and adoption of these models have made the challenges and conse-
quences of those challenges outlined in this card imminent. As a result, we especially encourage
more research into:
• Economic impacts of AI and increased automation, and the structures needed to make the
transition for soc... | gpt-4-system-card |
!"Social media and digital communities: Are there new ways of expressing ourselves using generative tools? New applications like Midjourney are creating new social experiences as consumers learn to create in public. | Generative AI A Creative New World Sequoia Capital |
questions or concerns, please don’t hesitate to reach out to a qualified healthcare
professional or a trusted adult for guidance. | Llama2 |
Abigail See, Peter J. Liu, and Christopher D. Manning. Get To The Point: Summarization with
Pointer-Generator Networks. In Proceedings of the 55th Annual Meeting of the Association for
Computational Linguistics (Volume 1: Long Papers), pp. 1073–1083, Vancouver, Canada, July
2017. Association for Computational Linguisti... | DISTIL-WHISPER |
discriminative model training without actual labeling. The difference between self-supervised
representation learning and unsupervised representation is highlighted in Figure 10. In contrast
to unsupervised representation learning, SSRL techniques are designed to generate these pseudo
labels for model training. The abi... | AReviewofDeepLearningTechniquesforSpeechProcessing |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.