Implementing a neural network algorithm on a neurosynaptic substrate based on criteria related to the neurosynaptic substrate

Patent metadata

Field	Value
Publication number	US-10204301-B2
Application number	US-201514662096-A
Country	US
Kind code	B2
Filing date	Mar 18, 2015
Priority date	Mar 18, 2015
Publication date	Feb 12, 2019
Grant date	Feb 12, 2019

Abstract

One embodiment of the invention provides a system for mapping a neural network onto a neurosynaptic substrate. The system comprises a reordering unit for reordering at least one dimension of an adjacency matrix representation of the neural network. The system further comprises a mapping unit for selecting a mapping method suitable for mapping at least one portion of the matrix representation onto the substrate, and mapping the at least one portion of the matrix representation onto the substrate utilizing the mapping method selected. The system further comprises a refinement unit for receiving user input regarding at least one criterion relating to accuracy or resource utilization of the substrate. The system further comprises an evaluating unit for evaluating each mapped portion against each criterion. Each mapped portion that fails to satisfy a criterion may be remapped to allow trades offs between accuracy and resource utilization of the substrate.

First claim

What is claimed is:

1. A system comprising:
at least one processor; and
a non-transitory processor-readable memory device storing instructions that when executed by the at least one processor causes the at least one processor to perform operations including:
reordering one or more dimensions of an adjacency matrix representation of a neural network based on one or more hardware constraints for one or more hardware elements of a neurosynaptic substrate to partition the adjacency matrix representation into multiple portions of a size satisfying the one or more hardware constraints, wherein the neural network comprises a set of neurons with an operating latency requirement that the set of neurons receive inputs at the same time;
for each portion of the multiple portions of the adjacency matrix representation:
selecting, from a plurality of mapping methods, a mapping method for mapping the portion onto the neurosynaptic substrate; and
mapping the portion onto the neurosynaptic substrate utilizing the mapping method selected;

implementing synchronization and uniform latency within the set of neurons to satisfy the operating latency requirement by adding one or more delays to one or more mapped portions representing the set of neurons;
receiving user input comprising one or more user-defined evaluation metrics relating to at least one of accuracy and resource utilization of the neurosynaptic substrate;
evaluating each mapped portion against the one or more user-defined evaluation metrics, wherein each mapped portion that fails to satisfy the one or more user-defined evaluation metrics is re-mapped, and the re-mapping is biased towards one of increased accuracy or decreased resource utilization of the neurosynaptic substrate based on the one or more user-defined evaluation metrics;
composing all mapped portions including all added delays into an output file representing an executable neural network that satisfies the one or more hardware constraints; and
programming the neurosynaptic substrate in accordance with the output file, wherein the programmed neurosynaptic substrate satisfies the operating latency requirement.

2. The system of claim 1 , wherein the neurosynaptic substrate comprises one or more interconnected core circuits, and each core circuit comprises a plurality of electronic neurons, a plurality of electronic axons, and a plurality of synapses interconnecting the neurons to the axons.

3. The system of claim 2 , wherein the one or more user-defined evaluation metrics include at least one hardware constraint related to the neurosynaptic substrate.

4. The system of claim 3 , wherein a hardware constraint is related to one of the following: neuron characteristics of the neurosynaptic substrate, synaptic weights of the neurosynaptic substrate, neuronal fan-in of the neurosynaptic substrate, and neuronal fan-out of the neurosynaptic substrate.

5. The system of claim 1 , wherein the one or more user-defined evaluation metrics include at least one user-specified error criterion relating to accuracy, and each user-specified error criterion may be used to identify and re-map one or more portions of the adjacency matrix representation that exceed a pre-determined error threshold.

6. The system of claim 5 , wherein a user-specified error criterion is based on one of the following: synaptic weights of the neurosynaptic substrate, and dynamic range of the synaptic weights of the neurosynaptic substrate.

7. The system of claim 1 , wherein the one or more user-defined evaluation metrics include at least one user-specified resource utilization criterion relating to resource utilization of the neurosynaptic substrate, and each user-specified resource utilization criterion may be used to identify and re-map one or more portions of the adjacency matrix representation that exceed a pre-determined resource utilization threshold.

8. The system of claim 7 , wherein a user-specified resource utilization criterion is based on one of the following: neurons of the neurosynaptic substrate, axons of the neurosynaptic substrate, synaptic weights of the neurosynaptic substrate, power consumption of the neurosynaptic substrate, surface area of neurosynaptic substrate.

9. The system of claim 1 , wherein the operations further comprise:
marking each mapped portion as mapped.

10. The system of claim 9 , wherein, within the adjacency matrix representation, each entry of each mapped portion is marked as mapped by replacing said entry with zero.

11. The system of claim 1 , wherein implementing synchronization and uniform latency within the set of neurons to satisfy the operating latency requirement further comprises:
for each neuron of the set of neurons, determining a corresponding latency on the neuron based on a mapped portion representing the neuron;
determining which neuron of the set of neurons has the largest latency based on each corresponding latency determined; and
for each neuron of the set of neurons that does not have the largest latency, adjusting a corresponding latency on the neuron to match the largest latency by adding one or more delays to a mapped portion representing the neuron.

12. The system of claim 1 , wherein the dimensions of the adjacency matrix representation are reordered to produce blocks of a pre-determined size that substantially match a specified crossbar size of a synaptic crossbar of the neurosynaptic substrate.

13. The system of claim 1 , wherein the operations further comprise:
reordering the dimensions of the adjacency matrix representation based on a weight-descent blocking method.

14. The system of claim 1 , wherein the operations further comprise:
reordering the dimensions of the adjacency matrix representation based on a pair-wise centroid distance minimization method.

15. A method comprising:
reordering one or more dimensions of an adjacency matrix representation of a neural network based on one or more hardware constraints for one or more hardware elements of a neurosynaptic substrate to partition the adjacency matrix representation into multiple portions of a size satisfying the one or more hardware constraints, wherein the neural network comprises a set of neurons with an operating latency requirement that the set of neurons receive inputs at the same time;
for each portion of the multiple portions of the adjacency matrix representation:
selecting, from a plurality of mapping methods, a mapping method for mapping the portion onto the neurosynaptic substrate; and
mapping portion onto the neurosynaptic substrate utilizing the mapping method selected;
implementing synchronization and uniform latency within the set of neurons to satisfy the operating latency requirement by adding one or more delays to one or more mapped portions representing the set of neurons;
receiving user input comprising one or more user-defined evaluation metrics relating to at least one of accuracy and resource utilization of the neurosynaptic substrate;
evaluating each mapped portion against the one or more user-defined evaluation metrics, wherein each mapped portion that fails to satisfy the one or more user-defined evaluation metrics is re-mapped, and the re-mapping is biased towards one of increased accuracy or decreased resource utilization of the neurosynaptic substrate based on the one or more user-defined evaluation metrics;
composing all mapped portions including all added delays into an output file representing an executable neural network that satisfies the one or more hardware constraints; and
programming the neurosynaptic substrate in accordance with the output file, wherein the programmed neur

Assignees

• IBM

Inventors

• Amir Arnon
• Appuswamy Rathinakumar
• Datta Pallab
• Flickner Myron D
• Merolla Paul A
• Modha Dharmendra S
• Shaw Benjamin G

Classifications

• G06N3/063 Primary

  using electronic means · CPC title

• G06N3/065

  Analogue means · CPC title

• G06N3/049 Primary

  Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06N3/0495

  Quantised networks; Sparse networks; Compressed networks · CPC title