Partial activation of multiple pathways in neural networks
A device, system, and method for approximating a neural network comprising N synapses or filters. The neural network may be partially-activated by iteratively executing a plurality of M partial pathways of the neural network to generate M partial outputs, wherein the M partial pathways respectively comprise M different continuous sequences of synapses or filters linking an input layer to an output layer. The M partial pathways may cumulatively span only a subset of the N synapses or filters such that a significant number of the remaining the N synapses or filters are not computed. The M partial outputs of the M partial pathways may be aggregated to generate an aggregated output approximating an output generated by fully-activating the neural network by executing a single instance of all N synapses or filters of the neural network. Training or prediction of the neural network may be performed based on the aggregated output.
1 . A method for approximating a neural network, the method comprising:
partially-activating a neural network comprising N synapses or filters by iteratively executing a plurality of M partial pathways of the neural network to generate M partial outputs, wherein each of the M partial pathways propagates through X M respective synapses or filters, to compute a total of Σ i M X M synapses or filters in each run of the partially-activated neural network that is less than a total number N of all synapses or filters executed in each run of the fully-activated neural network; and
aggregating the M partial outputs of the plurality of M partial pathways to generate an aggregated output approximating an output generated by the fully-activated neural network that executes a single instance of all N synapses or filters of the neural network.
2 . The method of claim 1 comprising performing training or prediction of the neural network based on the aggregated output.
3 . The method of claim 1 comprising:
storing, in permanent memory, a random number generator, which when executed, randomly generates the neural network to be a random neural network stored in temporary memory; and
approximating an output of a target neural network that is not stored by iteratively executing the plurality of M partial pathways of the random neural network that approximate a subset of synapses or filters of the non-stored target neural network.
4 . The method of claim 1 comprising approximating an ensemble of a plurality of neural networks by dividing the M partial pathways among the plurality of neural networks.
5 . The method of claim 1 , wherein the plurality of M partial pathways are arranged in a hierarchy of N levels, wherein the neural network is partially-activated in N multiple phases for the N respective levels, wherein the partial outputs of executing partial pathways in a stage for a relatively higher level is used to determine which partial pathways to execute in a subsequent stage for a relatively lower level.
6 . The method of claim 1 comprising selecting the plurality of M partial pathways in the neural network by a technique selected from the group consisting of: randomly, semi-randomly, a hybrid or both randomly and semi-randomly, based on individual characteristics of each input into the neural network, during a training phase by learning optimal partial pathways to execute or optimal execution probabilities of partial pathways, based on evolutionary constraints that select partial pathways by competing different pathways against each other using a genetic algorithm, based on heuristic constraints that select optimal partial pathways based on tested accuracy of partial outputs of partial pathways, based on heuristic constraints that select optimal partial pathways that maximize output signals for a specific final result, and based on probabilities of activating synapses or filters along partial pathways proportional to the synapse or filter distance, columnar patterns of randomness, or strengths of the various synapse or filters weights.
7 . The method of claim 1 comprising tuning the number M of partial pathways and the number of synapses or filters X M per pathway such that the aggregated output of the partially-activated neural network best approximates the output of the fully-activated neural network.
8 . The method of claim 1 comprising:
partially-activating the neural network only in training mode and fully-activating the neural network only in prediction mode,
partially-activating the neural network only in prediction mode and fully-activating the neural network only in training mode, or
partially-activating the neural network in both training mode and prediction mode.
9 . The method of claim 1 comprising performing a combination of partially-activating and fully-activating the neural network, wherein the neural network is partially or fully activated for different layers, filters, channels, iterations, or modes of operation, of the neural network.
10 . The method of claim 1 comprising performing training or prediction of the neural network based on the aggregated output.
11 . A system for approximating a neural network, the system comprising:
one or more processors configured to:
partially-activate a neural network comprising N synapses or filters by iteratively executing a plurality of M partial pathways of the neural network to generate M partial outputs, wherein each of the M partial pathways propagates through X M respective synapses or filters, to compute a total of Σ i M X M synapses or filters in each run of the partially-activated neural network that is less than a total number N of all synapses or filters executed in each run of the fully-activated neural network, and
aggregate the M partial outputs of the plurality of M partial pathways to generate an aggregated output approximating an output generated by the fully-activated neural network that executes a single instance of all N synapses or filters of the neural network.
12 . The system of claim 11 comprising:
one or more permanent memories; and
one or more temporary memories,
wherein the one or more permanent memories store a random number generator, which when executed by the one or more processors randomly generates the neural network to be a random neural network that is stored in the one or more temporary memories, and
wherein the one or more processors are configured to approximate an output of a target neural network that is not stored in the permanent and temporary memories by iteratively executing the plurality of M partial pathways of the random neural network that approximate a subset of synapses or filters of the non-stored target neural network.
13 . The system of claim 11 , wherein the one or more processors are configured to tune the number M of partial pathways and the number of synapses or filters X M per pathway such that the aggregated output of the partially-activated neural network best approximates the output of the fully-activated neural network.
14 . The system of claim 11 , wherein the one or more processors are configured to perform training or prediction of the neural network based on the aggregated output.
15 . The system of claim 11 , wherein the one or more processors are configured to approximate an ensemble of a plurality of neural networks by dividing the M partial pathways among the plurality of neural networks.
16 . The system of claim 11 , wherein the plurality of M partial pathways are arranged in a hierarchy of N levels, wherein the neural network is partially-activated in N multiple phases for the N respective levels, wherein the partial outputs of executing partial pathways in a stage for a relatively higher level is used to determine which partial pathways to execute in a subsequent stage for a relatively lower level.
17 . The system of claim 11 , wherein the one or more processors are configured to select the plurality of M partial pathways in the neural network by a technique selected from the group consisting of: randomly, semi-randomly, a hybrid or both randomly and semi-randomly, based on individual characteristics of each input into the neural network, during a training phase by learning optimal partial pathways to execute or optimal execution probabilities of partial pathways, based on evolutionary constraints that select partial pathways by competing different pathways against each other using a genetic algorithm, based on heuristic constraints that select optimal partial pathways based on tested accuracy of partial outputs of partial pathways, based on heuristic constraints that select optimal partial pathways that maximize output signals for a specific final result, and based on probabilities of activating synapses or filters along partial pathways proportional to the synapse or filter distance, columnar patterns of randomness, or strengths of the various synapse or filters weights.
18 . The system of claim 11 , wherein the one or more processors are configured to:
partially-activate the neural network only in training mode and fully-activate the neural network only in prediction mode,
partially-activate the neural network only in prediction mode and fully-activate the neural network only in training mode, or
partially-activate the neural network in both training mode and prediction mode.
19 . The system of claim 11 , wherein the one or more processors are configured to perform a combination of partially-activating and fully-activating the neural network, wherein the neural network is partially or fully activated for different layers, filters, channels, iterations, or modes of operation, of the neural network.
20 . A non-transitory computer readable medium storing instructions, which when executed by a processor, cause the processor to:
partially-activate a neural network comprising N synapses or filters by iteratively executing a plurality of M partial pathways of the neural network to generate M partial outputs, wherein each of the M partial pathways propagates through X M respective synapses or filters, to compute a total of Σ i M X M synapses or filters in each run of the partially-activated neural network that is less than a total number N of all synapses or filters executed in each run of the fully-activated neural network; and
aggregate the M partial outputs of the plurality of M partial pathways to generate an aggregated output approximating an output generated by the fully-activated neural network that executes a single instance of all N synapses or filters of the neural network.