神经生物学研究【20260008】

C++ Technical Documentation: Hodgkin-Huxley Neural Network for Low‑Pass Filtering

Project: HHMLP – Neuromorphic Low‑Pass Filter Using HH Neurons + Linear Readout
Language: C++17
Framework: LibTorch (PyTorch C++ API)
Platform: Ubuntu 24.04 / Fedora, CPU/GPU

1. Introduction

This document describes a complete C++ implementation of a neuromorphic low‑pass filter based on a recurrent network of Hodgkin‑Huxley (HH) neurons. The model learns to map a noisy square wave to its ideal low‑pass filtered version, demonstrating the applicability of biologically plausible spiking neural networks to signal processing tasks.

Key features:

• Fully differentiable Hodgkin‑Huxley neuron model
• Recurrent connections among 20 neurons (trainable weights)
• Single linear readout layer with sigmoid activation
• End‑to‑end training using LibTorch’s automatic differentiation
• CPU and GPU support

2. System Architecture

The system consists of four main components:

2.1 Hodgkin‑Huxley Neuron

State variables: membrane potentialV VV, sodium activationm mm, sodium inactivationh hh, potassium activationn nn. Dynamics:

C m d V d t = I ext − g Na m 3 h ( V − E Na ) − g K n 4 ( V − E K ) − g L ( V − E L ) C_m \frac{dV}{dt} = I_{\text{ext}} - g_{\text{Na}} m^3 h (V-E_{\text{Na}}) - g_{\text{K}} n^4 (V-E_{\text{K}}) - g_{\text{L}} (V-E_{\text{L}})Cm​dtdV​=Iext​−gNa​m3h(V−ENa​)−gK​n4(V−EK​)−gL​(V−EL​)

Integration: forward Euler with (dt = 0.01) ms. All operations are performed ontorch::Tensor, allowing batching and automatic differentiation.

2.2 Recurrent HH Layer

• Number of neurons: 20
• Recurrent weight matrix:W rec ∈ R 20 × 20 W_{\text{rec}} \in \mathbb{R}^{20\times 20}Wrec​∈R20×20(trainable)
• At each time step, input current for neuroni ii:
  I ext , i ( t ) = I external ( t ) + ∑ j W rec , j , i ⋅ s j ( t − 1 ) I_{\text{ext},i}(t) = I_{\text{external}}(t) + \sum_j W_{\text{rec},j,i} \cdot s_j(t-1)Iext,i​(t)=Iexternal​(t)+j∑​Wrec,j,i​⋅sj​(t−1)
  wheres j ( t − 1 ) s_j(t-1)sj​(t−1)is the spike (0/1) from neuronj jjat previous step (threshold crossing detection).
• Output: membrane potential sequence(batch, T, 20)

2.3 Readout Layer

• Linear layer mapping 20‑dim membrane potential to a single scalar.
• Sigmoid activation to ensure output in ([0,1]) (target normalized range).
• Final output shape:(batch, T).

3. Core Implementation Details

3.1 HH Neuron (HHNeuronImpl)

classHHNeuronImpl:publictorch::nn::Module{public:HHNeuronImpl(doubledt=0.01);torch::Tensorforward(torch::Tensor I_ext,intstep=1);voidreset_state(int64_tbatch_size=1);private:doubledt,Cm,gNa,gK,gL,ENa,EK,EL,V_rest;torch::Tensor V,m,h,n;// gate rate functionstorch::Tensoralpha_m(torch::Tensor V);torch::Tensorbeta_m(torch::Tensor V);// ... similar for h, n};TORCH_MODULE(HHNeuron);

Theforwardmethod performs Euler integration forsteptime steps and updates internal state.

3.2 Recurrent HH Layer (RecurrentHHLayerImpl)

classRecurrentHHLayerImpl:publictorch::nn::Module{public:RecurrentHHLayerImpl(intn_neurons,doubledt=0.01);torch::Tensorforward(torch::Tensor input_current);private:intn_neurons;std::vector<HHNeuron>neurons;torch::Tensor W_rec;// trainable recurrent weights};TORCH_MODULE(RecurrentHHLayer);

• input_currentshape:(batch, T, n_neurons)
• Internal loop over time steps:
  • Compute recurrent current using previous spikes
  • Update each neuron one step
  • Detect spikes (V > 0) for next step
• Returns membrane potentials for all time steps.

3.3 Full Model (HHMLPFilterImpl)

classHHMLPFilterImpl:publictorch::nn::Module{public:HHMLPFilterImpl(intn_neurons=20,doubledt=0.01);torch::Tensorforward(torch::Tensor x);private:RecurrentHHLayer hh_layer{nullptr};torch::nn::Linear fc{nullptr};};TORCH_MODULE(HHMLPFilter);

Forward pass:

1. Expand input(batch, T)to(batch, T, n_neurons)(all neurons receive same external current)
2. Pass throughhh_layer→ membrane potentials(batch, T, n_neurons)
3. Apply linear layer and sigmoid along the neuron dimension →(batch, T)

3.4 Data Generation

Functiongenerate_datacreates synthetic training data:

• Input: Square wave (1–10 Hz) + Gaussian noise + random impulse noise, clamped to ([-0.8, 0.8]).
• Target: Low‑pass filtered version using a first‑order IIR filter (cutoff 5 Hz).
• Normalization: Input scaled to ([-5, 5]) (suitable for HH neuron current sensitivity). Target remains in ([-0.8, 0.8]) (output sigmoid will be later mapped).

4. Training Pipeline

4.1 Hyperparameters

4.2 Training Loop

for(intepoch=1;epoch<=epochs;++epoch){doubletotal_loss=0.0;for(intbatch=0;batch<n_batches;++batch){autox=inputs.slice(0,start,end);autoy=targets.slice(0,start,end);optimizer.zero_grad();autopred=model->forward(x);autoloss=torch::mse_loss(pred,y);loss.backward();optimizer.step();total_loss+=loss.item<double>()*(end-start);}if(epoch%10==0)std::cout<<"Epoch "<<epoch<<", Loss = "<<total_loss/n_samples<<std::endl;}

4.3 Model Saving

torch::save(model,"lowpass_model.pt");

Model can be reloaded for inference or further training.

5. Compilation and Execution

5.1 Prerequisites

• LibTorch (CPU or GPU version) – Download
• CMake ≥ 3.18
• C++17 compiler (GCC ≥ 9, Clang ≥ 7)

5.2 CMake Configuration (CMakeLists.txt)

cmake_minimum_required(VERSION 3.18) project(LowpassTrain) set(CMAKE_CXX_STANDARD 17) set(CMAKE_PREFIX_PATH "/path/to/libtorch") # e.g., /home/user/libtorch_cpu find_package(Torch REQUIRED) add_executable(train_lowpass train_lowpass.cpp) target_link_libraries(train_lowpass ${TORCH_LIBRARIES})

5.3 Build & Run

g++-std=c++17 -I/home/x/pro/libtorch/include -I/home/x/pro/libtorch/include/torch/csrc/api/include -I/home/x/pro/BioSight/core -L/home/x/pro/libtorch/lib -Wl,-rpath,/home/x/pro/libtorch/lib train_lowpass.cpp../core/HHNeuron.cpp-ltorch-ltorch_cpu-lc10-otrain_lowpass

mkdirbuild&&cdbuild cmake..make./train_lowpass

Expected output:

Generating data... Data shape: inputs [1500, 50], targets [1500, 50] Training... Epoch 10/100, Loss = 0.0234 ... Epoch 100/100, Loss = 0.00612 Model saved to lowpass_model.pt Test sample MAE: 0.0456

6. Performance Considerations

• Complexity: Each training epoch processes 1500 samples × 50 time steps × 20 neurons × ~10 floating‑point ops per neuron ≈ 15 million operations. On a modern CPU, 100 epochs take ~5 minutes.
• Parallelization: LibTorch automatically uses OpenMP for tensor operations (e.g., matrix multiplication, element‑wise ops). The HH neuron loop itself is sequential; for larger neuron counts (e.g., 200), one may add#pragma omp parallel for– but for 20 neurons, overhead is negligible.
• GPU acceleration: To use GPU, replacetorch::kCPUwithtorch::kCUDAfor tensors and link GPU‑enabled LibTorch. The same code works unchanged thanks to LibTorch’s device abstraction.

7. Extending to Fuzzy Control

The same architecture can be adapted to the fuzzy controller task (Mamdani approximation) by:

• Modifying the data generation to produce(e, edot)inputs and normalized control output.
• Changing the forward pass to accept two scalar inputs instead of a time series.
• The recurrent HH layer still provides temporal dynamics (useful for smoothing).

Example adaptation: remove the time dimension, treatn_injectandn_freestages as in the Python version. The readout remains linear + sigmoid.

8. Conclusion

This C++/LibTorch implementation demonstrates that a small recurrent network of Hodgkin‑Huxley neurons with a single linear readout can learn a low‑pass filtering task end‑to‑end. The code is self‑contained, compiles with standard tools, and serves as a foundation for more complex neuromorphic signal processing and control applications.

Repository: https://gitee.com/waterruby/ANNA.git
License: Apache 2.0

Appendix: Complete Code Listing

The full source code (train_lowpass.cpp) is available in the repository undercpp-src/. It includes:

• HHNeuronImpl,RecurrentHHLayerImpl,HHMLPFilterImplclass definitions
• Data generation routine
• Training loop with checkpointing (optional)
• Model saving and evaluation

For the latest version, please refer to the online repository.