GNN-BPU: Biofidelic Graph Neural Architecture

• GNN-BPU is a biofidelic graph neural network architecture that embeds a fixed, biologically-derived recurrent network as its core.
• It maps task-specific data into the sensory subspace using trainable input encoders while preserving the empirical wiring of the Biological Processing Unit.
• Empirical benchmarks show that GNN-BPU achieves competitive performance with MLPs and Transformers on tasks like image classification and chess puzzles, underscoring its parameter efficiency.

The GNN-BPU model refers to a class of graph neural network (GNN) architectures in which the “BPU” component stands for Biological Processing Unit—an empirically derived, fixed-weight recurrent network based directly on the complete connectome of the Drosophila larva brain. This model leverages the dense, heterogeneous, and multi-modal wiring patterns observed in a real animal nervous system, in contrast to conventional GNNs whose architectures are typically synthetic and task-driven. In the GNN-BPU paradigm, task-specific data—such as graph-based inputs, images, or symbolic representations—are mapped through trainable input encoders into the sensory subspace of the BPU, which then propagates activation through its static biological graph for a fixed number of recurrent steps; outputs are read out via trainable projection layers. Only the input and output projections are learnable, while the BPU “core” remains unmodified apart from sign polarity to reflect excitatory/inhibitory synapses. The architecture has demonstrated competitive or superior performance to size-matched multi-layer perceptrons (MLPs) and parameter-matched Transformers in supervised and sequential reasoning benchmarks, motivating further investigation into biofidelic neural computing and the scaling of connectome-derived networks (Yu et al., 15 Jul 2025).

1. Biological Processing Unit: Derivation and Structure

The foundational component of the GNN-BPU model is the Biological Processing Unit (BPU), instantiated from the full Drosophila larva brain connectome as reconstructed from electron microscopy. This network comprises approximately 3,000 nodes (neurons) and 65,000 directed, weighted synaptic connections. The neurons are separated into three functional modules:

• Sensory neurons ($N=430$): receive external signals corresponding to environmental stimuli.
• Output neurons ($N=218$): provide outputs to motor and neuroendocrine circuits.
• Internal neurons ($N=2,304$): form a densely interconnected recurrent core mediating the bulk of computation.

The BPU is used as a non-parametric, fixed-weight reservoir. Its update equations over discrete time $t$ for the population activities $S(t)$ (sensory), $I(t)$ (internal), and $O(t)$ (output) are:

$$\begin{aligned} S(t+1) &= f\left(W_{ss} S(t) + W_{rs} I(t) + W_{os} O(t) + E(t)\right) \ I(t+1) &= f\left(W_{sr} S(t) + W_{rr} I(t) + W_{or} O(t)\right) \ O(t+1) &= f\left(W_{so} S(t) + W_{ro} I(t) + W_{oo} O(t)\right) \end{aligned}$$

Here, $W_{xy}$ are fixed synaptic weight matrices from neuron group $y$ to $x$, $E(t)$ encodes external input injected at $t=0$, and $f(\cdot)$ is a typically nonlinear element-wise function (e.g., ReLU). Synaptic weights’ polarity (excitatory/inhibitory) is derived from the biological data. All weights within the recurrent core are non-trainable, faithfully reflecting the empirical connectivity.

2. Model Integration and Task Encoders

To flexibly support diverse input modalities, the GNN-BPU model employs trainable projections for both input encoding and output readout:

• Input projection: Data are pre-encoded (e.g., using linear transformations, GNN layers, or CNNs) into a vector with the same dimensionality as the sensory neuron population. For instance, board states in chess are encoded via a GNN with two GINEConv layers and pooling to yield a $256$-dimensional embedding, which is linearly mapped to the $430$-dimensional sensory space; for images, pixel vectors are projected directly.
• Output projection: After a finite number of recurrent BPU steps, the activities of the output neuron group are mapped (via a learnable linear or MLP layer) to task-dependent outputs (class probabilities, move logits, etc.).

Only the projections are updated during supervised training, typically by gradient descent, while the BPU core weights are strictly fixed. This architectural constraint isolates the effect of the biological wiring and preserves the inductive bias conferred by evolutionarily optimized circuit motifs.

3. Performance and Empirical Benchmarks

The GNN-BPU demonstrates competitive performance across wide-ranging tasks:

• Image classification: On MNIST, the unmodified BPU achieves $98\%$ accuracy, surpassing a two-layer MLP (97%) with the same number of parameters. On CIFAR-10, it attains $58\%$ (vs 52% for MLP), indicating that biological connectivity confers a nontrivial inductive advantage.
• Chess puzzle solving (ChessBench): As a GNN-BPU (with $\sim$232,912 trainable parameters), the model achieves $60\%$ move accuracy after training on only $10,000$ games, representing an order of magnitude improvement over Transformer models with comparable or larger parameter counts. A CNN-BPU with $\sim$2M parameters outperforms a similarly sized Transformer, and augmented with a depth-6 minimax search achieves $91.71\%$ accuracy—exceeding the $9$M-parameter Transformer baseline.

A scaling experiment using degree-corrected Stochastic Block Model (DCSBM) expansion of the connectome further increases accuracy on CIFAR-10, preserving biological block structure and yielding monotonic performance gains, a property not generally observed in standard architectures (Yu et al., 15 Jul 2025).

4. Comparison to Standard Neural and GNN Architectures

The GNN-BPU contrasts sharply with canonical MLPs, Transformers, and message-passing GNNs:

• Wiring: The BPU’s connectivity arises from measured biological circuits, exhibiting non-random, modular, and heterogeneous motifs absent in synthetic deep networks.
• Parameter efficiency: With only trainable input and output projections, GNN-BPU achieves or surpasses the performance of MLP/Transformer models with several times more parameters.
• Flexibility: The architecture supports both symbolic (e.g., chess graphs) and perceptual (e.g., images) input modalities by varying the encoder while keeping the biofidelic core intact.
• Search enhancement: The architecture can be profitably combined with classical algorithms, such as alpha-beta pruned minimax search, to yield high-accuracy decision-making systems.

5. Connectome Scaling and Block Model Expansion

To evaluate scalability, the model applies the DCSBM to expand the original connectome by up to $5\times$, maintaining the statistical properties of neural subpopulation connectivity:

Connectome Size	CIFAR-10 Accuracy (%)	Baseline MLP (%)
Original BPU	58	52
%%%%24$I(t)$25%%%%	Higher (monotonic)	Lower/inferior

This strategy demonstrates that scaling biofidelic graphs using empirical motif statistics (block sizes, connection densities) preserves and enhances architectural benefits, a phenomenon rarely observed when scaling generic random graph-based networks (Yu et al., 15 Jul 2025).

6. Biofidelic AI: Functional Significance and Implications

The GNN-BPU architecture illustrates several foundational implications:

• Biofidelic wiring as an inductive prior: Evolutionarily optimized neural circuits efficiently support nontrivial AI benchmarks, including challenging perceptual and symbolic tasks.
• Parameter minimization by leveraging structure: Retention of fixed recurrent weights shifts the optimization problem to the “interface” layers, resulting in improved data efficiency and avoidance of overfitting.
• Potential for multimodal and sequential computation: The natural dynamics of biological circuits are particularly suited for processing temporally and modally diverse patterns, aligning with emergent needs in sensor-fusion, planning, or embodied AI.
• Scalability through structured expansions: DCSBM-based connectome scaling offers a principled avenue for bridging small animal circuits and larger, more cognitively capable reservoirs.

A plausible implication is that further leveraging empirical connectomes from increasingly complex organisms and integrating biological motifs (e.g., recurrent feedback, inhibitory control) may furnish architectures with improved generalization, robustness, and learning efficiency across AI domains.

7. Future Directions and Open Problems

The GNN-BPU model opens new directions in neural architecture design rooted in empirical neurobiology:

• Larger connectomes: Pursuing mammalian-scale BPUs as higher-resolution mouse, primate, or synthetic connectomes become available.
• Task-specialized circuit ablations: Mapping the contribution of discrete biological modules (e.g., modality-specific sensory circuits) to different behavioral tasks, as already demonstrated by ablation in the Drosophila BPU for performance analysis.
• Integration with neuromorphic hardware: Implementing fixed-weight, recurrent BPUs on energy-efficient, brain-inspired silicon to capitalize on the biological regularity of the architecture.
• Theoretical analysis: Developing deeper understanding of why and how natural circuits outperform random or hand-designed architectures, particularly in the presence of limited data or multimodal inputs.

These research directions are motivated by the observed advantages in data efficiency, generalization, and strategic reasoning conferred by biologically motivated neural design, as demonstrated empirically in the GNN-BPU framework on classical AI benchmarks (Yu et al., 15 Jul 2025).