AI Breakdown

arxiv Preprint - An Image is Worth Multiple Words: Learning Object Level Concepts using Multi-Concept Prompt Learning

In this episode we discuss An Image is Worth Multiple Words: Learning Object Level Concepts using Multi-Concept Prompt Learning
by Chen Jin, Ryutaro Tanno, Amrutha Saseendran, Tom Diethe, Philip Teare. The paper proposes a framework called Multi-Concept Prompt Learning (MCPL) to address the challenge of integrating multiple object-level concepts within one scene using prompt learning. The authors introduce three regularization techniques to enhance word-concept correlation. The MCPL framework is evaluated through image generation, editing, and attention visualization, and is compared to a previous approach that can only learn a single concept from each image.

arxiv Preprint - Retrieval meets Long Context Large Language Models

In this episode we discuss Retrieval meets Long Context Large Language Models
by Peng Xu, Wei Ping, Xianchao Wu, Lawrence McAfee, Chen Zhu, Zihan Liu, Sandeep Subramanian, Evelina Bakhturina, Mohammad Shoeybi, Bryan Catanzaro. This paper compares two methods for handling long context in large language models (LLMs): retrieval-augmentation and extending the context window. The study finds that LLMs with a 4K context window using retrieval-augmentation achieve similar performance to LLMs with a 16K context window through positional interpolation, while requiring less computation. Moreover, the authors demonstrate that retrieval significantly improves LLM performance regardless of the context window size.

arxiv Preprint - Contrastive Prefence Learning: Learning from Human Feedback without RL

In this episode we discuss Contrastive Prefence Learning: Learning from Human Feedback without RL
by Joey Hejna, Rafael Rafailov, Harshit Sikchi, Chelsea Finn, Scott Niekum, W. Bradley Knox, Dorsa Sadigh. Traditional approaches to Reinforcement Learning from Human Feedback (RLHF) assume that human preferences align with reward, but recent research suggests they align with regret under the user's optimal policy. This flawed assumption complicates the optimization of the learned reward function using RL. Contrastive Preference Learning (CPL) is proposed as a new approach that learns optimal policies directly from preferences without the need for RL, using maximum entropy and a contrastive objective. CPL is off-policy, applicable to various problems, and can handle high-dimensional and sequential RLHF tasks.

arxiv Preprint - BitNet: Scaling 1-bit Transformers for Large Language Models

In this episode we discuss BitNet: Scaling 1-bit Transformers for Large Language Models
by Hongyu Wang, Shuming Ma, Li Dong, Shaohan Huang, Huaijie Wang, Lingxiao Ma, Fan Yang, Ruiping Wang, Yi Wu, Furu Wei. The paper introduces BitNet, an architecture for large language models that addresses concerns about energy consumption and deployment challenges. BitNet utilizes 1-bit weights and introduces a BitLinear layer to replace the nn.Linear layer. Experimental results show that BitNet achieves competitive performance while reducing memory footprint and energy consumption. It also exhibits a scaling law similar to full-precision Transformers, suggesting its potential for scaling to larger language models efficiently. Detailed graphs and tables are provided to showcase the advantages of BitNet in terms of model size, energy cost reduction, and loss.

arxiv Preprint - Automatic Prompt Optimization with ”Gradient Descent” and Beam Search

In this episode we discuss Automatic Prompt Optimization with "Gradient Descent" and Beam Search
by Reid Pryzant, Dan Iter, Jerry Li, Yin Tat Lee, Chenguang Zhu, Michael Zeng. The paper introduces ProTeGi, a method for improving prompts used in large language models. It utilizes mini-batches of data to generate "natural language gradients" that provide feedback on the prompt. ProTeGi uses beam search and bandit selection to efficiently modify the prompt, resulting in improved performance on benchmark NLP tasks and a novel LLM jailbreak detection problem. This method reduces manual effort and enhances task performance by automatically optimizing prompts.

arxiv Preprint - Understanding Retrieval Augmentation for Long-Form Question Answering

In this episode we discuss Understanding Retrieval Augmentation for Long-Form Question Answering
by Hung-Ting Chen, Fangyuan Xu, Shane A. Arora, Eunsol Choi. This paper examines the impact of retrieval-augmented language models on long-form question answering. The authors compare the generated answers using the same evidence documents to analyze how retrieval augmentation affects different language models. They also investigate the quality of the retrieval document set and its effect on the generated answers.

arxiv Preprint - On the Connection between Pre-training Data Diversity and Fine-tuning Robustness

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.

arxiv Preprint - Fair Diffusion: Instructing Text-to-Image Generation Models on Fairness

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.

arxiv Preprint - In-Context Pretraining: Language Modeling Beyond Document Boundaries

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.

ICCV 2023 - Sigmoid Loss for Language Image Pre-Training

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.