AI in neuroscience
Applying language models and machine learning to neurophysiology data, built on the same open standards as the rest of our work.
Standardized data is what makes AI useful in neuroscience. An agent can only write an analysis for a dataset it can read, and a model can only be trained across labs whose data shares a format. Our AI work builds on NWB and the DANDI Archive, and like everything else we write, it is open source.
Agents for processing steps
Language-model agents that carry out repetitive, well-specified steps in a lab's pipeline, such as drafting conversion code for a new format, checking metadata, or generating analysis notebooks, with a person reviewing the result.
Data curation for foundation models
Assembling standardized, well-annotated neurophysiology data from many labs into training sets, which is the same NWB and DANDI work we do for data sharing, aimed at a different consumer.
Training in agentic code generation
Hands-on training for lab members in using coding agents on scientific software: where they save time, where they fail, and how to check their output.
Facilitating analysis of open neurophysiology data on the DANDI Archive using large language model tools
An agentic chat assistant and a notebook generation pipeline for exploring DANDI datasets with large language models.
Read the paper Open-source toolDANDI Semantic Atlas
A browsable map of the DANDI Archive. Every public Dandiset is placed by the meaning of its metadata, using sentence embeddings of its title, description, and experimental details, so datasets that describe similar science sit near one another even when they use different vocabulary. Topic regions are discovered by clustering, points can be colored by topic or species, and the map rebuilds itself from the DANDI API every night.
Open Open-source toolEncore Science
An agentic science project, in which AI agents carry out scientific work end to end. Encore Science is a public registry of attempts to reproduce the key findings of published papers from their open data, with agents doing the work. An agent reads each paper and writes its claims down before running any analysis, then tries to reproduce them from the dataset the paper deposited, starting with the DANDI Archive. Independent agents check every attempt, and each claim gets a recorded outcome whether or not it reproduced, along with an executed notebook and any problems found in the data itself. A failure to reproduce may be a failure of the agent and is not a judgement of the paper. The same checks can be run on a lab's own data with the encore command.
Open Open-source toolLLM Cost Frontier
A dashboard that tracks how cheaply each level of language-model capability can be bought. It plots the Pareto frontier of the Artificial Analysis Intelligence Index against measured cost per task, shows how that frontier has moved over time, and lists the cheapest model that reaches each capability tier. The data refreshes every week.
Open Contributiontorch_brain
torch_brain is a Python library for various deep learning models designed for neuroscience. We contribute to this open-source project that provides tools for building and training AI models tailored to neuroscience data.
GitHub Teaching · NIH R25NeuroData AI Summer School
A five-year NIH-funded residential summer school on reanalysis of open neurophysiology data and AI methods for neural data.
Funded project The TransmitterShould neuroscientists vibe code?
Exploring the role of AI in scientific programming.
ReadConsidering an AI project?
Tell us what the lab does by hand today and what data it has. A free 30-minute call is enough to tell whether an agent or a model would help.