# US 12,223,432 B2 · Using disentangled learning to train an interpretable deep learning model

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

- Number: US 12,223,432 B2
- Document ID: US12223432B2
- Status: Granted
- Kind: B2
- Application number: 17/133,437
- Filed: 2020-12-23
- Published: 2022-06-23
- Granted: 2025-02-11
- Earliest claimed: 2020-12-23
- Assignees: International Business Machines Corporation
- Applicants: International Business Machines Corporation
- Inventors: Supriyo Chakraborty; Seraphin Bernard Calo; Jiawei Wen
- Web page: https://patentgrep.com/patent/US12223432B2

## CPC

- G06N 3/08: Learning methods
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06F 18/2137: based on criteria of topology preservation, e.g. multidimensional scaling or self-organising maps
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/0475: Generative networks
- G06N 3/09: Supervised learning
- G06V 30/274: {Syntactic or semantic context, e.g. balancing}
- G06N 5/01: Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
- G06N 20/20: Ensemble learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

A method and system of training an interpretable deep learning model includes receiving an input set of data, which may be complex. The input set of data is provided to deep learning model for feature extraction. In an exemplary embodiment, the deep learning model generates a disentangled latent space of features from the feature extraction. The features may comprise semantically meaningful data which is then provided to a low-complexity learning model. The low-complexity learning model generates output based on a specified task (for example, classification or regression). Being a low-complexity learning model provides confidence that the data output from the deep learning model is inherently interpretable.

## Claims

### Claim 1

A method of training an interpretable deep learning model for a machine learning system, comprising:
receiving an input set of data;
providing the input set of data to a deep neural network model;
extracting features from the deep neural network model;
generating a latent space of vectors comprising the extracted features;
feeding the latent space of vectors generated from extracted features of the deep neural network model to a task-specific model, wherein the task-specific model is a low-complexity and linear learning model;
generating interpretable predictions of feature dimensions from the task-specific model;
reconstructing the input set of data using a decoder module;
determining a reconstruction error loss from reconstructing the input set of data;
determining a classification loss or a regression loss from a task specific output set of data; and
training an autoencoder, the decoder module, and the low-complexity learning model, using a combination of (i) the reconstruction error loss and (ii) the classification loss or the regression loss.

### Claim 2 (depends on 1)

The method of claim 1, wherein the features are extracted using an encoder module.

### Claim 3 (depends on 1)

The method of claim 1, wherein the latent space of vectors are disentangled representations of the input set of data.

### Claim 4 (depends on 1)

The method of claim 1, wherein the autoencoder is a Beta variational autoencoder, and the method further comprising using the Beta variational autoencoder to extract the features from the deep neural network.

### Claim 5 (depends on 1)

The method of claim 1, further comprising:
associating a feature dimension for each vector in the latent space with a semantically meaningful characteristic; and
generating the interpretable predictions of feature dimensions based on the semantically meaningful characteristic of each vector.

### Claim 6

A computer program product for training an interpretable deep learning model for a machine learning system, the computer program product comprising:
one or more non-transitory computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising:
receiving an input set of data;
providing the input set of data to a deep neural network model;
extracting features from the deep neural network model;
generating a latent space of vectors comprising the extracted features;
feeding the latent space of vectors generated from extracted features of the deep neural network model to a task-specific model, wherein the task-specific model is a low-complexity and linear learning model;
generating interpretable predictions of feature dimensions from the task-specific model;
reconstructing the input set of data using a decoder module;
determining a reconstruction error loss from reconstructing the input set of data;
determining a classification loss or a regression loss from a task specific output set of data; and
training an autoencoder, the decoder module, and the low-complexity learning model, using a combination of (i) the reconstruction error loss and (ii) the classification loss or the regression loss.

### Claim 7 (depends on 6)

The computer program product of claim 6, wherein the features are extracted using an encoder module.

### Claim 8 (depends on 6)

The computer program product of claim 6, wherein the latent space of vectors are disentangled representations of the input set of data.

### Claim 9 (depends on 6)

The computer program product of claim 6, wherein:
the autoencoder is a Beta variational autoencoder; and
the program instructions further comprise using the Beta variational autoencoder to extract the features from the deep neural network.

### Claim 10 (depends on 6)

The computer program product of claim 6, wherein the program instructions further comprise:
associating a feature dimension for each vector in the latent space with a semantically meaningful characteristic; and
generating the interpretable predictions of feature dimensions based on the semantically meaningful characteristic of each vector.

### Claim 11

A computer server, comprising:
a network connection;
one or more computer readable storage media;
a processor coupled to the network connection and coupled to the one or more computer readable storage media; and
a computer program product comprising program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising:
receiving an input set of data;
providing the input set of data to a deep neural network model;
extracting features from the deep neural network model;
generating a latent space of vectors comprising the extracted features;
feeding the latent space of vectors generated from extracted features of the deep neural network model to a task-specific model to a task-specific model, wherein the task-specific model is a low-complexity and linear learning model;
generating interpretable predictions of feature dimensions from the task-specific model;
reconstructing the input set of data using a decoder module;
determining a reconstruction error loss from reconstructing the input set of data;
determining a classification loss or a regression loss from a task specific output set of data; and
training an autoencoder, the decoder module, and the low-complexity learning model, using a combination of (i) the reconstruction error loss and (ii) the classification loss or the regression loss.

### Claim 12 (depends on 11)

The computer server of claim 11, wherein the features are extracted using an encoder module.

### Claim 13 (depends on 11)

The computer server of claim 11, wherein the latent space of vectors are disentangled representations of the input set of data.

### Claim 14 (depends on 13)

The computer server of claim 13, wherein:
the autoencoder is a Beta variational autoencoder; and
the program instructions further comprise using the Beta variational autoencoder to extract the features from the deep neural network.

### Claim 15 (depends on 11)

The computer server of claim 11, wherein the program instructions further comprise:
associating a feature dimension for each vector in the latent space with a semantically meaningful characteristic; and
generating the interpretable predictions of feature dimensions based on the semantically meaningful characteristic of each vector.

### Claim 16

A method of training an interpretable deep learning model for a machine learning system, comprising:
receiving an input set of data;
providing the input set of data to a Beta variational autoencoder;
generating by the Beta variational autoencoder, an output representation of the input set of data;
processing the output representation generated by the Beta variational autoencoder, using a low-complexity linear learning model;
determining by the low-complexity learning model a task specific output set of data;
providing an interpretation of the input set of data based on the task specific output set of data;
reconstructing the input set of data using a decoder module;
determining a reconstruction error loss from reconstructing the input set of data;
determining a classification loss or a regression loss from the task specific output set of data; and
training an autoencoder, the decoder module, and the low-complexity learning model, using a combination of (i) the reconstruction error loss and (ii) the classification loss or the regression loss.

### Claim 17 (depends on 16)

The method of claim 16, wherein the output representation of the input set of data generated by the Beta variational autoencoder, is a latent space of dimensional vectors, organized by features with a semantic relationship.

### Claim 18 (depends on 16)

The method of claim 16, wherein the low-complexity learning model is one of: a parametric model, a non-parametric model, a decision tree, a regression tree, or an ensemble model.

### Claim 19

A computer program product for training an interpretable deep learning model for an artificial intelligence computing system, the computer program product comprising:
one or more non-transitory computer readable storage media, and program instructions collectively stored on the one or more computer readable storage media, the program instructions comprising:
receiving an input set of data;
providing the input set of data to a Beta variational autoencoder;
generating by the Beta variational autoencoder, an output representation of the input set of data;
processing the output representation generated by the Beta variational autoencoder using a low-complexity learning model that is linear;
determining by the low-complexity learning model a task specific output set of data;
providing an interpretation of the input set of data based on the task specific output set of data;
reconstructing the input set of data using a decoder module;
determining a reconstruction error loss from reconstructing the input set of data;
determining a classification loss or a regression loss from the task specific output set of data; and
training the Beta variational autoencoder, the decoder module, and the low-complexity learning model using a combination of (i) the reconstruction error loss and (ii) the classification loss or the regression loss.

### Claim 20 (depends on 19)

The computer program product of claim 19, wherein the low-complexity learning model is one of: a parametric model, a non-parametric model, a decision tree, a regression tree, or an ensemble model.