NSF AI Institute for Artificial and Natural Intelligence (ARNI) is excited to cohost the Workshop on Emerging Trends in AI on May 5-6.

The workshops will dive into two key emerging themes:
🧠 The connection between brain resilience and algorithmic robustness
📊 The impact of ML-generated data in the social sciences and its potential to accelerate scientific discovery

On day 1, ARNI Director, Richard Zemel, will be moderating a panel focused on how insights from machine learning and neuroscience can inform one another in the pursuit of building more resilient and robust systems.

On day 2, there will be a second panel discussing the ethical and practical implications of synthetic data in shaping research and policy

AER Symposium

The ARNI Club, ARNI Emerging Researchers, will host its very first student Symposium on April 25th!

This event welcomes all ARNI trainees and graduate students ONLY involved in ARNI-related projects. The symposium is designed to support career development and foster valuable networking opportunities. Additionally, it aims to create a space for open dialogue around complex, challenging, and sometimes controversial topics at the intersection of intelligence and NeuroAI.

If you are an ARNI graduate student or postdoc and want to register, please email [email protected]. 

A groundbreaking moment in Neuroscience and AI!

Huge congratulations to our own Andreas Tolias and XaqPitkow (https://xaqlab.com/) and the entire MICrONS team on the release of a first-of-its-kind functional foundation model of the mouse visual cortex—based on neural activity from ~75,000 neurons and detailed electron microscopy mapping.

This remarkable model was trained across data from multiple mice and cortical areas an showed robust generalization to new neurons, animals, and even entirely new stimulus domains. It also predicted new properties like anatomically defined cell types. The team built functional digital twins of mouse brains, revealing that neurons don’t connect randomly—even when they physically can. Instead, they preferentially form connections based on functional similarity (“what” they do) rather than just spatial proximity (“where” they are). This challenges long-standing assumptions in cortical wiring.

This foundational work represents a breakthrough in systematically decoding neural activity, with profound implications for neuroscience and the future of AI- for understanding natural intelligence and building brain-inspired AI. ARNI can leverage this publicly available dataset to train and test our own functional models, and use the anatomy to provide more structure to these
models.

Congratulations to the incredible collaboration across institutions including Stanford, BCM, Princeton, Allen Institute, and others, and to IARPA and the BRAIN Initiative for supporting this ambitious project.

This body of work is published in Nature:
"Foundation model of neural activity predicts response to new stimulus types"
"Functional connectomics spanning multiple areas of mouse visual cortex?"

ARNI at COSYNE2025: NeuroFM Workshop

ARNI postdoctoral fellows Mehdi Azabou and Cole Hurwitz are among the co-organizers of the workshop "Building a Foundation Model for the Brain: Datasets, Theory, and Models" at COSYNE. This workshop is bringing together experts who are at the forefront of neuro-foundation model research.

Workshop Summary:
Advances in neurotechnology have enabled the collection of large-scale neural recordings during animal behavior. To extract insights from these datasets, researchers are now faced with significant challenges when comparing data across different brain regions, subjects, and tasks. Inspired by recent successes in large-scale "foundation" modeling of natural language and computational biology, there has been a push towards developing "neuro-foundation models" that can be trained across diverse datasets and then fine-tuned for downstream tasks and analyses including neural encoding, decoding, cell-type classification, and activity prediction. In this workshop, we aim to provide an overview of current neuro-foundation model approaches and to bring together experimentalists, theoreticians, and model builders to discuss how this paradigm can lead to a deeper understanding of the brain. In particular, we plan to highlight the utility of neuro-foundation models in analyses across recording modalities, brain regions, individuals, species, and contexts. In addition to these talks, we will also host a panel to discuss challenges, opportunities, and risks of building these models.

ARNI Working Groups

• Summary: Neural network models are typically designed with a fixed architecture, determining the number of nodes, connectivity, and timesteps for backpropagation. While this approach helps limit resource requirements and optimize performance, it restricts the ability to explore tradeoffs between space, time, energy, and error. To address this, we aim to develop methods for quantifying resource costs and optimizing models to balance multiple constraints efficiently, benefiting both neuroscience and AI development.
• Next session: Event details

We would like to highlight ARNI achievements in future newsletters. Please share with us your events, papers, presentation, or any news you want to share with the ARNI community!