AI Breakdown

In this episode we discuss On the Connection between Pre-training Data Diversity and Fine-tuning Robustness by Vivek Ramanujan, Thao Nguyen, Sewoong Oh, Ludwig Schmidt, Ali Farhadi. The paper investigates the impact of different factors in pre-training data on the robustness of fine-tuned models. The authors find that the primary factor influencing robustness is data quantity, whereas other factors like label space, image diversity, and data domains have limited significance. The study uses pre-training distributions from natural and synthetic data sources and focuses on the iWildCam-WILDS distribution shift to test downstream robustness.

In this episode we discuss Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness by Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, Kristian Kersting. The paper proposes a strategy called Fair Diffusion to address biases in text-to-image models after deployment. This approach allows users to adjust biases in any direction based on human instructions, enabling the training of generative models on fairness. The authors also conduct an audit of existing text-to-image models for biases and suggest methods to address and mitigate them. Fair Diffusion provides a practical solution for achieving different notions of fairness in generative models.

In this episode we discuss In-Context Pretraining: Language Modeling Beyond Document Boundaries by Weijia Shi, Sewon Min, Maria Lomeli, Chunting Zhou, Margaret Li, Victoria Lin, Noah A. Smith, Luke Zettlemoyer, Scott Yih, Mike Lewis. This paper introduces a new approach called IN-CONTEXT PRETRAINING for training large language models. It addresses the limitation of current LM training pipelines that concatenate random sets of short documents without providing signal for predicting the next document. IN-CONTEXT PRETRAINING reorders the pretraining data by combining semantically related documents to create coherent input contexts, resulting in improved performance in tasks that require complex contextual reasoning.

In this episode we discuss Sigmoid Loss for Language Image Pre-Training by Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, Lucas Beyer. The paper introduces a pairwise Sigmoid loss for Language-Image Pre-training (SigLIP), which operates on image-text pairs and allows for scaling up batch size without the need for global pairwise similarities. By combining SigLIP with Locked-image Tuning, the authors achieve high ImageNet zero-shot accuracy in just two days of training. The authors also discuss the impact of batch size and find that a batch size of 32k is sufficient.

In this episode we discuss Walking Down the Memory Maze: Beyond Context Limit through Interactive Reading by Howard Chen, Ramakanth Pasunuru, Jason Weston, Asli Celikyilmaz. The paper introduces MEMWALKER, an approach to address the limitations of the self-attention mechanism in large language models (LLMs) when processing long sequences. MEMWALKER treats the LLM as an interactive agent that iteratively reads the text, processing the long context into a tree of summary nodes. The model is then able to navigate this tree to gather relevant information and respond to queries. The paper demonstrates that MEMWALKER outperforms existing methods for long-text question answering tasks and enhances explainability by highlighting reasoning steps and relevant text segments.

In this episode we discuss HyperAttention: Long-context Attention in Near-Linear Time by Insu Han, Rajesh Jayaram, Amin Karbasi, Vahab Mirrokni, David P. Woodruff, Amir Zandieh. The paper introduces "HyperAttention," an approximate attention mechanism for handling long contexts in Large Language Models (LLMs). It proposes two parameters to measure problem difficulty and presents a linear time sampling algorithm for attention. Empirical results demonstrate that HyperAttention outperforms existing methods, significantly speeding up inference time while maintaining comparable perplexity. The paper concludes by highlighting the scalability limitations of exact computation in attention layers and discussing the potential of HyperAttention to overcome these limitations.

In this episode we discuss InstructCV: Instruction-Tuned Text-to-Image Diffusion Models as Vision Generalists by Yulu Gan, Sungwoo Park, Alexander Schubert, Anthony Philippakis, Ahmed M. Alaa. The paper proposes a unified language interface for computer vision tasks that allows for task execution through natural language instructions. The approach involves training a text-to-image diffusion model using a multi-modal and multi-task training dataset created through paraphrasing prompt templates. Experimental results show that the model, called InstructCV, performs competitively compared to other vision models and exhibits strong generalization capabilities.

In this episode we discuss Large Language Models Cannot Self-Correct Reasoning Yet by Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, Denny Zhou. The paper explores the effectiveness of self-correction in Large Language Models (LLMs) for improving the accuracy and appropriateness of generated content. It specifically focuses on the role of self-correction in reasoning tasks. The study reveals that LLMs struggle to self-correct without external feedback and, in some cases, their performance declines after self-correction. Possible areas for further research and practical applications in this domain are also discussed.

In this episode we discuss Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution by Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, Tim Rocktäschel. The paper presents PROMPTBREEDER, a method for evolving and adapting prompts for Large Language Models (LLMs) in order to enhance their reasoning abilities. It uses an LLM to mutate a population of task-prompts and evaluates their fitness on a training set. The mutation of task-prompts is guided by self-improved mutation-prompts generated by the LLM, leading to improved performance in tasks such as arithmetic, commonsense reasoning, and hate speech classification.

In this episode we discuss Self-Taught Optimizer (STOP): Recursively Self-Improving Code Generation by Eric Zelikman, Eliana Lorch, Lester Mackey, Adam Tauman Kalai. The paper presents a method called Self-Taught Optimizer (STOP) that utilizes a language model to enhance a scaffolding program for solving optimization problems. The language model suggests self-improvement strategies like beam search, genetic algorithms, and simulated annealing. The study demonstrates the success of STOP by comparing the improved program to its original version in various downstream tasks and analyzes the potential risks associated with bypassing a sandbox in the generated code.