Authors: Romain Ferrand, Maximilian Baronig, Thomas Limbacher and Robert Legenstein paper link
This paper explores the integration of bottom-up and top-down information in neocortical pyramidal neurons, highlighting its importance for cortical computations. We propose a simple spiking neuron model that captures the properties of top-down modulated activity. To investigate the impact of this model on context-dependent computations, we evaluate recurrently connected networks of such neurons. Our results demonstrate the enhanced capability of spike-based context-dependent computations.
Spiking Neural Networks, Pyramidal Neurons, Apical Modulation, Context-Dependent Computations
To run the simulation program used in this research, you will need to install the following Python package:
- pytorch
- torchmetrics
- tonic
$ python run.py --help
usage: CSNN [-h] [--seed SEED] [--nb_worker NB_WORKER] [--data_path DATA_PATH] [--batch_size BATCH_SIZE] [--dataset {dvsgesture,shd}] [--seq_len SEQ_LEN]
[--task_type {recognition,memory}] [--model_type {clif,add_clif,lif,lsnn}] [--apical_rec {true,t,1,yes,y,false,f,0,no,n}]
[--apical_tau {short,long,distributed}] [--lr LR] [--lr_decay LR_DECAY] [--l2_decay L2_DECAY] [--epoch EPOCH] [--target_spike_prob TARGET_SPIKE_PROB]
[--spike_reg_l2 SPIKE_REG_L2]
options:
-h, --help show this help message and exit
--seed SEED Random seed for PyTorch and Numpy.
--nb_worker NB_WORKER
Number of workers for the pytorch dataloaders.
--data_path DATA_PATH
Path for dataset download and caching location.
--batch_size BATCH_SIZE
Batch size
--dataset {dvsgesture,shd}
Dataset choice.
--seq_len SEQ_LEN Number of item in the sequence, in the paper we evaluate for 1 and 5.
--task_type {recognition,memory}
Select either the memorization of recognition task.
--model_type {clif,add_clif,lif,lsnn}
Model choice.
--apical_rec {true,t,1,yes,y,false,f,0,no,n}
Determine if we want to use apical recurrence for clif models.
--apical_tau {short,long,distributed}
Determine the time scale of the apical membrane time constant short = 20ms, long = 200ms, distributed unif(20, 200)ms
--lr LR Learning step.
--lr_decay LR_DECAY Exponential decay rate of the learning rate.
--l2_decay L2_DECAY Weights decay factor.
--epoch EPOCH Number of epoch.
--target_spike_prob TARGET_SPIKE_PROB
Target spike probability for regularization.
--spike_reg_l2 SPIKE_REG_L2
Target spike probability for regularization.