Readme

Deep SNNs with various neural coding methods

Spiking Neural Network library based on TensorFlow (V2)

DNN-to-SNN conversion based

Neural Coding

How to run?

Please refer to provided shell script (run.sh).

DNN-to-SNN conversion

1. Run DNN inference w/ fused batchnorm. and to collect activation statistics.

• related configurations

in run.sh
nn_mode=ANN

in ./configs/weight_norm.conf
f_write_stat_train_data=True

2. Run SNN w/ data-based norm.

in run.sh
nn_mode=SNN

in ./configs/weight_norm.conf
f_write_stat_train_data=False
f_fused_bn=True

Set neural coding

in run.sh

input_spike_mode={'REAL',POISSON',WEIGHTED_SPIKE','BURST','TEMPORAL'}
neural_coding={POISSON',WEIGHTED_SPIKE','BURST','TEMPORAL'}

Gradient-based optimization of temporal kernel parameters (T2FSNN, DAC-20)

1. Set nn_mode=SNN, f_train_time_const=True in run.sh

2. Set the number of train epoch epoch_train_time_const and save interval ```time_const_save_interval``

   (Total numver of train data = (the number of train epoch) x (save interval))

3. Run and Train

4. Set f_train_time_const=False, f_load_time_const=True, and time_const_num_trained_data=# of trained data

5. Run and Inference

(if f_train_time_const=True and f_load_time_const=True, load and train kernel parameters)

Early firing (T2FSNN, DAC-20)

• Conventional method time_fire_start == time_fire_duration
• Early firing time_fire_start < time_fire_duration (e.g.,time_fire_start=40,time_fire_duration=80)

Download models

https://www.dropbox.com/sh/6ubl8y3s8jdpj6v/AACf0IIcNhYnPUGDn8ELKKRja?dl=0

(models should be unzipped and located in ./models_ckpt directory)

Publications

Fast and Efficient Information Transmission in Deep Spiking Neural Networks (DAC-19) (https://dl.acm.org/citation.cfm?id=3316781.3317822)

T2FSNN: Deep Spiking Neural Networks with Time-to-first-spike Coding (DAC-20) (https://dl.acm.org/doi/10.5555/3437539.3437564)