keras-spiking 0.2.0

Spiking neuron integration for Keras

KerasSpiking

KerasSpiking provides tools for training and running spiking neural networks directly within the Keras framework. The main feature is keras_spiking.SpikingActivation, which can be used to transform any activation function into a spiking equivalent. For example, we can translate a non-spiking model, such as

inp = tf.keras.Input((5,))
dense = tf.keras.layers.Dense(10)(inp)
act = tf.keras.layers.Activation("relu")(dense)
model = tf.keras.Model(inp, act)

into the spiking equivalent:

# add time dimension to inputs
inp = tf.keras.Input((None, 5))
dense = tf.keras.layers.Dense(10)(inp)
# replace Activation with SpikingActivation
act = keras_spiking.SpikingActivation("relu")(dense)
model = tf.keras.Model(inp, act)

Models with SpikingActivation layers can be optimized and evaluated in the same way as any other Keras model. They will automatically take advantage of KerasSpiking’s “spiking aware training”: using the spiking activations on the forward pass and the non-spiking (differentiable) activation function on the backwards pass.

KerasSpiking also includes various tools to assist in the training of spiking models, such as additional regularizers and filtering layers.

If you are interested in building and optimizing spiking neuron models, you may also be interested in NengoDL. See this page for a comparison of the different use cases supported by these two packages.

Documentation

Check out the documentation for

• Installation instructions

• More detailed example introducing the features of KerasSpiking

• API reference

Release history

0.2.0 (February 18, 2021)

Compatible with TensorFlow 2.1.0 - 2.4.0

Added

• Added the keras_spiking.Alpha filter, which provides second-order lowpass filtering for better noise removal for spiking layers. (#4)

• Added keras_spiking.callbacks.DtScheduler, which can be used to update layer dt parameters during training. (#5)

• Added keras_spiking.default.dt, which can be used to set the default dt for all layers that don’t directly specify dt. (#5)

• Added keras_spiking.regularizers.RangedRegularizer, which can be used to apply some other regularizer (e.g. tf.keras.regularizers.L2) with respect to some non-zero target point, or a range of acceptable values. This functionality has also been added to keras_spiking.regularizers.L1L2/L1/L2 (so they can now be applied with respect to a single reference point or a range). (#6)

• Added keras_spiking.regularizers.Percentile which computes a percentile across a number of examples, and regularize that statistic. (#6)

• Added keras_spiking.ModelEnergy to estimate energy usage for Keras Models. (#7)

Changed

• keras_spiking.SpikingActivation and keras_spiking.Lowpass now return sequences by default. This means that these layers will now have outputs that have the same number of timesteps as their inputs. This makes it easier to process create multi-layer spiking networks, where time is preserved throughout the network. The spiking fashion-MNIST example has been updated accordingly. (#3)

• Layers now support multi-dimensional inputs (e.g., output of Conv2D layers). (#5)

Fixed

• KerasSpiking layers’ reset_state now resets to the value of get_initial_state (as documented in the docstring), rather than all zeros. (#12)

• Fixed a bug with keras_spiking.Alpha on TensorFlow 2.1, where a symbolic tensor in the initial state shape could not be converted to a Numpy array. (#16)

0.1.0 (August 14, 2020)

Compatible with TensorFlow 2.1.0 - 2.3.0

Initial release