Research map / Probabilistic and scientific modeling / Dynamical systems
A differentiable brain simulator bridging brain simulation and brain-inspired computing
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces BrainPy, a differentiable brain simulator built on JAX/XLA that adds sparse and event-driven operators, JIT connectivity, a synapse-projection abstraction, multi-scale model building, and object-oriented JIT, bridging brain simulation and brain-inspired computing.
Problem
Traditional brain simulators lack differentiability for gradient-based training, while deep learning frameworks and SNN libraries lack the biophysical realism, sparse/event-driven efficiency, and scalability needed for brain dynamics, so no shared programming framework exists for brain simulation an
Approach
Build a differentiable brain simulator on top of JAX/XLA by providing dedicated sparse and event-driven primitive operators, memory-free JIT connectivity operators, an AlignPre/AlignPost synaptic projection abstraction that decouples dynamics from communication, and an object-oriented JIT transformation.
Key results
- Event-driven csrmv operator is reported as two to five orders of magnitude faster than other sparse/dense operators on CPU and GPU.
- JIT connectivity reservoir scaled to 30,000 nodes reaches 94.4% on KTH (vs. 91.3% prior with 16,384 units); 50,000-node reservoir reaches 98.9% on MNIST.
- COBA-LIF network scaled to over 4 million neurons with 80 synapses per neuron on a single GPU; the GIF PFC spiking network reaches near 100% accuracy on the DMS task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use event-driven matrix-vector multiplication that computes only at presynaptic spike events | BrainPy synaptic computation · Synaptic matrix-vector multiplication | speedup: two to five orders of magnitude faster | Helped |
| Use JIT connectivity instead of dense or sparse matrix multiplication | BrainPy connectivity computation · JIT connectivity matrix-vector multiplication | speed: one to two orders of magnitude greater speed | Helped |
| Increase reservoir size to 30,000 nodes using JIT connectivity | BrainPy reservoir model · KTH | testing accuracy: 91.3% -> 94.4% | Helped |
| Increase reservoir size to 50,000 nodes | BrainPy reservoir model · MNIST | accuracy: 98.9% | Helped |
| Use JIT connectivity instead of sparse connectivity | BrainPy reservoir model · Reservoir inference | speed: twice as fast | Helped |
| Use JIT connectivity instead of dense connectivity for reservoirs larger than 10,000 nodes | BrainPy reservoir model · Reservoir inference | speed | Helped |
| Add automatic AlignPost synapse merging (ablation) | BrainPy multi-area spiking network · 30-area spiking neural network | memory usage | Helped |
| Add automatic AlignPost synapse merging (ablation) | BrainPy multi-area spiking network · 30-area spiking neural network | compilation and simulation time | Helped |
| Train the biologically grounded GIF network with back-propagation | BrainPy spiking network · Delayed match-to-sample working memory task | testing accuracy: nearly 100% | Helped |
Limitations
- Speed claims are based on specific micro-benchmarks and networks (COBA-LIF/HH) rather than broad workloads.
- Biologically detailed training demo is limited to a single working memory task.
Experimental setup
Datasets: KTH MNIST delayed match-to-sample working memory task
Benchmarks: KTH MNIST COBA-LIF COBA-HH delayed match-to-sample
Metrics: speed (running time) memory usage classification accuracy training loss compilation time
Baselines: NEURON NEST Brian2 ANNArchy BindsNet snnTorch SpikingJelly Norse
Builds on: JAX XLA Flax
Closest papers by meaning
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.85)
- Brain Harmony: A Multimodal Foundation Model Unifying Morphology and Function into 1D Tokens (NeurIPS 2025 · similarity 0.84)
- Continuous Thought Machines (NeurIPS 2025 · similarity 0.84)
- BrainVista: Modeling Naturalistic Brain Dynamics as Multimodal Next-Token Prediction (NeurIPS 2026 · similarity 0.83)
- POCO: Scalable Neural Forecasting through Population Conditioning (NeurIPS 2025 · similarity 0.83)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.