Granted Patent B2

US 12,608,604 · App. 17/533,679 · Granted Apr 21, 2026

Inventors: Ye Bin Moon (Daejeon, KR); Tae Hyun Oh (Daejeon, KR); Yong Jin Kwon (Daejeon, KR)

Assignee: ELECTRONICS AND TELECOMMUNICATIONS RESEARCH INSTITUTE

G06N3/08G06F18/10G06F18/22

Abstract

Claims (50)

1 . A computer-implemented method for training artificial intelligence based on an episodic memory, the method comprising:

constructing an episodic memory to include a representation of a feature vector stored in a full memory, the feature vector being associated with a training dataset;

obtaining output data by inputting query data into an artificial intelligence model;

deriving a similarity between the output data and the representation in the episodic memory; and

deriving an episode loss function based on the similarity,

wherein the feature vector is generated by inputting the training dataset into a pretrained deep neural network and includes a plurality of numerical elements,

wherein constructing the episodic memory, obtaining the output data, deriving the similarity, and deriving the episode loss function are repeated a plurality of times during training to iteratively update one or more parameters of the artificial intelligence model using backpropagation based on the episode loss function and an optimization algorithm, and

wherein, after each iteration of the training, the full memory is configured to update the feature vector stored in the full memory by:

inputting at least a portion of the training dataset into the artificial intelligence model after the one or more parameters are updated to generate updated output data;

generating an updated feature vector based on the updated output data; and

storing the updated feature vector in the full memory in association with the training dataset prior to constructing the episodic memory for a next iteration of the training.

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

3 . The computer-implemented 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 computer-implemented 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 computer-implemented method of claim 4 , 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 computer-implemented method of claim 1 , wherein the artificial intelligence model comprises a first artificial intelligence model and a second artificial intelligence model, 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.

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

8 . The computer-implemented method of claim 7 , further comprising deriving a final loss function by applying a weight to the episode loss function and the knowledge distillation loss function.

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

10 . The computer-implemented method of claim 1 , wherein the full memory stores a class label corresponding to the feature vector.

11 . The computer-implemented method of claim 10 , wherein a class label of a feature vector with a highest similarity is allocated as a class label of the query data.

12 . The computer-implemented method of claim 1 , wherein the training dataset comprises the query data and random data.

13 . The computer-implemented method of claim 1 , wherein a plurality of the query data forms mini batch.

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

15 . The computer-implemented method of claim 1 , wherein the episodic memory is constructed only once per training iteration.

16 . An apparatus for training artificial intelligence based on an episodic memory, the apparatus comprising:

a full memory comprising a feature vector associated with a training dataset; and

a processor configured to control the full memory,

wherein the processor is further configured to:

construct an episodic memory to include a representation of the feature vector;

obtain output data by inputting query data into an artificial intelligence model;

derive a similarity between the output data and the representation in the episodic memory; and

derive an episode loss function based on the similarity,

wherein the feature vector is generated by inputting the training dataset into a pretrained deep neural network and includes a plurality of numerical elements,

wherein the processor is configured to repeatedly construct the episodic memory, obtain the output data, derive the similarity, and derive the episode loss function a plurality of times during training to iteratively update one or more parameters of the artificial intelligence model using backpropagation based on the episode loss function and an optimization algorithm, and

wherein, after each iteration of the training, the processor is configured to update the feature vector stored in the full memory by:

inputting at least a portion of the training dataset into the artificial intelligence model after the one or more parameters are updated to generate updated output data;

generating an updated feature vector based on the updated output data; and

storing the updated feature vector in the full memory in association with the training dataset prior to constructing the episodic memory for a next iteration of the training.

17 . A non-transitory computer-readable medium storing a computer program, the computer program implementing:

constructing an episodic memory to include a representation of a feature vector stored in a full memory, the feature vector being associated with a training dataset;

obtaining output data by inputting query data into an artificial intelligence model;

deriving a similarity between the output data and the representation in the episodic memory; and

deriving an episode loss function based on the similarity,

wherein the feature vector is generated by inputting the training dataset into a pretrained deep neural network and includes a plurality of numerical elements,

wherein constructing the episodic memory, obtaining the output data, deriving the similarity, and deriving the episode loss function are repeated a plurality of times during training to iteratively update one or more parameters of the artificial intelligence model using backpropagation based on the episode loss function and an optimization algorithm, and

wherein, after each iteration of the training, the computer program is configured to update the feature vector stored in the full memory by:

inputting at least a portion of the training dataset into the artificial intelligence model after the one or more parameters are updated to generate updated output data;

generating an updated feature vector based on the updated output data; and

storing the updated feature vector in the full memory in association with the training dataset prior to constructing the episodic memory for a next iteration of the training.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (28)