Oct 20, 2024

• AI engineers claim new algorithm reduces AI power consumption by 95%

• BitEnergy AI's new algorithm, Linear-Complexity Multiplication (L-Mul), replaces complex floating-point multiplication with integer addition, achieving a potential 95% reduction in AI power consumption while maintaining accuracy.

• Why do random forests work? They are self-regularizing adaptive smoothers

• Tree ensembles, viewed as adaptive and self-regularizing smoothers, enhance prediction smoothness beyond individual trees, adjusting their smoothness based on input dissimilarity at test-time, which offers a fresh perspective on their effectiveness.

• Implementing neural networks on the "3 cent" 8-bit microcontroller

• Neural networks can achieve over 90% accuracy on MNIST inference using the ultra-low-cost PMS150C microcontroller, demonstrating that effective machine learning can be implemented even in constrained environments with minimal resources.

• The Languages of English, Math, and Programming

• The Triplets.ipynb notebook explores how different AI models perform in solving the problem of listing distinct positive integer triplets that multiply to 108, revealing that 7 out of 9 models succeeded when prompted to write a program, compared to only 2 out of 9 for a direct listing request.

• Sabotage Evaluations for Frontier Models

• Anthropic's new evaluations assess AI models for potential sabotage capabilities, including misleading human decisions and inserting bugs into code, to prepare for future risks as AI systems become more advanced.

• [Project] Tsetlin Machine for Deep Logical Learning and Reasoning With Graphs (finally, after six years!)

• The Graph Tsetlin Machine is a novel deep learning model that enables multimodal learning and reasoning across directed and labeled multigraphs, marking a significant advancement after six years of development.

• All-optical switch device paves way for faster fiber-optic communication

• An ultrafast all-optical switch developed by a University of Michigan team utilizes circularly polarized light to control light signals without electrical conversion, enhancing speed and energy efficiency in fiber-optic communication.

• [P] NHiTs: Deep Learning + Signal Processing for Time-Series Forecasting

• NHiTs is a state-of-the-art deep learning model for time-series forecasting, effectively integrating past observations, future known inputs, and static exogenous variables to enhance predictive accuracy.

• [D] Last Week in Medical AI: Top LLM Research Papers/Models (October 12 - October 19)

• MedLFQA introduces a benchmark dataset for evaluating the factuality of long-form answers from medical LLMs, enhancing the reliability of AI-generated medical information.