Distributed state via cascades of tensor decompositions and neuron activation binding on neuromorphic hardware
View Patent ↗Distributed state via cascades of tensor decompositions and neuron activation binding on neuromorphic hardware is provided. In various embodiments, a kernel is divided into a plurality of subkernels. Each subkernel has less than a predetermined size. The plurality of subkernels are distributed, each to one of a plurality of neurosynaptic processors. By each of the plurality of neurosynaptic processors, one of the subkernels is applied to an input to generate a partial convolution. The partial convolutions from each of the plurality of neurosynaptic processors are combined to determine an activation.
1. A method comprising:
dividing a kernel into a plurality of subkernels, each subkernel having less than a predetermined size;
distributing the plurality of subkernels, each to one of a plurality of neurosynaptic processors;
by each of the plurality of neurosynaptic processors, applying one of the subkernels to an input to generate a partial convolution, wherein generating the partial convolutions comprises computing partial sums and wherein the partial sums are represented in rate code;
combining the partial convolutions from each of the plurality of neurosynaptic processors to determine an activation.
2. The method of claim 1 , wherein combining the partial convolutions comprises summing the partial sums.
3. The method of claim 2 , wherein combining the partial convolutions comprises thresholding.
4. The method of claim 1 , wherein each of the plurality of neurosynaptic processors comprises a plurality of synapses.
5. The method of claim 4 , wherein the synapses are trinary.
6. The method of claim 1 , wherein each of the plurality of subkernels has about the same sparsity.
7. The method of claim 1 , further comprising: clipping the partial convolutions.
8. The method of claim 1 , wherein the input comprises an image.
9. The method of claim 1 , wherein the input comprises the output of a layer of a neural network.
10. A system comprising:
a plurality of neurosynaptic processors comprising a plurality of synapses, each of the plurality of neurosynaptic processors adapted to:
receive one of a plurality of subkernels of a kernel, each subkernel having less than a predetermined size;
apply the one of the plurality of subkernels to an input to generate a partial convolution, wherein generating the partial convolutions comprises computing partial sums and wherein the partial sums are represented in rate code;
a neurosynaptic processor adapted to receive the partial convolutions from each of the plurality of neurosynaptic processors and to determine an activation therefrom.
11. The system of claim 10 , wherein combining the partial convolutions comprises summing the partial sums.
12. The system of claim 11 , wherein combining the partial convolutions comprises thresholding.
13. The system of claim 10 , wherein each of the plurality of neurosynaptic processors comprises a plurality of synapses.
14. The system of claim 13 , wherein the synapses are trinary.
15. The system of claim 10 , wherein each of the plurality of subkernels has about the same sparsity.
16. A computer program product for tensor decomposition and neuron activation binding, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to perform a method comprising:
dividing a kernel into a plurality of subkernels, each subkernel having less than a predetermined size;
distributing the plurality of subkernels, each to one of a plurality of neurosynaptic processors;
by each of the plurality of neurosynaptic processors, applying one of the subkernels to an input to generate a partial convolution, wherein generating the partial convolutions comprises computing partial sums and wherein the partial sums are represented in rate code;
combining the partial convolutions from each of the plurality of neurosynaptic processors to determine an activation.
17. The computer program of claim 16 , wherein combining the partial convolutions comprises summing the partial sums.
18. The computer program of claim 16 , wherein each of the plurality of neurosynaptic processor comprises a plurality of synapses.
19. The computer program of claim 17 , wherein the synapses are trinary.
20. The computer program of claim 16 , wherein each of the plurality of subkernels has about the same sparsity.