May 31, 2024

News Highlights

• Imprecise language models are smaller, speedier, and nearly as accurate.

• 1-bit large language models (LLMs) offer a promising solution to AI's energy consumption by being smaller and faster, yet maintaining near-par accuracy with traditional models.

• Better RAG Results with Reciprocal Rank Fusion and Hybrid Search

• Reciprocal Rank Fusion (RRF) and Hybrid Search significantly enhance the performance of Retrieval Augmented Generation (RAG) systems by combining keyword and vector search results for more accurate query responses.

• Legal models hallucinate in 1 out of 6 (or more) benchmarking queries

• AI-driven legal research tools like LexisNexis and Thomson Reuters still produce incorrect information in more than 17% of cases, despite claims of reducing errors compared to general-purpose models like GPT-4.

• Data drift detection methods fail to accurately predict model performance degradation.

• I ran 580 model-dataset experiments to show that, even if you try very hard, it is almost impossible to know that a model is degrading just by looking at data drift results.

• YOLOv5 on FPGA with Hailo-8 and 4 Pi Cameras

• The multi-camera YOLOv5 project on Zynq UltraScale+ leverages the Hailo-8 AI accelerator for enhanced performance in machine vision applications, demonstrating a significant improvement over FPGA-based neural network implementations.

• Superconducting Computer: Imec's plan to shrink datacenters

• Imec's superconducting computer technology promises to reduce data center sizes to that of a shoebox by leveraging superconductors' zero-resistance properties for energy-efficient computing, potentially revolutionizing AI processing and cloud-based training.

• NPGA: Neural Parametric Gaussian Avatars – high-fidelity digital faces

• NPGA introduces a data-driven approach to create high-fidelity, controllable avatars from multi-view video recordings, leveraging 3D Gaussian Splatting for efficient rendering and topological flexibility.

• Incorporating RoPE for positional encoding and Flash Attention could modernize BERT, enhancing its understanding of position and attention efficiency.

• KAN == multi-layer GAM?

• The KAN paper introduces a method for stacking multiple layers of Generalized Additive Models (GAMs), suggesting that KANs can be viewed as multi-layer GAMs with the incorporation of an activation function.

• Lipreading with LipNet: End-to-End Sentence-level Lipreading

• The LipNet model has been re-implemented from scratch to predict sentences by analyzing lip movements, now using a 3DConv-LSTM (bi-directional) architecture instead of the original 3DConv-GRU.

• Machine learning introspection aims to equip AI with human-like intuition for problem-solving, a concept with roots in Newell and Simon's 1958 "General Problem Solver."

• ANAH: Analytical Annotation of Hallucinations in Large Language Models

• ANAH is a bilingual dataset designed to measure and annotate hallucinations in Large Language Models (LLMs) within Generative Question Answering, aiming to address the critical issue of hallucination in LLM applications.

• NVIDIA’s New AI: 5,000x Faster Virtual Worlds!

• NVIDIA's new AI technique transforms text into 3D worlds with unprecedented speed and quality, outperforming previous methods by being 5,000 times faster.

• DITTO-2: Distilled Diffusion Inference-Time T-Optimization for Music Generation

• DITTO-2 significantly speeds up music generation by distilling a pre-trained diffusion model, enabling faster-than-real-time generation with enhanced control over music inpainting, outpainting, intensity, melody, and structure.

• LLMs exhibit sensitivity to the sequence of options in multiple-choice questions

• This observation underscored by experiments with the LLAMA3-8B model and supported by research, including studies found on arXiv and arXiv.