'Mind-captioning' technique can read human thoughts from brain scans

Reading brain activity with advanced technologies is not a new concept. However, most techniques have focused on identifying single words associated with an object or action a person is seeing or thinking of, or matching up brain signals that correspond to spoken words. Some methods used caption databases or deep neural networks, but these approaches were limited by database word coverage or introduced information not present in the brain. Generating detailed, structured descriptions of complex visual perceptions or thoughts remains difficult.

Team develops high-speed, ultra-low-power superconductive neuron device

A research team has developed a neuron device that holds potential for application in large-scale, high-speed superconductive neural network circuits. The device operates at high speeds with ultra-low-power consumption and is tolerant to parameter fluctuations during circuit fabrication.

Paired Neural Network for Matching Experimental and Predicted Infrared Spectra

We present a novel machine learning (ML)-based scoring technique for determining the similarity between experimental and predicted infrared (IR) spectra for identification purposes. IR spectroscopy is a powerful technique used to identify the molecular structure and composition of a sample by measuring the unique vibrational frequency pattern of the molecule’s functional groups. Molecular identifications are often made by comparing experimental and reference spectra. However, the limited number of reference spectra available in spectral libraries can confound the identification process. Alternative identification procedures rely on in silico techniques to simulate spectra for a wide range of molecules. However, scoring spectral similarity between an experimental query and computationally predicted reference remains a significant challenge. Our proposed ML-based scoring technique overcomes these barriers by accurately and efficiently determining spectral similarity.

LLM-Deflate: Extracting LLMs Into Datasets

Large Language Models compress massive amounts of training data into their parameters. This compression is lossy but highly effective—billions of parameters can encode the essential patterns from terabytes of text. However, what’s less obvious is that this process can be reversed: we can systematically extract structured datasets from

Defeating Nondeterminism in LLM Inference

Reproducibility is a bedrock of scientific progress. However, it’s remarkably difficult to get reproducible results out of large language models. For example, you might observe that asking ChatGPT the same question multiple times provides different results. This by itself is not surprising, since getting a result from a language model involves “sampling”, a process that converts the language model’s output into a probability distribution and probabilistically selects a token. What might be more surprising is that even when we adjust the temperature down to 0This means that the LLM always chooses the highest probability token, which is called greedy sampling. (thus making the sampling theoretically deterministic), LLM APIs are still not deterministic in practice (see past discussions here, here, or here). Even when running inference on your own hardware with an OSS inference library like vLLM or SGLang, sampling still isn’t deterministic (see here or here).

Neural Nets vs. Cellular Automata

Neural Nets vs. Cellular Automata

ARC Prize - What is ARC-AGI?

Learn more about the only AI benchmark that measures AGI progress.