Our Work
Intent-aligned AI systems deplete human agency
Presented at ICML as a spotlight poster, this work highlights the neuroscience of agency and the impossibility of making AI systems safe by
aligning them to human intent. It provides a conceptual algorithm for separation of utility from human agency optimization as
the only sound approach to developing and deploying AI systems. We also published a longer-form version on
arXiv.
[View ICML paper]
The path of scalable Whole Brain Emulation
Our roadmap paper lays out the foundation for building functional models of the brain from mesoscale to microscale
neural data, highlighting technical bottlenecks and challenges. It also discusses metrics for evaluating whole-brain neural foundation models
against biological organisms.
[View paper on arXiv]
We have started building mesoscale neural foundation models (FM) of cortex-wide neural dynamics based on our previous work published in eLife. We now are training the first generation of neuro-behavioural FMs on neural datasets of millions of datapoints from mesoscale widefield [ca] imaging.
Our preliminary findings show that mesoscale neuro-behavioural FMs - aka proto-WBEs - can capture unique neurodynamics present in real but not shuffled or "control" datasets. These findings support the path of scalable WBEs for achieving whole-brain neuro-behavioural dynamical models and we are in the process of expanding this work to include 10s of millions of neurobehavioural datapoints and matching behavioural data.
