Certainty-based classification networks

Patent metadata

Field	Value
Publication number	US-2023289654-A1
Application number	US-202118016914-A
Country	US
Kind code	A1
Filing date	Jul 19, 2021
Priority date	Jul 24, 2020
Publication date	Sep 14, 2023
Grant date	—

Abstract

A certainty-based prediction apparatus and method are provided. A plurality of main classifier (MC) modules each predict an MC predicted class based on input data, and determine an MC certainty. Each MC module processes a pre-trained, machine learning main classifier having at least one expert class and a plurality of non-expert classes. An expert classifier (EC) module associated with each expert class predicts an EC predicted class based on the input data. Each EC module processes a pre-trained, machine learning expert classifier having two classes including an associated expert class and a residual class that includes any non-associated expert classes and the plurality of non-expert classes. A final predicted class decision module determines a final predicted class and a final certainty based on each MC predicted class, each MC certainty and each EC predicted class. The final predicted class and the final certainty are output.

First claim

What is claimed is:

1 . A hardware accelerator, comprising:
a plurality of main classifier (MC) modules, each MC module to process a pre-trained, machine learning main classifier having at least one expert class and a plurality of non-expert classes, each MC module configured to predict an MC predicted class based on input data, determine an MC certainty, and output the MC predicted class and the MC certainty; an expert classifier (EC) module associated with each expert class, each EC module to process a pre-trained, machine learning expert classifier having two classes including an associated expert class and a residual class that includes any non-associated expert classes and the plurality of non-expert classes, each EC module configured to predict an EC predicted class based on the input data, and output the EC predicted class; and a final predicted class decision module, coupled to each MC module and each EC module, configured to receive each MC predicted class, each MC certainty and each EC predicted class, determine a final predicted class and a final certainty based on each MC predicted class, each MC certainty and each EC predicted class, and output the final predicted class and the final certainty, where each MC certainty is a binary value that indicates whether the MC predicted class is certain or uncertain, and the final certainty is a binary value that indicates whether the final predicted class is certain or uncertain.

2 . The hardware accelerator according to claim 1 , where:
each main classifier is an artificial neural network that includes an input layer, one or more hidden layers and an output layer having a plurality of output nodes, each output node generating a probability for an associated class; and each MC certainty is calculated based on an entropy of the probabilities of the associated classes.

3 . The hardware accelerator according to claim 2 , where the entropy is calculated based on a sum of each output node probability times a value approximately equal to a binary logarithm of the output node probability.

4 . The hardware accelerator according to claim 3 , where each MC certainty is certain when the entropy is less than a predetermined threshold, and uncertain when the entropy is equal to or greater than the predetermined threshold.

5 . The hardware accelerator according to claim 4 , where the output node probabilities are between 0 and 1, and the predetermined threshold is determined during training.

6 . The hardware accelerator according to claim 1 , where, when each MC certainty indicates that the MC predicted class is certain and each MC predicted class is the same, the final predicted class is the MC predicted class, and the final certainty indicates that the final predicted class is certain.

7 . The hardware accelerator according to claim 6 , where, when each MC certainty indicates that the MC predicted class is certain, at least one MC predicted class is different, at least one MC predicted class is an expert class and at least one EC predicted class is the expert class, the final predicted class is the EC predicted class, and the final certainty indicates that the final predicted class is certain.

8 . The hardware accelerator according to claim 7 , where, when at least one MC certainty indicates that the MC predicted class is uncertain, at least one MC predicted class is an expert class and at least one EC predicted class is the expert class, the final predicted class is the EC predicted class, and the final certainty indicates that the final predicted class is certain.

9 . The hardware accelerator according to claim 1 , where, when each MC certainty indicates that the MC predicted class is certain and each MC predicted class is the same, each EC module does not predict and output the EC predicted class.

10 . A method, comprising:
predicting, by a plurality of main classifier (MC) modules, a plurality of MC predicted classes based on input data, each MC module processes a pre-trained, machine learning main classifier having at least one expert class and a plurality of non-expert classes; determining, by each MC module, an MC certainty; predicting, by an expert classifier (EC) module associated with each expert class, an EC predicted class based on the input data, each EC module processes a pre-trained, machine learning expert classifier having two classes including an associated expert class and a residual class that includes any non-associated expert classes and the plurality of non-expert classes; determining, by a final predicted class decision module, a final predicted class and a final certainty based on each MC predicted class, each MC certainty and each EC predicted class; and outputting, by the final predicted class decision module, the final predicted class and the final certainty, where each MC certainty is a binary value that indicates whether the MC predicted class is certain or uncertain, and the final certainty is a binary value that indicates whether the final predicted class is certain or uncertain.

11 . The method according to claim 10 , where:
each main classifier is an artificial neural network that includes an input layer, one or more hidden layers and an output layer having a plurality of output nodes, each output node generating a probability for an associated class; and said determining the MC certainty includes calculating an entropy of the probabilities of the associated classes.

12 . The method according to claim 11 , where said calculating the entropy is based on a sum of each output node probability times a value approximately equal to a binary logarithm of the output node probability.

13 . The method according to claim 12 , where each MC certainty is certain when the entropy is less than a predetermined threshold, and uncertain when the entropy is equal to or greater than the predetermined threshold.

14 . The method according to claim 13 , where the output node probabilities are between 0 and 1, and the predetermined threshold is determined during training.

15 . The method according to claim 10 , where, when each MC certainty indicates that the MC predicted class is certain and each MC predicted class is the same, the final predicted class is the MC predicted class, and the final certainty indicates that the final predicted class is certain.

16 . The method according to claim 15 , where, when each MC certainty indicates that the MC predicted class is certain, at least one MC predicted class is different, at least one MC predicted class is an expert class and at least one EC predicted class is the expert class, the final predicted class is the EC predicted class, and the final certainty indicates that the final predicted class is certain.

17 . The method according to claim 16 , where, when at least one MC certainty indicates that the MC predicted class is uncertain, at least one MC predicted class is an expert class and at least one EC predicted class is the expert class, the final predicted class is the EC predicted class, and the final certainty indicates that the final predicted class is certain.

18 . The method according to claim 10 , where, when each MC certainty indicates that the MC predicted class is certain and each MC predicted class is the same, each EC module does not predict and output the EC predicted class.

Assignees

• Advanced Risc Mach Ltd

Inventors

• VENU Balaji
• Eyole Mbou

Classifications

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/063 Primary

  using electronic means · CPC title

• G06N20/00 Primary

  Machine learning · CPC title