20240013522. METHOD AND SYSTEM FOR IDENTIFYING AND MITIGATING BIAS WHILE TRAINING DEEP LEARNING MODELS simplified abstract (Tata Consultancy Services Limited)

METHOD AND SYSTEM FOR IDENTIFYING AND MITIGATING BIAS WHILE TRAINING DEEP LEARNING MODELS

Organization Name

Tata Consultancy Services Limited

Inventor(s)

Jayavardhana Rama Gubbi Lakshminarasimha of Bangalore (IN)

Vartika Sengar of Bangalore (IN)

Vivek Bangalore Sampathkumar of Bangalore (IN)

Gaurab Bhattacharya of Bangalore (IN)

Balamuralidhar Purushothaman of Bangalore (IN)

Arpan Pal of Kolkata (IN)

METHOD AND SYSTEM FOR IDENTIFYING AND MITIGATING BIAS WHILE TRAINING DEEP LEARNING MODELS - A simplified explanation of the abstract

This abstract first appeared for US patent application 20240013522 titled 'METHOD AND SYSTEM FOR IDENTIFYING AND MITIGATING BIAS WHILE TRAINING DEEP LEARNING MODELS

Simplified Explanation

The disclosed patent application is about a method for identifying and mitigating bias in deep learning models. Conventional methods lack effective techniques for bias identification and require predefined concepts and rules for bias mitigation. The embodiments of this disclosure train an auto-encoder to generate a generalized representation of an input image by decomposing it into a set of latent embeddings. These latent embeddings are used to learn the shape and color concepts of the input image. The auto-encoder is then trained to reconstruct the input image using the shape embedding modulated by the color embedding, enabling feature specialization. To identify bias, a permutation invariant neural network is trained for a classification task, and attribution scores corresponding to each concept embedding are computed. The method also includes de-biasing the classifier by training it with a set of counterfactual images generated by modifying the latent embeddings learned by the auto-encoder.

• The disclosed patent application proposes a method for identifying and mitigating bias in deep learning models.
• Conventional methods lack effective techniques for bias identification and require predefined concepts and rules for bias mitigation.
• The method involves training an auto-encoder to produce a generalized representation of an input image by decomposing it into a set of latent embeddings.
• The set of latent embeddings are used to learn the shape and color concepts of the input image.
• Feature specialization is achieved by training an auto-encoder to reconstruct the input image using the shape embedding modulated by the color embedding.
• A permutation invariant neural network is trained for a classification task to identify bias, and attribution scores corresponding to each concept embedding are computed.
• The method also includes de-biasing the classifier by training it with a set of counterfactual images generated by modifying the latent embeddings learned by the auto-encoder.

Potential Applications

• Bias identification and mitigation in deep learning models.
• Improving fairness and reducing bias in automated decision-making systems.
• Enhancing the accuracy and reliability of image recognition and classification systems.

Problems Solved

• Conventional methods lack effective techniques for identifying bias in deep learning models.
• Existing approaches require predefined concepts and rules for bias mitigation.
• Bias in automated decision-making systems can lead to unfair outcomes and discrimination.
• Deep learning models may exhibit biased behavior due to the training data used.

Benefits

• Enables the identification of bias in deep learning models without relying on predefined concepts and rules.
• Provides a method for mitigating bias by training the auto-encoder and classifier with counterfactual images.
• Enhances the fairness and accuracy of automated decision-making systems.
• Improves the reliability and trustworthiness of image recognition and classification systems.

Original Abstract Submitted

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.