US 11,816,565 B2
Semantic coherence analysis of deep neural networks
Moussa Doumbouya, Santa Clara, CA (US); Xavier Suau Cuadros, Barcelona (ES); Luca Zappella, Sunnyvale, CA (US); and Nicholas E. Apostoloff, San Jose, CA (US)
Assigned to Apple Inc., Cupertino, CA (US)
Filed by Apple Inc., Cupertino, CA (US)
Filed on Feb. 17, 2020, as Appl. No. 16/792,835.
Claims priority of application No. ES201930916 (ES), filed on Oct. 16, 2019.
Prior Publication US 2021/0117778 A1, Apr. 22, 2021
Int. Cl. G06N 20/10 (2019.01); G06N 3/045 (2023.01); G06N 3/08 (2023.01)
CPC G06N 3/08 (2013.01) [G06N 3/045 (2023.01); G06N 20/10 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A method comprising:
obtaining a neural network having a plurality of layers that was trained to perform a first inference task;
labeling individual samples in an input dataset for the neural network with class labels for a second inference task to assign the individual samples to classes corresponding to the second inference task;
applying the neural network to the input dataset to obtain intermediate feature vectors for the first inference task from an intermediate layer of the plurality of layers;
clustering the intermediate feature vectors into a plurality of clusters;
comparing the clusters of intermediate feature vectors for the first inference task with the classes of samples labeled according to the class labels for the second inference task to determine a coherence score of the intermediate layer, wherein the coherence score indicates a semantic coherence of features produced by the intermediate layer for performing the second inference task; and
outputting the coherence score for the intermediate layer.