Method and apparatus for training artificial intelligence based on episode memory

Patent metadata

Field	Value
Publication number	US-2023041614-A1
Application number	US-202117533679-A
Country	US
Kind code	A1
Filing date	Nov 23, 2021
Priority date	Aug 9, 2021
Publication date	Feb 9, 2023
Grant date	—

Abstract

The present disclosure relates to a method and apparatus for training artificial intelligence based on an episodic memory. According to an embodiment of the present disclosure, a method for training artificial intelligence based on an episodic memory may include: constructing an episodic memory by using a feature vector of a training dataset stored in a full memory; obtaining output data by inputting query data into an artificial intelligence model; deriving a similarity between the output data and a feature vector in the constructed episodic memory; and deriving an episode loss function based on the similarity.

First claim

What is claimed is:

1 . A method for training artificial intelligence based on an episodic memory, the method comprising:
constructing an episodic memory by using a feature vector of a training dataset stored in a full memory; obtaining output data by inputting query data into an artificial intelligence model; deriving a similarity between the output data and a feature vector in the constructed episodic memory; and deriving an episode loss function based on the similarity.

2 . The method of claim 1 , wherein the episodic memory comprises an index in the full memory for a feature vector comprised in the episodic memory and a matrix representation of the feature vector.

3 . The method of claim 1 , wherein the episode loss function is derived based on at least one of a hard-attention loss function and a soft-attention loss function.

4 . The method of claim 3 , wherein the hard-attention loss function is derived based on a probability that an arbitrary slot in the episodic memory corresponds to the query data.

5 . The method of claim 3 , wherein, in response to the arbitrary slot in the episodic memory corresponding to the query data, the soft-attention loss function is derived based on a difference between the output data that are obtained through the artificial intelligence model by using a probability that another arbitrary slot in the episodic memory corresponds to the query data.

6 . The method of claim 1 , wherein the artificial intelligence model comprises a first artificial intelligence model and a second artificial intelligence model, and
wherein the second artificial intelligence model is a teacher model that forwards intrinsic knowledge to the first artificial intelligence model through knowledge distillation.

7 . The method of claim 6 , wherein the second artificial intelligence model comprises a pretrained artificial neural network that performs a same type of task as the first artificial intelligence model.

8 . The method of claim 6 , further comprising deriving a knowledge distillation loss function by using the second artificial intelligence model.

9 . The method of claim 8 , further comprising deriving a final loss function by applying a weight to the episode loss function and the knowledge distillation loss function.

10 . The method of claim 1 , wherein the artificial intelligence model is based on a convolutional neural network (CNN) or an autoencoder.

11 . The method of claim 1 , wherein the full memory stores a class label corresponding to the feature vector.

12 . The method of claim 11 , wherein a class label of a feature vector with a highest similarity is allocated as a class label of the query data.

13 . The method of claim 1 , wherein a feature vector of the training dataset is stored in the full memory and is updated at a predetermined interval.

14 . The method of claim 1 , wherein the training dataset comprises the query data and random data.

15 . The method of claim 1 , wherein a plurality of the query data forms mini batch.

16 . The method of claim 1 , further comprising initializing the artificial intelligence model and the full memory before the episodic memory is constructed.

17 . The method of claim 1 , further comprising performing back propagation by updating a parameter of the artificial intelligence model after the episode loss function is derived.

18 . The method of claim 1 , further comprising reconstructing the episodic memory by using a feature vector of the training dataset stored in the full memory, after the episode loss function is derived.

19 . An apparatus for training artificial intelligence based on an episodic memory, the apparatus comprising:
a memory constructed to store a feature vector of a training dataset; and a processor configured to control the memory, wherein the processor is further configured to: construct an episodic memory by using a feature vector of a training dataset stored in a full memory, obtain output data by inputting query data into an artificial intelligence model, derive a similarity between the output data and a feature vector in the constructed episodic memory, and derive an episode loss function based on the similarity.

20 . A computer program stored in a non-transitory computer-readable medium, the computer program implementing:
constructing an episodic memory by using a feature vector of a training dataset stored in a full memory; obtaining output data by inputting query data into an artificial intelligence model; deriving a similarity between the output data and a feature vector in the constructed episodic memory; and deriving an episode loss function based on the similarity.

Assignees

• Electronics & Telecommunications Res Inst

Inventors

• Moon Ye Bin
• Oh Tae Hyun
• Kwon Yong Jin

Classifications

• G06N3/0895 Primary

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

• G06N3/096

  Transfer learning · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title

• G06N3/0455 Primary

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title