HelioX - A GPU-Native Framework for Simulation and Training of Biophysically Detailed Networks

Sun, 10 May 2026 00:00:00 +0000

Accepted at ICML 2026.

HelioX is a GPU-native framework for simulation and training of biophysically detailed neural networks. It targets the mismatch between irregular dendritic computation and conventional deep learning stacks, and provides an end-to-end pipeline for both forward simulation and gradient-based training.

Highlights:

Multi-stream GPU execution for ionic current calculation, ODE construction, and conductance update
Efficient spike handling path designed for sparse spike events
End-to-end differentiable training for deep biophysical neural architectures
Strong runtime and scalability gains over traditional simulators in large-scale settings

HelioX: A GPU-Native Framework for Simulation and Training of Biophysically Detailed Networks

Sun, 10 May 2026 00:00:00 +0000

Accepted at ICML 2026.

HelioX introduces a GPU-native path to train and simulate biophysically detailed neural networks at practical scale. Instead of forcing biological models into dense ANN-oriented execution paths, it aligns runtime design with dendritic structure and simulation dynamics.

Teaser

Key Contributions

GPU-to-Biophysics Design: Custom-fused CUDA kernels for dendritic hierarchical scheduling and gradient propagation
Unified Simulation + Training Runtime: End-to-end loop from simulation state updates to parameter optimization
Efficiency at Scale: Significant improvements in throughput and memory usage across simulation and training workloads
Large-Scale Feasibility: Demonstrated on deep biophysical MLPs and organism-scale C. elegans fitting tasks on consumer GPUs

Numerical Fidelity and Throughput

Organism-Scale Training

Why It Matters

Biophysically detailed neurons carry rich temporal and spatial computation that point-neuron abstractions cannot directly capture. HelioX lowers the engineering and hardware barriers for using these models in trainable AI systems, making detailed-neuron research more practical for both computational neuroscience and brain-inspired learning.