Hey Folks!

This newsletter will discuss some cool AI research papers and AI tools. Happy learning!

➡️ Researchers from Stanford, NVIDIA, and UT Austin Propose Cross-Episodic Curriculum (CEC): A New Artificial Intelligence Algorithm to Boost the Learning Efficiency and Generalization of Transformer Agents

How can the learning efficiency and generalization of Transformer agents be enhanced? This paper introduces the Cross-Episodic Curriculum (CEC) algorithm, which embeds cross-episodic experiences into a Transformer's context to form an evolving curriculum. By arranging online learning sessions and mixed-quality demonstrations sequentially, CEC crafts curricula that illustrate the progression and skill enhancement across episodes. When merged with the robust pattern recognition of Transformers, this results in an impactful cross-episodic attention mechanism. The algorithm proves effective in scenarios like multi-task reinforcement learning with discrete control in DeepMind Lab and imitation learning with varied data quality in RoboMimic. CEC-trained policies consistently outperform others in performance and generalization.

➡️  Can We Generate Hyper-Realistic Human Images? This AI Paper Presents HyperHuman: A Leap Forward in Text-to-Image Models

Given the limitations of existing models like Stable Diffusion and DALL-E 2, how can we improve the hyper-realism of generated human images? Addressing this, this paper introduces "HyperHuman", a novel framework designed to harness the inherent multi-level structure of human images, from the broad body skeleton to intricate spatial geometry. The approach involves creating a vast dataset, "HumanVerse", comprising 340M images annotated with pose, depth, and surface details. Using this, the proposed "Latent Structural Diffusion Model" concurrently denoises and synthesizes RGB images, integrating appearance, spatial relationships, and geometry. Additionally, a "Structure-Guided Refiner" enhances resolution and visual quality. The results reveal unparalleled hyper-realism in generated human images across varied contexts.

➡️  Researchers from the National University of Singapore propose Show-1: A Hybrid Artificial Intelligence Model that Marries Pixel-Based and Latent-Based VDMs for Text-to-Video Generation

How can we efficiently generate high-quality videos with precise text-video alignment from large-scale pre-trained Diffusion Models (VDMs)? Addressing existing limitations in pixel-based and latent-based VDMs, this paper introduces "Show-1", a pioneering hybrid model. Show-1 combines the best of both VDMs: it initially employs pixel-based VDMs to create a low-resolution video with robust text-video correlation, then leverages a novel expert translation technique with latent-based VDMs to upscale this video to high resolution. The result is a system that offers precise alignment akin to pixel VDMs but at a fraction of the computational cost, with GPU memory usage dropping dramatically from 72G to 15G during inference. This model outperforms on standard benchmarks.

➡️ Researchers from UC Berkeley Propose RingAttention: A Memory-Efficient Artificial Intelligence Approach to Reduce the Memory Requirements of Transformers

Can Transformers efficiently handle long sequences given their memory constraints? Addressing this limitation, this new paper from UC Berkeley introduces "Ring Attention", a novel approach that employs blockwise computation of self-attention. This technique disperses lengthy sequences across multiple devices, synchronizing the communication of key-value blocks with blockwise attention computation. As a result, Ring Attention permits training and inference of sequences considerably longer than previous memory-efficient Transformers, essentially bypassing the individual device memory restrictions. Experimental outcomes from language modeling tasks validate Ring Attention's capacity to accommodate larger sequence input sizes while enhancing performance.

➡️ Researchers from NVIDIA Introduce Retro 48B: The Largest LLM Pretrained with Retrieval before Instruction Tuning

Can the effectiveness of instruction tuning and zero-shot generalization in auto-regressive LLMs be enhanced by further pretraining with retrieval? This paper from NVIDIA unveils "Retro 48B", an expansion on the existing pretrained retrieval-augmented LLMs, such as the 7.5B parameter Retro. By continuing to pretrain the 43B GPT model using the Retro augmentation method on an additional 100 billion tokens and retrieving from 1.2 trillion tokens, the resulting Retro 48B model significantly surpasses the original 43B GPT in perplexity. Post-instruction tuning, the InstructRetro model demonstrates marked improvements in zero-shot QA tasks. Notably, omitting the encoder from InstructRetro achieves similar outcomes, suggesting that retrieval pretraining bolsters the decoder's context processing for QA.

✅ Featured AI Tools For You

✅ Featured AI Startups For You

• Meet Unitary AI: A UK Based AI Startup Revolutionizing Content Moderation with a Multimodal Approach

• Meet E2B: A Startup Developing the Cloud for AI Agents

• Revolutionizing Digital Identities: Prins AI's Leap into the Future of AI

• Podcastle's Magic Dust AI Transforms Podcasting with Studio-Quality Sound

• Meet Character.AI: An AI Startup that Introduces Group Chats where People and Multiple AIs can Talk to Each Other

• Meet Pano AI: A Wildfire Detection Startup Powered by AI to Provide Actionable Intelligence for Wildfire Management