Toddler-inspired Bayesian learning method and computing apparatus for performing the same

Patent metadata

Field	Value
Publication number	US-12536454-B2
Application number	US-202117467971-A
Country	US
Kind code	B2
Filing date	Sep 7, 2021
Priority date	Nov 26, 2020
Publication date	Jan 27, 2026
Grant date	Jan 27, 2026

Abstract

A toddler-inspired Bayesian learning method according to an embodiment includes: collecting information related to at least one task while an agent is performing exploration; and performing Bayesian inference regarding the at least one task by using the collected information as Bayesian informative priors.

First claim

What is claimed is:

1 . A toddler-inspired Bayesian learning method performed by a computing device including at least one processor, a memory storing a toddler-inspired Bayesian learning model and result data of the toddler-inspired Bayesian learning model, and a communication module transmitting the result data of the toddler-inspired Bayesian learning model, wherein the computing device implements a virtual environment and virtual objects in the virtual environment, wherein the processor executes instructions to implement the toddler-inspired Bayesian learning model, the method for performing image classification and distance estimation for one of the virtual objects comprising:
collecting visual information related to the image classification and the distance estimation for the one virtual object while an agent of the virtual environment is performing exploration in the virtual environment through interaction with the virtual objects in the virtual environment; and performing Bayesian inference regarding the image classification and the distance estimation for the one virtual object by using the collected visual information as a prior distribution to be used for the Bayesian inference, wherein performing the Bayesian inference comprises:
obtaining a feature map by passing the collected visual information through an artificial neural network;
calculating the prior distribution to be used for the Bayesian inference by using the feature map and a weight matrix learned during the exploration of the agent; and
calculating posterior probabilities respectively related to the image classification and the distance estimation for the one virtual object by applying the prior distribution as a same prior distribution to a plurality of Bayesian Support Vector Machines (SVMs) respectively,
wherein obtaining the feature map comprises obtaining the feature map by inputting the collected visual information into a convolutional neural network in which learning related to the image classification and the distance estimation has been performed, wherein calculating the prior distribution comprises:
converting the feature map into a feature vector by using a flatten function; and
calculating the prior distribution by linearly projecting the feature vector through the weight matrix,
wherein calculating the prior distribution by linearly projecting the feature vector comprises calculating the same prior distribution by applying a result, obtained by multiplying the feature vector and a transpose of the weight matrix, to a radial basis function (RBF) kernel, wherein calculating the posterior probabilities respectively related to the image classification and the distance estimation for the one virtual object comprises:
inputting the same prior distribution to a first Bayesian Support Vector Machine for classifying images of the one virtual object, calculating a first posterior probability for the image classification through the first Bayesian Support Vector Machine, and outputting an inference result for the image classification of the one virtual object based on the first posterior probability; and
inputting the same prior distribution to a second Bayesian Support Vector Machine for estimating distances of the one virtual object, calculating a second posterior probability for the distance estimation through the second Bayesian Support Vector Machine, and outputting an inference result for the distance estimation of the one virtual object based on the second posterior probability.

2 . The toddler-inspired Bayesian learning method of claim 1 , wherein collecting the visual information comprises collecting the visual information in such a manner that the agent randomly performs a plurality of types of operations in the virtual environment according to a reinforcement learning algorithm and obtains a reward according to interaction with objects present in the virtual environment.

3 . A non-transitory computer-readable storage medium having stored thereon a program that, when executed by a computer, causes the computer to execute the toddler-inspired Bayesian learning method set forth in claim 1 .

4 . A computer program that is executed by a computing apparatus and stored in a medium in order to perform the toddler-inspired Bayesian learning method set forth in claim 1 .

5 . A computing apparatus for implementing a virtual environment and virtual objects in the virtual environment to perform image classification and distance estimation for one of the virtual objects by performing toddler-inspired Bayesian learning, the computing apparatus comprising:
a controller including at least one processor, and configured to perform the toddler-inspired Bayesian learning by executing a program; a memory storing a toddler-inspired Bayesian learning model and result data of the toddler-inspired Bayesian learning model; and a communication module transmitting the result data of the toddler-inspired Bayesian learning model, wherein the processor, by executing instructions to implement the toddler-inspired Bayesian learning model, is configured to collect visual information related to the image classification and the distance estimation for the one virtual object in a process in which an agent of the virtual environment is performing exploration in the virtual environment through interaction with the virtual objects in the virtual environment, and perform Bayesian inference regarding the image classification and the distance estimation for the one virtual object by using the collected visual information as a prior distribution to be used for the Bayesian inference, wherein when performing the Bayesian inference, the processor is configured to obtain a feature map by passing the collected visual information through an artificial neural network; calculate the prior distribution to be used for the Bayesian inference by using the feature map and a weight matrix learned during the exploration of the agent; and calculate posterior probabilities respectively related to the image classification and the distance estimation for the one virtual object by applying the prior distribution as a same prior distribution to a plurality of Bayesian Support Vector Machines (SVMs) respectively, wherein when obtaining the feature map, the toddler-inspired Bayesian learning model obtains the feature map by inputting the collected visual information into a convolutional neural network in which learning related to the image classification and the distance estimation has been performed, wherein when calculating the prior distribution, the toddler-inspired Bayesian learning model converts the feature map into a feature vector by using a flatten function, and calculates the prior distribution by linearly projecting the feature vector through the weight matrix, wherein when calculating the prior distribution by linearly projecting the feature vector, the toddler-inspired Bayesian learning model calculates the same prior distribution by applying a result, obtained by multiplying the feature vector and a transpose of the weight matrix, to a radial basis function (RBF) kernel, wherein when calculating the posterior probabilities respectively related to the image classification and the distance estimation for the one virtual object, the processor is configured to input the same prior distribution to a first Bayesian Support Vector Machine for classifying images of the one virtual object, calculate a first posterior probability for the image classification through the first Bayesian Support Vector Machine, and output an inference result for the image classification of the one virtual object based on the first posterior probability; and input the same prior distribution to a second Bayesian Support Vector Machine for estimating distances of the interacted

Assignees

• Seoul Nat Univ R&Db Foundation

Inventors

• Zhang Byoung-Tak
• Lee Youngki
• Park Junseok
• Park Kwanyoung
• Oh Hyunseok
• Jang You Won

Classifications

• G06N20/10 Primary

  using kernel methods, e.g. support vector machines [SVM] · CPC title

• G06N7/01 Primary

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• G06N3/0895

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

• G06N3/092

  Reinforcement learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title