NeuroTrailblazers banner

Overview

The NeuroAI thesis holds that understanding biological neural circuits can inspire better artificial intelligence, and that AI tools can accelerate neuroscience. Connectomics sits at the center of this exchange: it provides the structural data that constrains biological models and reveals architectural motifs that may inform AI design. But the bridge between structure and function has load limits. This document covers what connectomics can and cannot contribute to AI, and vice versa.

Instructor script: the bidirectional exchange

From neuroscience to AI

The history of AI is punctuated by ideas borrowed from neuroscience:

Convolutional neural networks ↔ visual cortex: Hubel & Wiesel (1962) discovered that neurons in cat visual cortex respond to oriented edges and are organized retinotopically (nearby visual field positions map to nearby cortical positions). This inspired Fukushima’s Neocognitron (1980) and later LeCun’s convolutional neural networks (1989). CNNs use local receptive fields and weight sharing — directly analogous to the retinotopic, feature-detecting architecture of early visual cortex.

Yamins et al. (2014) later showed that deep CNNs trained on object recognition develop internal representations that quantitatively match neural activity patterns in primate visual cortex — a striking convergence suggesting that similar computational problems produce similar solutions.

Recurrent neural networks ↔ cortical recurrence: The mammalian cortex is massively recurrent — feedback connections are as abundant as feedforward connections. Yet early neural networks were purely feedforward. Recurrent neural networks (RNNs, LSTMs, and now transformers with attention) better capture temporal dynamics and context dependence. Connectomics reveals the specific recurrent architecture: lateral connections within layers, feedback from higher to lower areas, and cell-type-specific recurrence patterns.

Feedback connections ↔ generative models: Rao & Ballard (1999) proposed that the brain implements predictive coding: higher cortical areas generate predictions (via feedback connections) that are compared to incoming sensory data (via feedforward connections), with only prediction errors propagated forward. This idea influenced variational autoencoders and other generative models.

Lateral inhibition ↔ normalization: Retinal and cortical circuits feature lateral inhibition, where activated neurons suppress their neighbors. This implements a form of divisive normalization (Carandini & Heeger 2012) that is now standard in deep learning as batch normalization, layer normalization, and similar techniques.

Sparse connectivity ↔ efficient networks: Biological neural circuits are sparse — each neuron connects to a tiny fraction of potential partners. This has inspired sparse neural networks and pruning methods in AI, which achieve similar performance with far fewer parameters.

From AI to neuroscience

The exchange flows both ways:

Deep learning for segmentation: Modern connectomics depends entirely on deep learning for automated image segmentation. Flood-filling networks (Januszewski et al. 2018), U-Net variants, and transformer-based models are what make large-scale reconstruction feasible.

Graph neural networks for cell-type classification: GNNs can learn to classify neuron types from connectivity patterns, bypassing the need for morphological reconstruction. Train a GNN on neurons with known types; apply it to classify neurons with unknown types based on their connectivity fingerprints.

Large language models for literature and analysis: LLMs are increasingly used for connectomics literature review, hypothesis generation, and even interactive data exploration.

Structure-function relationships

The connectome as a constraint

The connectome specifies which neurons CAN communicate directly. It does not specify:

Analogy: Knowing the circuit board of a computer (which chips are connected to which) constrains what the computer CAN do, but doesn’t tell you what software it’s running. The connectome constrains possible computations but doesn’t determine which computation is active at any moment.

Despite these limitations, structure DOES predict function in important ways:

Teaching point: “The question is not whether structure predicts function — it does, to a degree. The question is how much, and for which aspects of function.”

Connectome-constrained computational models

The approach

Build a biophysically realistic spiking neural network where:

  1. The connectivity (which neurons connect to which) comes from the connectome
  2. The neuron models (leaky integrate-and-fire, Hodgkin-Huxley, etc.) come from electrophysiology
  3. The synapse models (strength, dynamics) are either measured or treated as free parameters

Then test: can this model reproduce observed neural activity or behavior?

Example: Allen Institute visual cortex model

Billeh et al. (2020) built a model of mouse V1 with ~230,000 neurons, using connectivity constraints from experimental data (not full EM connectomics, but statistically matched). The model reproduced orientation selectivity, contrast response functions, and layer-specific activity patterns. It demonstrated that realistic connectivity is a useful constraint for building functional models.

Example: C. elegans whole-brain models

Several groups have built whole-nervous-system models of C. elegans using the complete connectome. These models can reproduce some locomotion patterns (Izquierdo & Beer 2016) and sensory responses, but require tuning of synapse strengths (not available from EM alone) to match behavior.

The synapse strength problem

EM connectomics tells you which neurons are connected and how many synapses they share. It does NOT directly tell you synapse strength (how much current flows per synapse). Synapse count correlates with strength but is not a reliable predictor — a 10-synapse connection is not necessarily 10× stronger than a 1-synapse connection.

Approaches to infer strength:

Graph neural networks on connectomes

Cell-type prediction

Given a partially labeled connectome (some neurons have known types, others don’t), can a GNN predict the unknown types?

Approach: Each neuron is a node with features (in-degree, out-degree, local clustering, morphological features if available). Edges carry synapse count weights. A GNN (e.g., GraphSAGE, GAT) learns to map node features + neighborhood structure to cell-type labels.

Application in Drosophila: Schlegel et al. (2024) used connectivity-based clustering (related to GNN approaches) to help assign ~8,000 cell types across the whole fly brain.

Anomaly detection

GNNs trained on normal connectivity patterns can flag neurons with unusual connection profiles — potential merge errors, rare cell types, or pathological features.

Latent space embeddings

Embedding neurons in a low-dimensional latent space based on connectivity creates a “connectivity landscape” where neurons with similar wiring patterns cluster together. This can reveal organizational principles not visible in the raw adjacency matrix.

Honest boundaries

What connectomics CAN contribute to AI

What connectomics CANNOT contribute

The overclaiming risk

Bargmann & Marder (2013): “From the connectome to brain function.” The central caution: “The same circuit can produce multiple outputs… Bhatt different circuits can produce similar outputs.” Connectomics provides necessary but not sufficient information for understanding brain function.

For NeuroAI specifically: Drawing a specific AI architecture from a connectome motif requires multiple leaps of abstraction. The motif must be genuinely computational (not an artifact of spatial proximity), the computation must be relevant to AI tasks, and the implementation must translate across the biological-to-silicon gap. Each leap introduces uncertainty.

Teaching point: “The best NeuroAI is honest about what the biological data actually constrains and what remains speculative. Overclaiming hurts both fields.”

Worked example: from cortical motif to RNN architecture

Observation: Layer 2/3 of mouse cortex shows 4× enrichment of reciprocal connections between excitatory neurons (Song et al. 2005).

Translation to AI:

  1. Standard RNNs have all-to-all recurrence. What if we enforce sparse, reciprocal-enriched recurrence?
  2. Build two RNN variants: (a) standard all-to-all, (b) reciprocal-enriched sparse (match the 4× enrichment ratio from cortex)
  3. Train both on a temporal pattern recognition task
  4. Compare: does the bio-inspired architecture show advantages in sample efficiency, stability, or generalization?

Caveat: Even if the bio-inspired version performs better on this task, we cannot conclude that reciprocal enrichment evolved “for” this computational advantage. We can only say the architectural motif is compatible with good performance.

Common misconceptions

Misconception Reality Teaching note
“The brain is a neural network” The brain uses neural networks, but also chemistry, anatomy, development, and embodiment Don’t reduce biology to the ML abstraction
“Connectomics will directly improve AI” The transfer is indirect: principles and constraints, not direct blueprints Look for principles, not wiring diagrams to copy
“AI models that match brain activity are correct models” Multiple models can fit the same data; model comparison is essential Matching activity is necessary, not sufficient
“We just need bigger connectomes” Scale helps, but the synapse-strength and dynamics gaps remain regardless of volume Complement connectomics with functional and molecular data

References