US 11,810,340 B2
System and method for consensus-based representation and error checking for neural networks
Pradip Bose, Yorktown Heights, NY (US); Alper Buyuktosunoglu, White Plains, NY (US); Schuyler Eldridge, Ossining, NY (US); Karthik V. Swaminathan, Mount Kisco, NY (US); and Swagath Venkataramani, Yonkers, NY (US)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by International Business Machines Corporation, Armonk, NY (US)
Filed on Nov. 29, 2017, as Appl. No. 15/825,660.
Prior Publication US 2019/0164048 A1, May 30, 2019
Int. Cl. G06V 10/82 (2022.01); G06N 3/08 (2023.01); G06N 20/00 (2019.01); G06F 18/214 (2023.01); G06F 18/20 (2023.01); G06N 3/045 (2023.01); G06V 10/774 (2022.01)
CPC G06V 10/82 (2022.01) [G06F 18/214 (2023.01); G06F 18/2148 (2023.01); G06F 18/285 (2023.01); G06N 3/045 (2023.01); G06N 3/08 (2013.01); G06N 20/00 (2019.01); G06V 10/7747 (2022.01)] 16 Claims
OG exemplary drawing
 
1. A system, comprising:
a memory that stores computer executable components;
a processor that executes computer executable components stored in the memory, wherein the computer executable components comprise:
a determination component that:
determines a first output for a first neural network of a set of neural networks based on a first input subset of data sub-sampled from an input set of data, and
iteratively, until an optimally complex neural network is determined, performs a process comprising:
determines a second output for a second neural network of the set of neural networks based on a second input subset of data sub-sampled from the input set of data, wherein the second input subset of data is different from and has a larger size than the first input subset of data and previous second input subsets of data from earlier iterations of the process, wherein the second neural network has greater complexity than the first neural network and previous second neural networks from the earlier iterations of the process, and complexity is a function of at least one of input pixel count or hidden layer sizes,
determines whether the second output has consensus, according to a consensus criterion, with at least one of the first output of the first neural network or any previous second outputs from the previous second neural networks from the earlier iterations of the process, comprising:
 determine, in an error-free state, a first consensus profile across the first input subset of data, the previous second input subset of data, and the second input subset of data,
 determine, in a presence of errors, a second consensus profile across the first input subset of data, the previous second input subset of data, and the second input subset of data,
 determine a delta between the first consensus profile and the second consensus profile, and
 wherein an error is reported if the delta is greater than the threshold or a no-error is reported if the delta is less than the threshold, and
in response to a determination that the second output has consensus, according to the consensus criterion, with at least one of the first output of the first neural network or any previous second outputs from the previous second neural networks from the earlier iterations of the process, determines that the second neural network is the optimally complex neural network.