Mo Shakiba

Scientific objectives
The initial state of an animal’s brain at birth isn’t a blank slate¹. Evolution has already built into the brain inductive biases and structures that facilitate learning and development in the animal’s natural environment and ecological niche²³. This is most strikingly revealed by the speed of learning: animals master ecologically relevant tasks—navigation, predator avoidance, foraging—in just a handful of trials, while the same animals require thousands of trials to learn arbitrary laboratory tasks². Equally striking is the ability to acquire new tasks without erasing previously learned ones—a failure mode known as catastrophic forgetting that plagues modern artificial neural networks (ANNs)⁴.

Yet no ANN has been simultaneously constrained by both the structural wiring and functional activity of a real cortical circuit—and evaluated on whether those constraints confer the learning advantages animals display in their ecological niche. The MICrONS dataset⁵ is a uniquely relevant resource for addressing this gap. It provides detailed anatomical information on over 200,000 cells and 523 million synapses, alongside two-photon functional imaging of approximately 75,000 neurons responding to visual stimuli.

We hypothesize that reproducing the structural and functional inductive biases of the visual cortex in an ANN will yield a model that learns faster, more efficiently, and forgets less on ecologically relevant tasks—compared to unconstrained baselines.

Methodology
We propose a family of ANNs sharing a core architecture: a convolutional neural network (CNN) front-end feeding into multiple recurrent neural network (RNN) modules, each corresponding to a distinct visual area. Models are constrained by MICrONS data in two ways—structurally, via connectivity initialized from anatomical wiring with spatial embedding of real neuronal coordinates⁶; and functionally, via a novel initialization method derived from neuronal activity correlations⁷.

We pre-train the networks with two objectives: (1) a self-supervised predictive coding loss on raw visual input, motivated by evidence that such objectives naturally produce ventral/dorsal functional specialization⁸; and (2) a representational alignment loss that drives network activations toward MICrONS-recorded neuronal responses. The pre-trained models are then evaluated on an ecologically relevant visual foraging task⁹, against structurally and functionally unconstrained baselines, to isolate each constraint’s contribution to learning efficiency and continual learning performance.

We conduct four analyses to mechanistically dissect why biological constraints would confer a learning advantage. First, systematic ablation—removing spatial embedding and functional initialization—identifies which constraints are load-bearing. Second, we characterize the dynamical regime of each model at initialization by measuring the spectral radius of the connectivity matrix, testing whether biological constraints naturally place the network near criticality—a regime associated with maximal information transmission and trainability¹⁰. Third, we apply the kernel alignment framework¹¹ to ask whether the dominant eigenmodes of the bio-constrained models are pre-aligned with the task kernel of the foraging environment—providing a principled, geometric explanation for any sample efficiency advantage. Finally, continual learning is quantified by sequentially exposing models to distinct foraging environments and computing the forgetting measure across biologically constrained and unconstrained baselines.

Originality and contribution
This project is the first to simultaneously constrain an artificial neural network by both the structural wiring and the functional activity of a real brain circuit—and evaluate whether those constraints produce the learning advantages seen in biological systems in their ecological niche. Previous work has imposed one or the other; none has done both, and none has tested the result on tasks the animal actually evolved to perform. The timing is not coincidental. The MICrONS dataset—released in 2025—is the first complete reconstruction of a mammalian cortical circuit with matched large-scale recordings. We are at the first moment it is possible to ask this question rigorously.

If the hypothesis holds, the implications are twofold. First, it would provide direct empirical evidence for a longstanding conjecture in neuroscience: that evolution has compressed the animal’s environment into its neural wiring. Second, it would reframe a central question in artificial intelligence: rather than pursuing general-purpose systems without regard for the tasks they must solve, we can look to biology—already optimized over hundreds of millions of years for specific ecological problems—as a principled design guide for building artificial systems that learn efficiently and adapt without forgetting—a central unsolved problem in modern artificial intelligence (AI).

If the hypothesis fails, the result is equally informative: biological structure alone does not suffice, and the burden of the brain’s learning efficiency falls on dynamics and plasticity—a finding that would reshape how we study continual learning. Ultimately, this project treats the mouse visual cortex not merely as a biological curiosity, but as a two-hundred-million-year-old solution to the problem of efficient, robust learning—one we are, for the first time, in a position to decode and translate into the design of artificial systems.

References