US 12,282,848 B2
Estimated online hard negative mining via probabilistic selection and scores history consideration
Ran Bakalo, Haifa (IL); and Dana Levanony, Tel Aviv (IL)
Assigned to International Business Machines Corporation, Armonk, NY (US)
Filed by INTERNATIONAL BUSINESS MACHINES CORPORATION, Armonk, NY (US)
Filed on Dec. 17, 2020, as Appl. No. 17/247,584.
Prior Publication US 2022/0198268 A1, Jun. 23, 2022
Int. Cl. G06N 3/08 (2023.01); G06F 9/30 (2018.01); G06F 9/38 (2018.01); G06F 18/214 (2023.01)
CPC G06N 3/08 (2013.01) [G06F 9/30029 (2013.01); G06F 9/3836 (2013.01); G06F 18/2148 (2023.01)] 20 Claims
OG exemplary drawing
 
1. A processor-implemented method for hard negative training, the method comprising:
receiving a training set, wherein the training set comprises training samples;
training a deep neural network (DNN) with the training set;
determining, using the training set as an input to the DNN, information for each of the training samples, wherein the information comprises one or more scores associated with the training samples, the one or more scores associated with the training samples being generated by the DNN in response to receiving the training set as the input and indicating whether the training samples are positive or negative;
generating a training epoch from the training samples based on the one or more scores associated with the training samples, wherein the training epoch comprises a subset of the training samples, wherein how many samples are included in the subset is determined based on a value of a hyper parameter;
updating the information based on using the training epoch as an input for the DNN, wherein the DNN is further trained via the using the training epoch as input to result in a further trained DNN model; and
performing classification by using the further trained DNN model on newly input samples.