nBP

Implementation of backprop on Intel's neuromorphic research chip Loihi using gated synfire chains.

This is the code used in the following publication:

Renner, A., Sheldon, F., Zlotnik, A., Tao, L., & Sornborger, A. The backpropagation algorithm implemented on spiking neuromorphic hardware. Nature Communications 15, 9691 (2024).
https://doi.org/10.1038/s41467-024-53827-9

Getting started

Please install nxsdk according to Intel's instructions. Please find more information on how to access the Loihi hardware and software here: https://intel-ncl.atlassian.net/wiki/spaces/INRC/pages/1810432001/Access+Intel+Loihi+Hardware (last checked 2024/11)

For downloading and preparing the dataset, we also require the following packages: mnist, scikit-image

The main file to run the code is SFC_backprop/SFC_backprop_main.py.

This code is written to allow rapid prototyping of gated synfire chains without writing code.

Structure

In this section, the configuration of the network is explained. You will not need this if you just want to use the 3-layer network. However, you need it, if you want to change the network topology.

The main file to run the code is SFC_backprop/SFC_backprop_main.py.

The actual network is built in the BackpropNet class that in turn uses the more generic ConnectedGroups. ConnectedGroups constructs a network based on a topology and parameters. A topology is a dictionary that contains connected_pairs and a layer mapping. connected_pairs is a list of tuples of string identifiers for each connection in the whole network. E.g.

connected_pairs = [('x','h','p'),('h','o','p2')]

means that there is a connection of type p between layer x and h and a connection of type p2 between layer h and o. The definition of what connection type p means, is given in the parameters. In the layer mapping, each layer gets assigned a type.

layers = {'x': ('in', 'n_sfc'),
          'h': ('hid', 'n_sfc'),
          'o': ('out', 'n_sfc')}

The first element of the tuple determines the layer type (e.g. how many neurons) and the second element determines the neuron type (e.g. the specific neuron model and parameters).

Here, we are providing 2 pre-defined topologies from the paper: A training network and an inference network.

This is the full topology of the 3-layer inference network:

connected_pairs_inference = [
    ('x', 'h1', 'f'),  # W1 forward
    ('h1', 'o', 'f'),  # W2 forward

    ('g01', 'x', 'g'), # gating x 
    ('g02', 'h1', 'g'), # gating h1
    ('g03', 'o', 'g'), # gating o

    ('input', 'x', '1e'), # spikegen tp provide input to x
    ('in_g', 'g00', '1ge'), # spikegen to initialize gating chain

    ('g00', 'g01', '1ge'), #gating chain
    ('g01', 'g02', '1ge'),
    ('g02', 'g03', '1ge'),
    ('g03', 'g00', '1ge')
]
layers_inference = {
    'input': ('in', 'n_inp'),
    'in_g': ('gat', 'n_inp'),
    'g00': ('gat', 'n_gat'),
    'g01': ('gat', 'n_gat'),
    'g02': ('gat', 'n_gat'),
    'g03': ('gat', 'n_gat'),
    'x': ('in', 'n_sfc'),
    'h1': ('hid', 'n_sfl'),
    'o': ('out', 'n_sfl')
}

It only contains an input (x), hidden (h1) and output (o) layer, a gating neuron for each layer (g00-03) and a spikegen layer to send in the inputs (input) and to kickstart the gating chain (in_g).

The ConnectedGroups class takes the topology and loops over all layers to create CompartmentGroups and over all connected pairs to create synapses.

In addition to the topology dictionary, a parameter dictionary is needed. The parameter dictionary contains general network parameters and the 2 sub-dictionaries neuron_types and connection_types.

neuron_types contains the neuron parameters, such as taus for each neuron type specified in the topology (e.g. 'n_sfc').

connection_types contains the connection parameters, such as the weight for each synapse type specified in the topology. Furthermore, it contains a connectivity type, such as 1:1 or a:a (for all-to-all connectivity) and a synapse class with which the synapse is created. These synapse classes are usually wrappers around the connect funtions of nxsdk.

The ConnectedGroups class is fully generic so that any kind of network can be created with it using the dictionary/string-based specification. The BackpropNet class which inherits from ConnectedGroups is less generic and includes certain specificities of the backprop network.