AI Breakdown

ICLR 2023 - DreamFusion: Text-to-3D using 2D Diffusion

In this episode we discuss DreamFusion: Text-to-3D using 2D Diffusion
by Ben Poole, Ajay Jain, Jonathan T. Barron, Ben Mildenhall. The paper presents DREAMFUSION, a method that uses a pretrained 2D text-to-image diffusion model to synthesize 3D objects from text. By optimizing a randomly-initialized 3D model using gradient descent and a loss based on probability density distillation, the authors generate 2D renderings of the model that closely match the input text. This approach eliminates the need for large-scale labeled 3D datasets and modifications to the image diffusion model, showcasing the power of pretrained models as priors for text-to-3D synthesis.

ICML 2023 - A Watermark for Large Language Models

In this episode we discuss A Watermark for Large Language Models
by John Kirchenbauer, Jonas Geiping, Yuxin Wen, Jonathan Katz, Ian Miers, Tom Goldstein. This paper presents a watermarking framework for large language models (LLMs), aiming to embed hidden signals in the generated text while remaining undetectable to humans. The approach involves selecting specific tokens and promoting their use in text generation. The proposed watermark is tested on a multi-billion-parameter LLM and its robustness and security are discussed, highlighting the need to detect and audit machine-generated text to mitigate potential malicious use.

arxiv Preprint - Skeleton-of-Thought: Large Language Models Can Do Parallel Decoding

In this episode we discuss Skeleton-of-Thought: Large Language Models Can Do Parallel Decoding
by Xuefei Ning, Zinan Lin, Zixuan Zhou, Huazhong Yang, Yu Wang. The paper proposes a method called "Skeleton-of-Thought" (SoT) to decrease the generation latency of large language models (LLMs). The sequential decoding approach used in current LLMs contributes to high latency. SoT guides LLMs to first generate the skeleton of the answer and then completes the contents of each skeleton point in parallel through API calls or batched decoding.

ICLR 2023 - Mastering the Game of No-Press Diplomacy via Human-Regularized Reinforcement Learning and Planning

In this episode we discuss Mastering the Game of No-Press Diplomacy via Human-Regularized Reinforcement Learning and Planning
by Anton Bakhtin, David J Wu, Adam Lerer, Jonathan Gray, Athul Paul Jacob, Gabriele Farina, Alexander H Miller, Noam Brown. The paper introduces a strategy called DiL-piKL that combines human imitation learning with reinforcement learning and planning to improve performance in the game of No-press Diplomacy. This algorithm regularizes a reward-maximizing policy towards a policy learned from human imitation, resulting in a no-regret learning algorithm. Building upon DiL-piKL, the paper proposes an extended self-play reinforcement learning algorithm called RL-DiL-piKL, which trains an agent that responds well to human play while also modeling human behavior.

arxiv Preprint - RLCD: Reinforcement Learning from Contrast Distillation for Language Model Alignment

In this episode we discuss RLCD: Reinforcement Learning from Contrast Distillation for Language Model Alignment
by Kevin Yang, Dan Klein, Asli Celikyilmaz, Nanyun Peng, Yuandong Tian. The paper presents a method called Reinforcement Learning from Contrast Distillation (RLCD) for aligning language models to natural language principles. RLCD trains a preference model using simulated preference pairs and uses reinforcement learning to improve an unaligned language model. Experimental results show that RLCD outperforms existing methods in three alignment tasks, confirming its effectiveness.

arxiv Preprint - DoG is SGD’s Best Friend: A Parameter-Free Dynamic Step Size Schedule

In this episode we discuss DoG is SGD's Best Friend: A Parameter-Free Dynamic Step Size Schedule
by Maor Ivgi, Oliver Hinder, Yair Carmon. The paper presents a dynamic SGD step size formula called DoG that does not require manual tuning. The authors analyze the DoG formula and demonstrate its strong convergence guarantees for stochastic convex optimization. Empirical evaluation shows that DoG performs comparably to SGD with tuned learning rate and even outperforms tuned SGD in a per-layer variant.

CVPR 2023 - LAVENDER: Unifying Video-Language Understanding as Masked Language Modeling

In this episode we discuss LAVENDER: Unifying Video-Language Understanding as Masked Language Modeling
by Linjie Li, Zhe Gan, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Ce Liu, Lijuan Wang. The paper presents LAVENDER, a unified video-language framework that uses Masked Language Modeling (MLM) as the common interface for pre-training and downstream tasks. LAVENDER simplifies the model architecture by using a lightweight MLM head on top of the multimodal encoder. Surprisingly, experimental results show that LAVENDER achieves competitive performance on various video-language benchmarks.

NeurIPS 2022 - Language Models with Image Descriptors are Strong Few-Shot Video-Language Learners

In this episode we discuss Language Models with Image Descriptors are Strong Few-Shot Video-Language Learners
by Zhenhailong Wang, Manling Li, Ruochen Xu, Luowei Zhou, Jie Lei, Xudong Lin, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Derek Hoiem, Shih-Fu Chang, Mohit Bansal, Heng Ji. VidIL is a few-shot video-language learner that combines image and language models to generalize to different video-to-text tasks with limited examples. It translates video content into frame captions, object, attribute, and event phrases, and combines them into a temporal-aware template. The language model is then prompted with a few in-context examples to generate a target output. Experimental results show that VidIL outperforms supervised models on video future event prediction.

arxiv preprint - MM-REACT: Prompting ChatGPT for Multimodal Reasoning and Action

In this episode we discuss MM-REACT: Prompting ChatGPT for Multimodal Reasoning and Action
by Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, Lijuan Wang. The paper introduces MM-REACT, a system that combines ChatGPT with expert vision models to tackle challenging visual tasks. MM-REACT utilizes a unique prompt design to enable language models to process multimodal information and interact with vision experts. Zero-shot experiments demonstrate the effectiveness of MM-REACT in achieving advanced visual understanding capabilities beyond existing models.

arxiv preprint - 3D-LLM: Injecting the 3D World into Large Language Models

In this episode we discuss 3D-LLM: Injecting the 3D World into Large Language Models
by Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, Chuang Gan. The paper proposes a new model called 3D-LLMs that integrates the 3D physical world into language models, allowing them to perform various 3D-related tasks such as captioning, question answering, and navigation. The authors employ three prompting mechanisms to collect a large dataset of 3D-language data efficiently and use a 3D feature extractor and 2D VLMs as the backbone for training the model. The experimental results demonstrate that the 3D-LLMs outperform existing baselines in terms of performance and capabilities.