Excitatory Synapse

class snnpytorch.synapse.exc_synapse.SynapseLayer(num_input_neurons=10, num_output_neurons=100, synapse_time_constant=0.05, conn_prob=0.5, initial_weight_config=None)

Bases: torch.nn.modules.module.Module

Class of a synapse layer.

The iterative model for the synapse is inspired from the following papers:

‘Enabling Deep Spiking Neural Networks with Hybrid Conversion and Spike Timing Dependent Backpropagation’ by Rathi et al. , ICLR 2019, https://openreview.net/forum?id=B1xSperKvH

‘Spatio-Temporal Backpropagation for Training High-Performance Spiking Neural Networks’ by Wu et al., https://www.frontiersin.org/articles/10.3389/fnins.2018.00331/full

In PyTorch, the total synaptic input current to a neuron / synapse state (i) update is performed in the following manner:

\[i^{t} = \lambda i^{t-1} + wI\]

where \(\lambda\) is the synaptic time constant, t is the present timestep/iteration, w is the synaptic weight, I is the input spike.

Parameters

  • num_input_neurons – Number of neurons in input layer

  • num_output_neurons – Number of neurons in output layer

  • synapse_time_constant – Synapse time constant

  • conn_prob – Connection probability from input to output layer

  • initial_weight_config – Weight initialisation method and value.

Default value of initial_weight_config : {“method”: “constant”, “value”: 0.25}

create_connectivity() → None

Create random connectivity between input and output layer neurons.

forward(x: torch.Tensor) → torch.Tensor

Forward pass for this synapse layer.

Parameters

x – Input neuron spikes

Returns

Synapse state / Input current to the output layer

initialize_states(model_device='cuda:0') → None

Initialise synapse layer weights, connectivity, and state.

Parameters

model_device – ‘cpu’ or ‘cuda:0’

update_synapse_states(x: torch.Tensor)

Update the synapse state.

Parameters

x – Input neuron spikes

Returns

Synapse state / Input current to the output layer