# US 2022/0044065 A1 · System and method for parameter compression of capsule networks using deep features

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

- Number: US 2022/0044065 A1
- Document ID: US20220044065A1
- Status: Published application
- Kind: A1
- Application number: 17/377,628
- Filed: 2021-07-16
- Published: 2022-02-10
- Earliest claimed: 2020-07-17
- Assignees: Tata Consultancy Services Limited
- Applicants: Tata Consultancy Services Limited
- Inventors: Chandan Kumar SINGH; Vivek Kumar GANGWAR; Anima MAJUMDER; Prakash Chanderlal AMBWANI; Rajesh SINHA
- Web page: https://patentgrep.com/patent/US20220044065A1

## CPC

- G06F 18/21: Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06V 10/454: {Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]}
- G06F 18/2431: Multiple classes
- G06N 3/045: Combinations of networks
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/0464: Convolutional networks [CNN, ConvNet]
- G06N 3/0495: Quantised networks; Sparse networks; Compressed networks
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06V 10/40: Extraction of image or video features
- G06N 3/088: Non-supervised learning, e.g. competitive learning
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

This disclosure relates generally to system and method for parameter compression of capsule networks using deep features. The conventional capsule networks have distinct capability of retaining spatial correlations between extracted features but that comes at a cost of intensive computational, cost, memory usage and bandwidth requirement. The embodiments herein disclose a system and method for employing a lightweight deep features based capsule network that is capable of compressing the parameters. In an embodiment, the system includes a deep feature based capsule network such that the capsule layer is preceded by feature blocks. Said feature blocks comprises convolutional operation with a kernel size 3, followed by convolutional operation with kernel of size 1, and a Batch Normalization layer, and hence are able to extract deep features. These deep features when input to the capsule layer facilitates in compressing the parameters of the capsule layer, while ensuring significant increase in recognition accuracy.

## Claims

### Claim 1

A processor implemented method, comprising:
employing a deep feature based Capsule Network for a task, via one or more hardware processors, the deep feature based Capsule Network comprising:
a feature extraction layer comprising a set of feature extraction units wherein each feature extraction unit of the set of feature extraction units comprises 3×3 convolutional operation with a kernel size 3, followed by convolutional operation with a kernel of size 1, and a batch normalization (BN) layer,
a primary capsule layer comprising a set of primary Capsules, and
a class capsule layer comprising a set of class capsules,
wherein employing the deep feature based Capsule Network comprises:
passing, successively through each of the set of feature extraction units, an input data to obtain a set of deep features;
performing, by the primary capsule layer, a convolution of the set of deep features into the set of primary capsules, wherein the set of primary capsules in the primary capsule layer comprises an optimal number of capsules based on a number of features in the set of deep features; and
predicting, at the class capsule layer, a set of classes from the set of deep features, wherein a number of capsules in the set of class capsules is equal to a number of classes associated with the task.

### Claim 2 (depends on 1)

The processor implemented method of claim 1, wherein the optimal number of capsules based on the number of deep features in the set of deep features comprises a minimum number of capsules satisfying a performance criteria and a parameter complexity criteria associated with the task, wherein the performance criteria is indicative of a performance expectation of the deep feature based Capsule Network, and wherein the parameter complexity criteria is indicative of complexity of parameters associated with the task, and wherein the set of deep features facilitates in parameter compression in the capsule layer.

### Claim 3 (depends on 1)

The processor implemented method of claim 1, the number of capsules in the class capsule layer is equal to the number of classes associated with the task, and wherein each capsule in the class capsule layer comprises a 1-D array with a dimension of 16.

### Claim 4 (depends on 1)

The processor implemented method of claim 1, wherein the deep feature based Capsule Network further comprises a decoder communicatively coupled with the class capsule layer, the decoder capable of receiving the set of classes predicted by the class capsule layer and reconstructing an image associated with the task based on the set of classes during a training phase of capsule based network.

### Claim 5

A system comprising:
one or more memories; and
one or more first hardware processors, the one or more first memories coupled to the one or more first hardware processors, wherein the one or more first hardware processors are configured to execute programmed instructions stored in the one or more first memories, to:
employ a deep feature based Capsule Network for a task, the Capsule Network comprising a feature extraction layer comprising a set of feature extraction units wherein each feature extraction unit of the set of feature extraction units comprises convolutional operation with a kernel size 3, followed by convolutional operation with a kernel of size 1, and a batch normalization (BN) layer, a primary capsule layer comprising a set of primary Capsules, and a class capsule layer comprising a set of class capsules, wherein to employ the deep feature based Capsule Network for the task, the one or more first hardware processors are configured by the instructions to;
pass, successively through each of the set of feature extraction units, an input data to obtain set of deep features;
perform, by the primary capsule layer, a convolution of the set of deep features into the set of primary capsules, wherein the set of primary capsules in the primary capsule layer comprises an optimal number of capsules based on a number of features in the set of deep features; and
predict, at the class capsule layer, a set of classes from the set of deep features, wherein a number of capsules in the set of class capsules is equal to a number of classes associated with the task.

### Claim 6 (depends on 5)

The system of claim 5, wherein the optimal number of capsules based on a number of deep features in the set of deep features comprises a minimum number of capsules satisfying a performance criteria and a parameter complexity criteria associated with the task, wherein the performance criteria is indicative of a performance expectation of the deep feature based Capsule Network, and wherein the parameter complexity criteria is indicative of complexity of parameters associated with the task, and wherein the set of deep features facilitates in parameter compression in the capsule layer.

### Claim 7 (depends on 5)

The system of claim 5, the number of capsules in the class capsule layer is equal to the number of classes associated with the task, and wherein each capsule in the class capsule layer comprises a 1-D array with a dimension of 16.

### Claim 8 (depends on 5)

The system of claim 5, wherein the deep feature based Capsule Network further comprises a decoder communicatively coupled with the class capsule layer, the decoder capable of receiving the set of classes predicted by the class capsule layer and reconstructing an image associated with the task based on the set of classes during a training phase of capsule based network.

### Claim 9

One or more non-transitory machine readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause managing a plurality of events, the instructions cause:
employing a deep feature based Capsule Network for a task, the deep feature based Capsule Network comprising:
a feature extraction layer comprising a set of feature extraction units wherein each feature extraction unit of the set of feature extraction units comprises 3×3 convolutional operation with a kernel size 3, followed by convolutional operation with a kernel of size 1, and a batch normalization (BN) layer,
a primary capsule layer comprising a set of primary Capsules, and
a class capsule layer comprising a set of class capsules,
wherein employing the deep feature based Capsule Network comprises:
passing, successively through each of the set of feature extraction units, an input data to obtain a set of deep features;
performing, by the primary capsule layer, a convolution of the set of deep features into the set of primary capsules, wherein the set of primary capsules in the primary capsule layer comprises an optimal number of capsules based on a number of features in the set of deep features; and
predicting, at the class capsule layer, a set of classes from the set of deep features, wherein a number of capsules in the set of class capsules is equal to a number of classes associated with the task.
