🥇 Top 5 AI Papers You Should Read This Week

AlphaSignal

Hey ,

Welcome back to AlphaSignal, where we bring you the latest developments in the world of AI. In the past few days, an impressive number of AI papers have been released, and among them, we have handpicked the top six that truly stand out.

We explore two applications using diffusion models. DreamDiffusion uses real-time EEG signals to reconstruct human visual perception, an exciting extension of prior work using fMRI signals. However, the practical utility of these reconstructions is a matter of debate.

GRADE-IF showcases another diffusion model application in the molecules and proteins field. It addresses the inverse problem of protein folding, demonstrating how generative AI can solve problems in biology.

A standout Language Model-based (LLM) paper, LENS, caught the attention of vision and language processing specialists. It proposes using pre-trained vision modules for detailed image descriptions, which the LLM then turns into accurate answers, matching state-of-the-art methods like Flamingo without additional fine-tuning.

If you have any questions, suggestions, or feedback, simply reply to this email to reach out to us, and we will promptly respond.

☁️ RESEARCH WORDCLOUD

📄 TOP PUBLICATIONS

Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias

Score: 9.9 • Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Krishna, Jiaming Shen, Chao Zhang

Topics
LLMs • Natural Language

Summary
The utilization of Language Model-based data generation has become the norm for fine-tuning language models towards specific tasks. Although the training strategy has undergone extensive study and development, the process of generating training data by leveraging Language Model-based methods has not been thoroughly explored. In this study, the authors introduce AttrPrompt, a novel approach for generating training data that produces diverse data of higher quality while mitigating bias. Comparatively, AttrPrompt demonstrates superior performance while reducing querying costs by a staggering 95% in comparison to the authors' previously mentioned simple class-conditional prompt method, referred to as Simprompt.

To identify and address bias in the generated data, the authors examine the data produced through SimPrompt using an attribute classifier. Their analysis reveals significant bias in the generated data, such as the disproportionately low occurrence of mentions related to Africa in comparison to North America within the NYT news dataset. To counteract this bias, AttrPrompt initially employs Language Models to generate various attributes that could be crucial for the considered dataset. After filtering the attributes through human interaction, the selected attributes are integrated into a more intricate and comprehensive prompt for dataset generation. In addition to specific attributes, the type of attribute values is also obtained by querying the Language Models once again, resulting in plausible answers. Ultimately, randomly selected attributes are merged with the template provided in AttrPrompt to generate a wider array of diverse samples.

Want to promote your company, product, job, or event to 100,000+ AI researchers and engineers? You can reach out here.

DreamDiffusion: Generating High-Quality Images from Brain EEG Signals

Score: 9.3 • Yunpeng Bai, Xintao Wang, Yan-pei Cao, Yixiao Ge, Chun Yuan, Ying Shan

Topics
Generative AI • Neuroscience

Summary
Several recent works demonstrated that it is possible to obtain high-quality reconstructions of what the human was watching solely from fMRI signals captured at the moment. In the work of DreamDiffusion, the authors take a step even further and show the possibility that image reconstruction is also possible from noisy EEG signals.

While EEG signal is much easier to collect than fMRI signals due to the modality being portable, and the cost being much lower, the signal is known to be notoriously noisy. In order to devise a plausible image reconstruction method from such low quality data, the authors propose a 2-step approach.

In the first stage, for representation learning, masked signal modeling is proposed as a means to train an autoencoder such that the encoder can be used later in the pipeline to produce the latents. In the second stage, the latents are used to fine-tune stable diffusion model with two losses: the usual denoising score matching loss, together with an alignment loss that maximizes the similarity of the EEG signal in the embedding space. The resulting model is shown to capture the coarse class-level semantics from the noisy EEG signals.

System-Level Natural Language Feedback

Score: 8.2 • Weizhe Yuan, Kyunghyun Cho, Jason Weston

Details
Natural language • Computation & Language

Summary
Instance-level natural language (NL) feedback has been shown to be effective in improving the performance of machine learning systems, providing much more general and direct feedback than binary or preference feedback. In this work, the authors step further and devise system-level NL feedback by aggregating instance-level NL feedback to improve language generation systems.

Specifically, given a quality checker that decides whether a response is satisfactory or not, the authors generate multiple responses that are unsatisfactory and cluster them into a few groups using k-means. Here, human-in-the-loop is triggered to generate criteria that are written in NL that can be used in the text refiner similar to prompt engineering. Moreover, for each clustered criterion, metrics that can quantify the behavior are also devised. Text refiner can be recursively used until we have a satisfactory response, which constructs a new dataset for supervised fine-tuning.

The resulting system-level NL feedback is shown to be effective with or without additionally using instance-level feedback. Further, in the effort to replace human feedback with LLM feedback, the authors find that while LLM feedbacks are often more verbose and contain less grammatical errors, it is often much less effective than human feedback due to the lack of diversity and being less straight to the point: recapitulating the current gap.

🏅 NOTABLE PAPERS

Supervised Pretraining Can Learn In-Context Reinforcement Learning
Score: 8.2 • Jonathan N. Lee, Annie Xie, Aldo Pacchiano, Yash Chandak, Chelsea Finn, Ofir Nachum, Emma Brunskill

Graph Denoising Diffusion for Inverse Protein Folding
Score: 7.3 • Kai Yi, Bingxin Zhou, Yiqing Shen, Pietro Liò, Yu Guang Wang

Towards Language Models That Can See: Computer Vision Through the LENS of Natural Language
Score: 6.9 • William Berrios, Gautam Mittal, Tristan Thrush, Douwe Kiela, Amanpreet Singh

Thank You