Parallelized Hierarchical Connectome: A Spatiotemporal Recurrent Framework for Spiking State-Space Models
A new model called PHCSSM integrates five biological constraints into a fully parallelizable training pipeline, cutting parameter complexity.
Researcher Po-Han Chiang has proposed a novel AI architecture called the Parallelized Hierarchical Connectome (PHC), a framework designed to bridge the gap between efficient machine learning models and biologically plausible neural networks. The work, detailed in a pre-print paper, fundamentally upgrades conventional State-Space Models (SSMs)—known for their fast, parallel sequence processing—by adding a spatial dimension. The PHC framework maps computations to a shared Neuron Layer and a shared Synapse Layer, organizing neurons into hierarchical regions. A key innovation is the Multi-Transmission Loop, which allows signals to propagate across this hierarchical connectome within a single timestep, enabling intra-slice spatial recurrence while maintaining the O(log T) parallelism that makes SSMs so fast.
The framework is instantiated as a model named PHCSSM, which is the first to successfully unify the dynamics of recurrent spiking neural networks (SNNs) with the parallel training efficiency of diagonal SSMs. Crucially, PHCSSM enforces five core biological constraints that are typically intractable for standard AI models, including adaptive leaky integrate-and-fire dynamics, Dale's Law (which separates excitatory and inhibitory neurons), and various forms of synaptic plasticity. Empirical testing on physiological benchmarks from the UEA multivariate time-series archive shows that PHCSSM performs competitively with state-of-the-art SSMs. A major efficiency breakthrough is its parameter reduction: where a standard L-layer stacked SSM architecture has complexity Θ(D²L), PHCSSM reduces this to Θ(D²), offering a more parameter-efficient path to powerful sequence modeling.
This research suggests that incorporating well-founded neuro-physical priors—the 'inductive biases' of biological brains—is a principled route to creating more efficient and capable AI systems. By opening the highly efficient diagonal SSM architecture to spatiotemporal recurrence, the PHC framework paves the way for training fully parallelizable, brain-inspired AI models that were previously too slow or complex to train at scale.
- The PHC framework adds spatial recurrence to fast State-Space Models (SSMs) via a hierarchical connectome and a Multi-Transmission Loop, preserving O(log T) parallelism.
- The resulting PHCSSM model enforces five biological constraints (e.g., Dale's Law, synaptic plasticity) and is the first to unify spiking neural network dynamics with parallel SSM training.
- It reduces parameter complexity from Θ(D²L) to Θ(D²) and achieves competitive performance on physiological time-series benchmarks, proving biological priors aid efficiency.
Why It Matters
It enables efficient training of complex, brain-inspired AI models, merging biological plausibility with the computational speed needed for practical applications.