Standard deep learning operates largely on the philosophy of the tabula rasa. We initialize a massive, overparameterized network with uniform or random connectivity—such as a fully connected multilayer perceptron or an all-to-all attention block—and rely on backpropagation to carve out functional pathways from data. While flexible, this approach requires vast amounts of training data and often yields uninterpretable "black boxes" that fail to generalize outside their narrow training distribution.
In contrast, biological brains are born with highly structured, non-random wiring diagrams. Millions of years of evolution have hard-coded specific physical layouts, layer hierarchies, and modular pathways that act as powerful computational priors. The research direction explored here—Anatomical Inductive Biases—seeks to bake these biological blueprints directly into the architecture of artificial neural networks. By constraining how information can flow, we can build models that are more sample-efficient, robust, and inherently interpretable.
Microscopic Constraints: Hardwiring the Connectome
The most granular way to introduce an anatomical bias is to copy the exact wiring diagram, or connectome, of a biological organism. Instead of allowing a weight matrix to be dense and fully unconstrained, researchers use empirical synaptic maps to mask the network's parameters, forcing non-zero weights to exist only where physical connections occur in nature.
Why does this work? It drastically shrinks the optimization search space to configurations that are evolutionarily predisposed to support specific functional dynamics. For instance, Lu Mi et al. demonstrated this in their Connectome-constrained Latent Variable Model of Whole-Brain Neural Activity, where masking synaptic weights based on the C. elegans nervous system allowed the model to accurately infer the activity of unobserved neurons and generalize across individuals far better than purely data-driven baselines. Similarly, Sunny Duan et al. showed in From Synapses to Dynamics: Obtaining Function from Structure in a Connectome Constrained Model of the Head Direction Circuit that a fruit fly's connectome acts as a precise structural prior capable of supporting complex continuous attractor dynamics with minimal parameter tuning; shuffling these weights entirely destroys the circuit's functional capacity.
Remarkably, these evolved structures possess a form of structural intelligence that transfers even to completely unrelated tasks. Shuo Yu et al. mapped the complete connectome of a fruit fly larva into a digital architecture in Biological Processing Units: Leveraging an Insect Connectome to Pioneer Biofidelic Neural Architectures, finding that these fixed-weight biological networks outperform standard multi-layer perceptrons on tasks like image recognition and chess. When applied to sensory processing, Tianfang Zhu et al. built a spiking neural network constrained by the 13 neuronal subtypes and connectivity matrices of the mouse barrel cortex in Biologically Constrained Barrel Cortex Model Integrates Whisker Inputs and Replicates Key Brain Network Dynamics, showing that biological wiring regularizes the network to spontaneously develop realistic firing selectivities and outperform standard CNNs or LSTMs in tactile discrimination tasks.
Macroscopic Constraints: Brain Atlases and Modularity
We do not always need single-synapse resolution to benefit from anatomical priors. At a macroscopic level, neuroimaging tools like fMRI and PET reveal that the brain is organized into distinct functional modules and regional hierarchies (such as the Default Mode Network). Standard Graph Neural Networks (GNNs) often treat anatomical regions as generic, homogeneous nodes, ignoring these hard-wired organizational principles.
By structuring a network's topology around established anatomical atlases, models can process inter-regional interactions through validated biological pathways. For example, Qianqian Wang et al. introduced the Leveraging Brain Modularity Prior for Interpretable Representation Learning of fMRI framework, which forces GNNs to respect the brain's natural modular boundaries. This structural constraint prevents the model from overfitting to imaging noise, significantly improving diagnostic accuracy for neurodevelopmental disorders. In a similar vein, Jiaming Guo et al. utilized hierarchical graph convolutions defined by anatomical atlases in Predicting Alzheimer's Disease by Hierarchical Graph Convolution from Positron Emission Tomography Imaging, allowing their model to be trained highly efficiently from scratch on small medical datasets without needing large-scale natural-image pre-training.
This concept extends naturally to multi-omic and gene-interaction networks. Zhihao Zhang et al. developed DEG-BRIN-GCN: interpretable graph convolutional framework with differentially expressed genes brain region interaction network prior for AD diagnosis, which constrains message-passing to known inter-regional gene expression interactions. By mapping functional connectivity directly from empirical brain data, researchers like Takuya Ito et al. have demonstrated that localized cognitive representations emerge mechanistically from information flow across these biologically defined networks (Discovering the Computational Relevance of Brain Network Organization; Constructing neural network models from brain data reveals representational transformations linked to adaptive behavior).
Domain-Specific and Bio-Inspired Structural Motifs
Anatomical inductive biases are not limited to modeling the brain itself; they serve as powerful architectural constraints across diverse domains like computer vision, medical report generation, and molecular biology.
In computer vision, Hosein Hasani et al. introduced a hard-coded center-surround connectivity motif inspired by visual receptive fields in Surround Modulation: A Bio-inspired Connectivity Structure for Convolutional Neural Networks, proving that fixed biological motifs can outperform learned parameters for spatial processing tasks. For facial analysis, Haoyu Chen et al. designed an architecture that decomposes and reconstructs facial movements based strictly on human facial anatomy in Prior Information Based Decomposition and Reconstruction Learning for Micro-Expression Recognition, achieving state-of-the-art micro-expression recognition while providing clear, anatomically grounded explanations for its classifications.
In medical and biological applications, constraining graph structures to follow physical branching or molecular pathways ensures that message passing aligns with underlying domain constraints. Tamara T. Mueller et al. highlighted this in A Survey on Graph Construction for Geometric Deep Learning in Medicine: Methods and Recommendations, noting that aligning graph edges with vascular branching or skeletal joints enforces physical plausibility. Similarly, Junzhong Ji et al. used a pathological graph constrained by the physical proximity of brain tissues to align visual features with medical text in Granularity Matters: Pathological Graph-driven Cross-modal Alignment for Brain CT Report Generation.
At the molecular scale, frameworks like those reviewed by Khoa Tran et al. (Deep learning in cancer diagnosis, prognosis and treatment selection), Heming Zhang et al. (MosGraphFlow: a novel integrative graph AI model mining signaling targets from multi-omic data), and Guadalupe Gonzalez et al. (Combinatorial prediction of therapeutic perturbations using causally inspired neural networks) initialize graph topologies using validated protein-protein interaction and gene regulatory databases. This ensures that the model's internal representations reflect true causal and functional biological pathways rather than statistical artifacts.
Why It Works: The Power of Structural Commitments
Why do these rigid architectural constraints consistently match or outperform highly flexible, unconstrained models?
- Representational Alignment: As Emanuele Luconi et al. demonstrated in Anatomically inspired digital twins capture hierarchical object representations in visual cortex, mirroring the physical hierarchy of the visual cortex allows models to build robust, invariant object representations that standard CNNs fail to achieve without overfitting to specific training stimuli.
- Task Generalization over Scaling: Ida Momennejad argues in A Rubric for Human-like Agents and NeuroAI that incorporating specific structural commitments—such as prefrontal cortex-inspired functional specialization—enables agents to generalize across tasks far more effectively than simply scaling model size.
- Isolating Biological Principles: By testing which architectural constraints (like recurrence or lateral connections) best predict empirical brain activity, as explored by Alessandro T. Gifford et al. (The Algonauts Project 2023 Challenge: How the Human Brain Makes Sense of Natural Scenes) and Greta Tuckute et al. (How to optimize neuroscience data utilization and experiment design for advancing brain models of visual and linguistic cognition?), researchers can isolate the precise inductive biases that make biological vision and cognition so robust and sample-efficient.
Ultimately, embedding anatomical inductive biases shifts the burden of discovery from the optimization algorithm to the architecture itself. By treating physical and biological structure as a fixed prior—a concept tracing back to early connectionist models like Daniel J. Rosen et al. (A Connectionist Model of the Owl's Sound Localization System)—we can design artificial systems that inherit the efficiency, robustness, and functional elegance of biological evolution.
Key Foundational Work
A Connectionist Model of the Owl's Sound Localization System (1993) — contributed 2 concept elements to this direction.
Disentangling 3D Pose in A Dendritic CNN for Unconstrained 2D Face Alignment (2018) — contributed 1 concept elements to this direction.