Method of converting neural network and recognition apparatus using the same
AssigneesSamsung Electronics Co., Ltd., Universitaet Zuerich
InventorsBodo RUCKAUER, Iulia-Alexandra LUNGU, Yuhuang HU, Michael PFEIFFER
Life of this patent
Estimating the term…
Abstract
A neural network conversion method and a recognition apparatus that implements the method are provided. A method of converting an analog neural network (ANN) to a spiking neural network (SNN) normalizes first parameters of a trained ANN based on a reference activation that is set to be proximate to a maximum activation of artificial neurons included in the ANN, and determines second parameters of an SNN based on the normalized first parameters.
Claims22
A method for converting an analog neural network (ANN) to a spiking neural network (SNN), the method comprising:
- normalizing first parameters of a trained ANN based on a reference activation which is determined based on a predetermined proximity with respect to a maximum activation of artificial neurons included in the ANN;
- and
- determining second parameters of an SNN based on the normalized first parameters.
A recognition apparatus, comprising:
- a processor configured to recognize input data by using a spiking neural network (SNN);
- and
- a memory including at least one instruction that is executable by the processor,
- wherein an output layer included in the SNN is configured to determine an artificial neuron to be used for outputting a spike based on current membrane potentials of artificial neurons included in the output layer obtained at a predetermined timing, and to output the spike via the determined artificial neuron.
A method for converting an analog neural network (ANN) to a spiking neural network (SNN), the method comprising:
- obtaining, for a plurality of neurons that belong to the ANN, a corresponding plurality of weights assigned to synapses of the plurality of neurons and a corresponding plurality of biases applied to the plurality of neurons;
- normalizing the obtained plurality of weights and the obtained plurality of biases based on a predetermined activation based on artificial neurons included in the ANN;
- and
- determining, based on the normalized weights and the normalized biases, a plurality of firing rates to be used for constructing the SNN.
Dependencies are inferred from the claim text.