CLIF: Complementary Leaky Integrate-and-Fire Neuron for Spiking Neural Networks (ICML 2024)

Commonly used LIF neurons suffer from the vanishing gradient problem in the temporal dimension. This work propose the Complementary Leaky Integrate-and-Fire (CLIF) Neuron without (hyper)parameters. CLIF creates extra paths to facilitate the backpropagation in computing temporal gradient while keeping binary output.

It is feasible to train a model using CLIF and subsequently deploy it or inference with LIF. We take pre-trained models of CLIF and LIF (Resnet18 with T=6) to perform inference on the CIFAR10 and CIFAR100 tasks. To compensate for CLIF’s enhanced reset process, we employ a hard reset with a bias as a hyperparameter. As can be seen in Table, this approach leads to an inference accuracy that surpasses that of a model directly trained with LIF.

Table. Directly convert the pre-trained CLIF/LIF model to an LIF neuron for inference.

CIFAR10	Soft Reset	Hard Reset
Reset Value	None	0	-0.02	-0.04	-0.06	-0.08	-0.1
CLIF pretrained (95.41%)	92.95%	93.41%	94.18%	94.54%	95.08%	94.84%	94.72%
LIF pretrained (94.51%)	94.51%	84.05%	76.68%	66.08%	52.16%	38.00%	27.04%
CIFAR100	Soft Reset	Hard Reset
Reset Value	None	0	-0.02	-0.04	-0.06	-0.08	-0.1
CLIF pretrained (78.36%)	68.72%	73.04%	74.64%	76.63%	76.55%	77.00%	76.54%
LIF pretrained (76.23%)	76.23%	47.74%	37.04%	27.56%	19.83%	13.22%	8.77%

Dependencies

• Python 3
• PyTorch, torchvision
• spikingjelly 0.0.0.0.12
• Python packages: pip install tqdm progress torchtoolbox thop

Training

We use single GTX4090 GPU for running all the experiments. Multi-GPU training is not supported in the current codes.

Setup

CIFAR-10, CIFAR-100, Tiny-Imagenet, DVS-CIFAR10, and DVS-Gesture:

# CIFAR-10
python train_BPTT.py -data_dir ./data_dir -dataset cifar10 -model spiking_resnet18 -T_max 200 -epochs 200 -weight_decay 5e-5 -neuron CLIF

# CIFAR-100
python train_BPTT.py -data_dir ./data_dir -dataset cifar100 -model spiking_resnet18 -T_max 200 -epochs 200 -neuron CLIF

# Tiny-Imagenet
python train_BPTT.py -data_dir ./data_dir -dataset tiny_imagenet -model spiking_vgg13_bn -neuron CLIF
   
# DVS-CIFAR10
python train_BPTT.py -data_dir ./data_dir -dataset DVSCIFAR10 -T 10 -drop_rate 0.3 -model spiking_vgg11_bn -lr=0.05  -mse_n_reg -neuron CLIF

# DVS-Gesture
python train_BPTT.py -data_dir ./data_dir -dataset dvsgesture -model spiking_vgg11_bn -T 20 -b 16 -drop_rate 0.4  -neuron CLIF

If you change the neuron, you can directly switch to LIF or PLIF by modifying the hyperparameters after -neuron.

For example to setup LIF neuron for CIFAR-10 task:

# LIF neuron for CIFAR-10
python train_BPTT.py -data_dir ./data_dir -dataset cifar10 -model spiking_resnet18 -amp -T_max 200 -epochs 200 -weight_decay 5e-5 -neuron LIF

Inference

The inference setup could refer file: run_inference_script

Acknowlegement

The code for data preprocessing is based on the spikingjelly repo. This repository is mainly modified from this repo.