# US 2020/0401900 A1 · Apparatus for performing class incremental learning and method of operating the apparatus

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0401900 A1
- Document ID: US20200401900A1
- Status: Published application
- Kind: A1
- Application number: 16/830,480
- Filed: 2020-03-26
- Published: 2020-12-24
- Earliest claimed: 2019-06-18
- Assignees: Samsung Electronics Co., Ltd.; Seoul National University R&Db Foundation
- Applicants: Samsung Electronics Co., Ltd.; Seoul National University R&DB Foundation
- Inventors: Donghyun LEE; Euntae CHOI; Kyungmi LEE; Kiyoung CHOI
- Web page: https://patentgrep.com/patent/US20200401900A1

## CPC

- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/09: Supervised learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method of performing a class incremental learning in a neural network apparatus, the method including training an autoencoder using first input embeddings with respect to a first class group, calculating a contribution value of each of parameters of the autoencoder and calculating a representative value with respect to each of at least one first class included in the first class group in the training of the autoencoder, retraining the autoencoder using second input embeddings with respect to a second class group, and updating the contribution value of the each of the parameters and calculating a representative value with respect to each of at least one second class included in the second class group in the retraining the autoencoder.

## Claims

### Claim 1

A method of performing a class incremental learning in a neural network apparatus, the method comprising:
training an autoencoder using first input embeddings with respect to a first class group;
calculating a contribution value of each of parameters of the autoencoder and calculating a representative value with respect to each of at least one first class included in the first class group in the training of the autoencoder;
retraining the autoencoder using second input embeddings with respect to a second class group; and
updating the contribution value of the each of the parameters and calculating a representative value with respect to each of at least one second class included in the second class group in the retraining the autoencoder.

### Claim 2 (depends on 1)

The method of claim 1, wherein
the training of the autoencoder comprises:
converting the first input embeddings into latent variables of low dimensions using an encoder;
generating first reconstructed embeddings from the latent variables using a decoder; and
training the autoencoder by minimizing a first loss function with respect to the autoencoder such that the first reconstruction embeddings coincide with the first input embeddings.

### Claim 3 (depends on 2)

The method of claim 2, wherein
the retraining of the autoencoder comprises:
converting the second input embeddings into latent variables of low dimensions using the encoder;
generating second reconstructed embeddings from the latent variables using the decoder; and
retraining the autoencoder by minimizing a second loss function with respect to the autoencoder such that the second reconstruction embeddings coincide with the second input embeddings, wherein the second loss function comprises a term related to regularization based on an updated contribution value of each of the parameters.

### Claim 4 (depends on 3)

The method of claim 3, wherein
the first loss function and the second loss function comprise terms related to cosine similarity with respect to paired latent variables.

### Claim 5 (depends on 3)

The method of claim 3, wherein
the first loss function and the second loss function comprise a term related to L1-norm.

### Claim 6 (depends on 3)

The method of claim 3, wherein:
the training of the autoencoder comprises converting input embeddings with respect to the at least one first class into first latent variables and calculating a first representative value representing the first latent variables; and
the retraining of the autoencoder comprises converting input embeddings with respect to the at least one second class into second latent variables and calculating a second representative value representing the second latent variables.

### Claim 7 (depends on 6)

The method of claim 6, further comprising:
acquiring a test latent variable, in response to inputting test embedding into the autoencoder;
determining a representative value, from among the first representative value and the second representative value, having a higher similarity with the test latent variable; and
classifying the test embedding into a class corresponding to the determined representative value.

### Claim 8 (depends on 1)

The method of claim 1, wherein
the training of the autoencoder comprises:
converting an input data set with respect to the first class group into the first input embedding representing a vector value; and
training the autoencoder using the first input embedding, and
the retraining of the autoencoder comprises:
converting an input data set with respect to the second class group into the second input embedding representing another vector value; and
retraining the autoencoder using the second input embedding.

### Claim 9 (depends on 2)

The method of claim 2, wherein the training of the autoencoder comprises:
converting input embeddings with respect to the at least one first class into latent variables and calculating a first representative value representing the latent variables;
selecting latent variables having a difference with the first representative value that is less than or equal to a threshold from among the latent variables; and
training an autoencoder based on input embeddings corresponding to the selected latent variables.

### Claim 10 (depends on 1)

A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the dialogue management method of claim 1.

### Claim 11

A neural network apparatus that performs a class incremental learning, the neural network apparatus comprising:
a processor configured to
train an autoencoder using first input embeddings with respect to a first class group,
calculate a contribution value of each of parameters of the autoencoder and a representative value with respect to each of at least one first class included in the first class group in the training of the autoencoder,
retrain an autoencoder using second input embeddings with respect to a second class group, and
update the contribution value of the each of the parameters and to calculate a representative value with respect to each of at least one second class included in the second class group in retraining the autoencoder.

### Claim 12 (depends on 11)

The neural network apparatus of claim 11, wherein the processor is further configured to:
convert first input embeddings into latent variables of low dimensions using an encoder,
generate first reconstructed embeddings from the latent variables using a decoder, and
train the autoencoder by minimizing a first loss function with respect to the autoencoder such that the first reconstruction embeddings coincide with the first input embeddings.

### Claim 13 (depends on 11)

The neural network apparatus of claim 11, wherein the processor is further configured to:
convert the second input embeddings into latent variables of low dimensions using the encoder,
generate second reconstructed embeddings from the latent variables using the decoder, and
retrain the autoencoder by minimizing a second loss function with respect to the autoencoder such that the second reconstructed embeddings coincide with the second input embeddings,
wherein the second loss function comprises a term related to regularization based on an updated contribution value of each of the parameters.

### Claim 14 (depends on 13)

The neural network apparatus of claim 13, wherein
the first loss function and the second loss function comprise terms related to cosine similarity with respect to paired latent variables.

### Claim 15 (depends on 13)

The neural network apparatus of claim 13, wherein
the first loss function and the second loss function comprise a term related to L1-norm.

### Claim 16 (depends on 13)

The neural network apparatus of claim 13, wherein the processor is further configured to:
converts input embeddings with respect to the at least one first class into first latent variables and to calculate a first representative value representing the first latent variables in training the autoencoder; and
convert input embeddings with respect to the at least one second class into second latent variables and to calculate a second representative value representing the second latent variables in retraining the autoencoder.

### Claim 17 (depends on 16)

The neural network apparatus of claim 16, wherein the processor is further configured to:
acquire a test latent variable, in response to inputting test embedding into the autoencoder,
determine a representative value, from among the first representative value and the second representative value, having a higher similarity with the test latent variable, and
classify the test embedding into a class corresponding to the determined representative value.

### Claim 18 (depends on 11)

The neural network apparatus of claim 11, wherein the processor is further configured to:
convert an input data set with respect to the first class group into the first input embedding representing a vector value,
train the autoencoder using the first input embedding,
convert an input data set with respect to the second class group into the second input embedding that represents another vector value, and
retrain the autoencoder using the second input embedding.

### Claim 19 (depends on 12)

The neural network apparatus of claim 12, wherein the processor is further configured to:
convert input embeddings with respect to the at least one first class into latent variables and to calculate a first representative value representing the latent variables,
select latent variables having a difference with the first representative value that is less than or equal to a threshold from among the latent variables, and
train an autoencoder based on input embeddings corresponding to the selected latent variables.

### Claim 20 (depends on 11)

The neural network apparatus of claim 11, further comprising
a memory configured to store the contribution value of each of the parameters and the representative value.