Computer-implemented training method, classification method and system and computer-readable recording medium

Patent metadata

Field	Value
Publication number	US-12073322-B2
Application number	US-202117327030-A
Country	US
Kind code	B2
Filing date	May 21, 2021
Priority date	May 22, 2020
Publication date	Aug 27, 2024
Grant date	Aug 27, 2024

Abstract

A computer-implemented method for training a classifier (Φη), including: training a pretext model (ΦΘ) to learn a pretext task, so as to minimize a distance between an output of a source sample via the pretext model (ΦΘ) and an output of a corresponding transformed sample via the pretext model (ΦΘ), the transformed sample being a sample obtained by applying a transformation (T) to the source sample; S 20 ) determining a neighborhood (NXi) of samples (Xi) of a dataset (SD) in the embedding space; S 30 ) training the classifier (Φη) to predict respective estimated probabilities Φηj(Xi), j=1 . . . C, for a sample (Xi) to belong to respective clusters (Cj), by using a second training criterion which tends to: maximize a likelihood for a sample and its neighbors (Xj) of its neighborhood (Nxi) to belong to the same cluster; and force the samples to be distributed over several clusters.

First claim

What is claimed is:

1. A computer-implemented training method for training a classifier (Φη), wherein:
a transformed sample being a sample obtained by applying a transformation (T) to a source sample, where the source sample is a datum of a source dataset (SD);
the training method comprises:
S 10 ) training a pretext model (ΦΘ) to learn a pretext task, based on a source dataset (SD), by using a first training criterion which tends to minimize, across the source samples of the source dataset, a distance between an output of a source sample via the pretext model (ΦΘ) and an output of a corresponding transformed sample via the pretext model (ΦΘ);
S 20 ) for at least one sample among the samples (Xi) of the source dataset (SD), determining a neighborhood (NXi) of the at least one sample;
wherein for the at least one sample, the neighborhood (NXi) of the at least one sample comprises K closest neighbors of the sample, K being an integer, K>=1, the K closest neighbors of the sample being K samples Xj of the dataset having smallest distances between ΦΘ(Xi) and ΦΘ(Xj);
S 30 ) training the classifier Φη to predict respective estimated probabilities Φηj(Xi), j=1 . . . C, for a sample to belong to respective clusters (Cj), by using a second training criterion which:
tends to maximize a likelihood for a sample and a neighbor (Xj) of the sample belonging to the neighborhood (NXi) of the sample to belong to the same cluster; and

tends to force the samples to be distributed over a plurality of clusters;
the second training criterion includes a summation:

Λ
=

-

1


D


⁢

∑

X
∈

D
⁢

⁢
k

⁢

∑

∈

N
⁢
X
⁢
i

⁢
f

<

Φ
⁢

η
⁡

(
X
)

,

Φ
⁢

η
⁡

(
k
)

>

where
f is an increasing continuous function, for instance a logarithm;
<, > is a dot product;
D is a dataset used for training the classifier at step S 30 ; and
|D| is the number of samples in the dataset.

2. The training method of claim 1 , wherein the first training criterion includes, for a considered sample, a term which increases when differences between a prediction ΦΘ(Xi) for the considered sample (Xi) and a prediction ΦΘ(Tj(Xi)) for the corresponding transformed sample (T(Xi)) increases.

3. The training method according to claim 1 , wherein at step S 20 , for at least a set of two or more similar samples, for which it has been determined that the similar samples should belong to the same cluster, a neighborhood is determined for each of the similar samples, the neighborhood comprising at least the other one(s) of the similar samples.

4. The training method according to claim 1 , wherein at step S 30 , the second training criterion is configured, for at least a set of two or more similar samples of the dataset, for which it has been determined that the similar samples should belong to the same cluster, to tend to maximize a likelihood for the similar samples of the considered set of similar samples to belong to the same cluster.

5. The training method of claim 1 , further comprising executing at least one time a fine-tuning step S 40 :
S 40 ) training the classifier (Φη) based on a third training criterion (Λ), the third training criterion being configured to maximize, for each considered sample (Xi) among high-confidence samples (X hc i) whose highest probability (Φη max (Xi)) is above a predetermined threshold (Thr), a probability (Φη j )(Xi)) for the considered sample to belong to the cluster (Cj) indicated by a maximum coordinate (Φη max (Xi)) of the prediction (Φη max (Xi)) for the considered sample (Xi).

6. The training method of claim 5 , wherein an execution of fine-tuning step(s) (S 40 ) is stopped when it is determined that the number of high-confidence samples does not increase anymore.

7. The training method according to claim 5 , wherein at step S 40 , the third training criterion is configured, for at least a set of two or more similar samples of the dataset, for which it has been determined that the similar samples should belong to the same cluster, to tend to maximize a likelihood for the similar samples of the considered set of similar samples to belong to the same cluster.

8. The training method according to claim 5 , wherein
a strong augmentation (SA(Xi)) of a high confidence sample (X HC i) is added to the dataset used for a step S 40 ; and
the third training criterion is configured to maximize a likelihood for the strong augmentation (SA(Xi)) to belong to the same cluster as the high confidence sample (X HC i).

9. The training method according to claim 1 , wherein
the training method comprises, at one or more of steps S 20 , S 30 and S 40 , taking into account prior knowledge that a plurality of samples, called similar samples, form a cluster; whereby,
at step S 20 , the similar samples are to be considered as neighbors when defining the neighborhoods, and/or
at step

Assignees

• Toyota Motor Co Ltd
• Univ Leuven Kath

Inventors

• Abbeloos Wim
• Othmezouri Gabriel
• Van Gansbeke Wouter
• Vandenhende Simon
• Proesmans Marc
• Georgoulis Stamatios
• Van Gool Luc

Classifications

• G06N3/0895

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

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/82

  using neural networks · CPC title

• G06F18/24137

  Distances to cluster centroïds · CPC title