Paper deep dive
Structure as Computation: Developmental Generation of Minimal Neural Circuits
Duan Zhou
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 98%
Last extracted: 4/18/2026, 1:49:37 AM
Summary
This paper introduces a generative developmental framework that simulates cortical neurogenesis using gene regulatory rules derived from mouse transcriptomic data. The process produces a minimal neural circuit of 85 mature neurons from 5,000 cells, which exhibits rapid learning capabilities on MNIST and CIFAR-10 datasets, suggesting that developmental rules encode domain-general structural priors for efficient computation.
Entities (5)
Relation Signals (4)
Duan Zhou → authored → Structure as Computation: Developmental Generation of Minimal Neural Circuits
confidence 100% · Structure as Computation: Developmental Generation of Minimal Neural Circuits Duan Zhou
Developmental Neural Network → trainedon → MNIST
confidence 100% · At iteration zero, this minimal circuit performs at chance level on MNIST.
Developmental Neural Network → trainedon → CIFAR-10
confidence 100% · The identical circuit, without any architectural modification or data augmentation, achieves 40.53% on CIFAR-10 after one epoch.
Developmental Neural Network → derivedfrom → GSE211140
confidence 95% · gene regulatory rules derived from mouse single-cell transcriptomic data... accession number GSE211140
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:This work simulates the developmental process of cortical neurogenesis, initiating from a single stem cell and governed by gene regulatory rules derived from mouse single-cell transcriptomic data. The developmental process spontaneously generates a heterogeneous population of 5,000 cells, yet yields only 85 mature neurons - merely 1.7% of the total population. These 85 neurons form a densely interconnected core of 200,400 synapses, corresponding to an average degree of 4,715 per neuron. At iteration zero, this minimal circuit performs at chance level on MNIST. However, after a single epoch of standard training, accuracy surges to over 90% - a gain exceeding 80 percentage points - with typical runs falling in the 89-94% range depending on developmental stochasticity. The identical circuit, without any architectural modification or data augmentation, achieves 40.53% on CIFAR-10 after one epoch. These findings demonstrate that developmental rules sculpt a domain-general topological substrate exceptionally amenable to rapid learning, suggesting that biological developmental processes inherently encode powerful structural priors for efficient computation.
Tags
Links
- Source: https://arxiv.org/abs/2604.15143v1
- Canonical: https://arxiv.org/abs/2604.15143v1
Trouble viewing inline? Open PDF directly →
Full Text
12,865 characters extracted from source content.
Expand or collapse full text
Structure as Computation: Developmental Generation of Minimal Neural Circuits Duan Zhou This work was conducted independently. Code available at: https://github.com/dataangel/developmental-n Abstract This work simulates the developmental process of cortical neurogenesis, initiating from a single stem cell and governed by gene regulatory rules derived from mouse single-cell transcriptomic data. The developmental process spontaneously generates a heterogeneous population of 5,000 cells, yet yields only 85 mature neurons—merely 1.7% of the total population. These 85 neurons form a densely interconnected core of 200,400 synapses, corresponding to an average degree of 4,715 per neuron. At iteration zero, this minimal circuit performs at chance level on MNIST. However, after a single epoch of standard training, accuracy surges to over 90%—a gain exceeding 80 percentage points—with typical runs falling in the 89–94% range depending on developmental stochasticity. The identical circuit, without any architectural modification or data augmentation, achieves 40.53% on CIFAR-10 after one epoch. These findings demonstrate that developmental rules sculpt a domain-general topological substrate exceptionally amenable to rapid learning, suggesting that biological developmental processes inherently encode powerful structural priors for efficient computation. I Introduction Traditional deep neural networks rely on end-to-end gradient-based optimization to simultaneously learn both network structure and synaptic weights. This approach, while remarkably successful, departs fundamentally from biological neural development, where the initial wiring of the brain is established through genetically encoded developmental programs long before sensory experience begins. In this work, we explore an alternative paradigm: rather than training both structure and weights, we generate network topology through a fixed, biologically-grounded developmental simulation. Starting from a single stem cell, we simulate the processes of division, migration, differentiation, and synaptogenesis using gene regulatory rules derived from mouse cortical transcriptomic data. The developmental process is purely generative—the rules are fixed and require no optimization. Our key finding is unexpected: the developmental process produces a heterogeneous population of 5,000 cells, yet only 85 mature neurons emerge—merely 1.7% of the total population. Despite this extreme sparsity, the resulting circuit achieves over 90% accuracy on MNIST after just one epoch of training. Strikingly, the identical circuit—without any architectural modification—achieves 40.53% on CIFAR-10 after one epoch. This rapid learning phenomenon suggests that developmental rules can sculpt topological substrates exceptionally amenable to gradient-based optimization across distinct visual domains. I Method I-A Developmental Data Source The gene regulatory rules governing cortical development are derived from a single-cell RNA sequencing timecourse of mouse embryonic stem cell differentiation and cortical neurogenesis, obtained from the Gene Expression Omnibus (GEO) under accession number GSE211140. The dataset comprises 10 sequential timepoints spanning from pluripotent stem cells to differentiated cortical neurons, capturing the binary activation states of 15 key neurodevelopmental genes. I-B Boolean Rule Inference Given the binary expression matrix ∈0,1G×TX∈\0,1\^G× T for G genes across T=10T=10 timepoints, we infer a Boolean regulatory rule for each target gene. The inference procedure respects two constraints: Temporal Causality: A candidate upstream regulator must exhibit a change in expression prior to or coincident with a change in the target gene. This ensures that inferred regulatory relationships are temporally plausible and not merely correlative. Agreement Maximization: For each target gene, we search over combinations of up to k=2k=2 upstream regulators. For each candidate combination, we compute the agreement score between the Boolean rule’s output and the observed expression trajectory: score(f)=1T∑t=1T[f(ℛ,t)=g,t]score(f)= 1T _t=1^TI [f(X_R,t)=X_g,t ] where f is a candidate Boolean function over the regulator set ℛR, and [⋅]I[·] is the indicator function. The rule achieving the highest agreement exceeding a confidence threshold θ=0.6θ=0.6 is selected. If no rule exceeds the threshold, the gene is assigned a constant default state. I-C Simulated Developmental Process Initialized with a single stem cell, the simulation proceeds for 60 discrete time steps. At each step: Division: Each cell divides with probability dependent on the expression of stemness-associated genes. Migration: Cells undergo a random walk: t+1=t+η⋅ϵp_t+1=p_t+η· ε, with η=0.04η=0.04 and ϵ∼(0,) ε (0,I). Differentiation: Gene states update synchronously according to the inferred Boolean regulatory rules. Maturation: After a developmental threshold, neuronal progenitors mature, activating mature neuronal markers. Synaptogenesis: Synaptic connections form between mature neurons based on gene expression compatibility and spatial proximity (r=0.12r=0.12). The initial synaptic weight is wij=Cij⋅(1−dij/r)w_ij=C_ij·(1-d_ij/r), where CijC_ij is the cosine similarity of gene expression vectors. I-D Network Integration and Training The developmentally generated topology is converted to a fixed-weight recurrent layer ∈ℝ85×85W ^85× 85, normalized to unit row sums. For MNIST, input images (28×2828× 28 grayscale) are flattened and projected via in∈ℝ784×85W_in ^784× 85. For CIFAR-10, the identical circuit is used without architectural modification; the only change is resizing the input projection to in∈ℝ3072×85W_in ^3072× 85 to accommodate the larger 32×32×332× 32× 3 inputs. No convolutional layers, spatial alignment, or data augmentation are employed. The network output is computed as: =softmax(out⋅ReLU(inT+ReLU(inT)T))y=softmax (W_out·ReLU (xW_in^T+ReLU(xW_in^T)W^T ) ) Only inW_in and outW_out are trainable. Training uses cross-entropy loss and the Adam optimizer with learning rate 10−310^-3, batch size 64, for 10 epochs (MNIST) and 100 epochs (CIFAR-10). I Experiments I-A Developmental Outcome The simulation generated 5,000 cells with distribution shown in Table I. Notably, only 85 neurons (1.7%) matured, embedded within a large progenitor pool. These 85 neurons form a densely interconnected core of 200,400 synapses, corresponding to an average degree of 4,715 per neuron. This extreme architecture—a small neuronal assembly with massive recurrent connectivity—emerges spontaneously from the developmental rules. TABLE I: Cellular Composition of Developmentally Generated Network Cell Type Count Proportion Neuronal progenitor 4,046 80.9% Oligodendrocyte progenitor 431 8.6% Stem cell 315 6.3% Undefined 123 2.5% Neuron (mature) 85 1.7% Total 5,000 100% Note: The 85 mature neurons form 200,400 synaptic connections (average degree: 4,715). I-B Rapid Learning on MNIST At iteration zero—with randomly initialized input/output projections and before any gradient updates—the network performs at chance level on MNIST. However, after a single epoch of standard training, accuracy surges to over 90% (92.15% in the reported run; typical runs fall within 89–94%, see Table I). This variability reflects the inherent stochasticity of the developmental process and underscores that the rapid learning phenomenon is a robust emergent property. Standard randomly-initialized networks of comparable size typically require 5–10 epochs to reach 90% accuracy on MNIST. TABLE I: Training Dynamics on MNIST Stage Accuracy Gain Iteration 0 (untrained) ∼ 10% – Epoch 1 92.15% +82% Epoch 5 96.33% +4.18% Epoch 10 (final) 96.85% +0.52% I-C Generalization to Natural Images To assess whether the rapid learning phenomenon is specific to MNIST or reflects a domain-general computational property, we evaluated the identical 85-neuron circuit on CIFAR-10. No architectural modification, no data augmentation, and no domain-specific tuning were employed. The 32×32×3 RGB images were simply flattened into 3072-dimensional vectors and fed directly into the network. The only change was resizing the input projection layer from 784→85784→ 85 to 3072→853072→ 85—a purely mechanical adjustment required by the difference in input dimensionality. At iteration zero, the network performs at chance level (∼ 10%). After a single epoch, accuracy reaches 40.53%—four times the random baseline (Table I). The network continues to improve, reaching 45.12% by epoch 5 and saturating at approximately 50% with extended training. We note that absolute performance is constrained by the minimal circuit size—85 neurons processing 3072-dimensional input without spatial inductive biases. This ceiling is expected and, crucially, serves as a control: it rules out overparameterization as an explanation for the rapid learning phenomenon. The fact that such a tiny circuit achieves 40% in one epoch—and approaches 50% asymptotically—underscores the efficiency of the developmental priors. TABLE I: Training Dynamics on CIFAR-10 (Same 85-Neuron Circuit) Stage Accuracy Gain Iteration 0 ∼ 10% – Epoch 1 40.53% +30% Epoch 2 43.00% +2.47% Epoch 5 45.12% +2.12% Asymptotic ∼ 50% – I-D Ablation: Random Topology A density-matched random topology (200,400 synapses, 85 neurons) achieves only chance-level accuracy at iteration zero and fails to exhibit the rapid learning phenomenon on either dataset. This confirms that the developmental rules—not mere sparsity or recurrence—are causally responsible for the observed effect. IV Discussion IV-A Why Does It Work? The developmental rules produce diverse gene expression profiles, yielding a synaptic matrix with rich spectral properties. This matrix acts as a fixed, high-dimensional projection that provides an exceptionally favorable loss landscape for gradient-based optimization. This aligns with the reservoir computing paradigm, where a fixed recurrent layer with appropriate spectral properties enables rapid linear readout learning. IV-B Domain-General Structural Priors The fact that the identical 85-neuron circuit supports rapid learning on both MNIST and CIFAR-10—despite the radical difference in input statistics—indicates that the developmental priors are domain-general. The topology does not encode task-specific features; rather, it provides a universal inductive bias that accelerates gradient-based learning across distinct visual domains. IV-C The Minimal Circuit Phenomenon The extreme sparsity of mature neurons (85/5,000) mirrors early cortical development, where a small cohort of pioneer neurons establishes the initial functional architecture amid a large progenitor pool. Our results suggest this developmental strategy may serve as an efficient circuit search mechanism: generate a large progenitor reservoir, let a select few mature, and rely on their dense interconnectivity to form a potent learning substrate. IV-D Limitations and Future Work This is a proof-of-concept on MNIST and CIFAR-10. Future work will extend to more complex datasets, investigate activity-dependent plasticity during development, and explore scaling to larger neuronal populations. V Conclusion We presented a generative developmental framework that produces functional neural circuits from a single stem cell using fixed gene regulatory rules. The resulting network—merely 85 neurons with 200,400 synapses—achieves over 90% accuracy on MNIST after a single epoch of training, and 40.53% on CIFAR-10 using the identical circuit without any architectural modification. Our findings demonstrate that developmental rules can sculpt minimal yet exceptionally learnable computational cores, offering a new paradigm for neural network initialization inspired by cortical neurogenesis. The complete source code is publicly available at: https://github.com/dataangel/developmental-n. References [1] Y. LeCun, C. Cortes, and C. J. Burges, “The MNIST database of handwritten digits,” 1998. [Online]. Available: http://yann.lecun.com/exdb/mnist/ [2] NCBI Gene Expression Omnibus, “Single-cell RNA-seq of mouse cortical development,” Accession: GSE211140, 2023. [Online]. Available: https://w.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE211140