Browse Prior Art Database

DERIVING GRAPH STRUCTURES WITHIN DEEP ARCHITECTURES FOR ROBUST HANDLING OF LIMITED TRAINING DATA

IP.com Disclosure Number: IPCOM000249657D
Publication Date: 2017-Mar-15
Document File: 4 page(s) / 383K

Publishing Venue

The IP.com Prior Art Database

Abstract

A technique for deriving graph structures within deep architecture for robust handling of limited training data is disclosed. The graph structures are created at each of the deep architecture layer, resulting in richer representation of patch training data using fewer parameters than regular deep convolutional neural networks (D-CNN) implementations. Relationship between the intermediate hidden layer nodes is used for constructing subsequent output nodes. Encoding hidden layer relationship is useful in capturing context and provides robustness in the presence of outlier training data, thereby resulting in lesser parameters while creating the output layer.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 48% of the total text.

DERIVING GRAPH STRUCTURES WITHIN DEEP ARCHITECTURES FOR ROBUST HANDLING OF LIMITED TRAINING DATA

BACKGROUND

 

The present disclosure relates generally to deep architectures and more particularly to a technique for deriving graph structures within deep architecture for robust handling of limited training data.

Deep architectures like deep convolutional neural networks (D-CNN or D-NN) are effective methods used in different classification and recognitions tasks. D-CNN is successfully used in modeling non-linear representations.

Further, deep learning is widely used in healthcare applications for quantification, characterization and precursor image analysis tasks including detection, classification, de-noising, segmentation, and super-resolution etc. However, one major challenge with healthcare application is pixel wise classification of the image. Effective modeling of non-linearity in classification and recognition task involves employment of excess free parameters, which are generally not abundantly available in small training data. Hence, the number of features (p) is greater than the number of training instances (N) resulting in overfitting and instability during training. This leads to inaccurate results during product deployment.

Several conventional techniques are used to overcome the problem of small training data. Some conventional techniques use regularization and reusing or initialization of network parameters from previous training tasks, while others deal with graph data using spectral convolutions, unstructured data or use graph data in an auxiliary manner. Also, graph structures are learnt for latent variables and used within factor analysis. However, conventional techniques do not effectively provide rich representation of training data using few parameters.

It would be desirable to have an improved technique of handing limited training data using graph structure from within the deep CNN.

BRIEF DESCRIPTION OF DRAWINGS

Figure 1 depicts a technique for deriving graph structures within each layer of deep convolutional neural networks.

DETAILED DESCRIPTION

A technique for deriving graph structures within deep architecture for robust handling of limited training data is disclosed. The graph structures are created at each of the deep architecture layer, resulting in richer representation of patch training data using fewer parameters than regular deep convolutional neural networks (D-CNN) implementations.

Relationship between the intermediate hidden layer nodes is used for constructing subsequent output nodes. Encoding hidden layer relationship is useful in capturing context and provides robustness in the presence of outlier training data, thereby resulting in lesser parameters while creating the output layer.

Figure 1 depicts the technique for deriving graph structures within each layer of D-CNN. At a certain layer, there are N input feature maps  and M output feature maps. Of M, one can select M1 feature maps  in the usual way usi...