Tata Consultancy Services Limited

Gaurab Bhattacharya

Jayavardhana Rama Gubbi Lakshminarasimha

Balamuralidhar Purushothaman

Nikhil Kilari

Bagyalakshmi Vasudevan

Arpan Pal

Vartika Sengar

Vivek Bangalore Sampathkumar

Bagya Lakshmi Vasudevan

Kalyan Prakash Baishya

Kuruvilla Abraham

This disclosure relates generally to identification and mitigation of bias while training deep learning models. Conventional methods do not provide effective methods for bias identification, and they require pre-defined concepts and rules for bias mitigation. The embodiments of the present disclosure train an auto-encoder to produce a generalized representation of an input image by decomposing into a set of latent embedding. The set of latent embedding are used to learn the shape and color concepts of the input image. The feature specialization is done by training an auto-encoder to reconstruct the input image using the shape embedding modulated by color embedding. To identify the bias, permutation invariant neural network is trained for classification task and attribution scores corresponding to each concept embedding are computed. The method also performs de-biasing the classifier by training it with a set of counterfactual images generated by modifying the latent embedding learned by the auto-encoder.

Method and system for identifying and mitigating bias while training deep learning models

Method and system for generating color variants for fashion apparels by providing a Fashion Apparel Regeneration—Generative Adversarial Network (FAR—GAN) to generate color variants of the fashion apparels. The FAR—GAN utilizes a two-step encoding process to encapsulate both an input image and an edge-map information along with a target color embedding branch to manipulate the color information present in the fashion apparel present in the input image that is to be changed to a desired target color. Furthermore, the color and structural information is disentangled by controlling them using a color consistency loss. The FAR—GAN can be trained end-to-end without incorporating complex multi-step process.

Method and system for generating color variants for fashion apparels

Product recommendation is a very important aspect of e-commerce applications. Traditional product recommendation systems recommend products similar to a query image provided by a user and allows minimum or no personalization. It is challenging to incorporate personalization due to the presence of overlapping fine-grained attributes, variations in attribute style and visual appearance, small inter-class variation and class imbalance in the images of products. Embodiments of present disclosure address these challenges by a method of personalized substitute product recommendation using Personalized Attribute Search Networks (PAtSNets) comprising neural network layers interleaved with Attentive Style Embedding (ASE) modules to generate attribute-aware feature representation vector of a query image provided by the user and conforming to the personalization instructions specified by the user. This feature representation vector is then used to recommend substitute products to the user. Thus, embodiments of present disclosure enable accurate substitute product recommendation suiting user requirements.

Method, system, and non-transitory machine-readable information storage medium for personalized substitute product recommendation

Traditional systems used for fashion attribute detection struggle to generate accurate predictions due to presence of large intra-class and relatively small inter-class variations in data related to the fashion attributes. The disclosure herein generally relates to image processing, and, more particularly, to a method and system for fashion attribute detection. The method proposes F-AttNet, an attribute extraction network to leverage the performance of fine-grained localized fashion attribute recognition. F-AttNet comprises Attentive Multi-scale Feature Encoder (AMF) blocks that encapsulate multi-scale fine-grained attribute information upon adaptive recalibration of channel weights. F-AttNet is designed by hierarchically stacking the AMF encoders to extract deep fine-grained information across multiple scales. A data model used by F-AttNet is trained using a novel γ-variant focal loss function for addressing the class imbalance problem by penalizing wrongly classified examples and incorporating separate importance to positive and negative instances.

Method and system for fashion attribute detection

Growing community of inventors

Bengaluru, India

Gaurab Bhattacharya