Tata Consultancy Services Limited

Lovekesh Vig

Gautam Shroff

Monika Sharma

Pankaj Malhotra

Puneet Kumar Agarwal

Shubham Paliwal

Arindam Chowdhury

Rohit Rahul

Priyanka Gupta

Ashwin Srinivasan

Mayur Patidar

Ramya Sugnana Murthy Hebbalaguppe

Garima Gupta

Prerna Khurana

Mrinal Rawat

Narendhar Gugulothu

Jyoti Narwariya

Manu Sheoran

Diksha Garg

Vishal Sunder

Gaurav Gupta

Venkataramana Runkana

Rekha Singhal

Shirish Subhash Karande

Aditya Pareek

Ehtesham Hassan

Vishnu Swaroopji Masampally

Sivakumar Subramanian

Rajat Kumar

Anand Sriraman

Amit Sangroya

Rupinder Singh Virk

Gunjan Sehgal

Ankit Sharma

Sarmimala Saikia

Solomon Pushparaj Manuelraj

Ankit Verma

Vartika Tewari

Animesh

Sanket Kadarkar

Siddharth Verma

Sharod Roy Choudhury

Samarth Mittal

Arnab Chatterjee

Vishnu Tv

Ranjitha Prasad

Vishwanath Doreswamy Gowda

Vishwanath Doreswamy

Mouli Rastogi

Syed Afshan Ali

Arushi Jain

Vaibhav Varshney

Meghal Dani

Manasi Malik

Swati

Shikha Gupta

Kushagra Mahajan

Vishnu Tankasala Veparala

Swati Jindal

Ankit Bansal

Nishit Raj

Vishnu T V

Vibhor Gupta

Ramya Sugnana Hebbalaguppe

Richa Rawat

Animesh,

Mohit Yadav

Swati,

Puneet Agarwal

Rupinder Virk

Swati Swati

The present disclosure detects lesions in different datasets using a semi-supervised domain adaptation manner with very few labeled target samples. Conventional approaches suffer due to domain-gap between source and target domain. Initially, the system receives an input image, and extracts a plurality of multi-scale feature maps from the input image. Further, a classification map is generated based on the plurality of multi-scale feature maps. Further, a 4D vector corresponding to each of a plurality of foreground pixels is computed. Further, an objectness score corresponding the plurality of foreground pixels is computed. After computing the objectness score, a centerness score is computed for each of the plurality of foreground pixels using a single centerness network. Further, an updated objectness score is computed for each of the plurality of foreground. Finally, a plurality of multi-sized lesions in the input image are detected using a trained few-shot adversarial lesion detector network.

Method and system for semi-supervised domain adaptation based universal lesion detection

Existing semi-supervised and unsupervised approaches for intent discovery require an estimate of the number of new intents present in the user logs. The present disclosure receives labeled utterances from known intents and update parameters of a pre-trained language model (PLM). Representation learning and clustering is performed iteratively using labeled and unlabeled utterances from known intents and unlabeled utterances from unknown intents to fine-tune PLM and a plurality of clusters is generated. Cluster merger algorithm is executed iteratively on generated plurality of clusters. A query cluster is obtained by randomly selecting one cluster from the plurality of clusters and by obtaining a corresponding plurality of nearest neighbors based on a cosine-similarity. A response for merging the query cluster and corresponding plurality of nearest neighbors is obtained, and a new cluster is created. The corresponding cluster representation is recalculated and each of the new cluster is interpreted as an intent.

System and method for intent discovery from user logs using deep semi-supervised contrastive clustering

Recommender Systems (RS) tend to recommend more popular items instead of the relevant long-tail items. Mitigating such popularity bias is crucial to ensure that less popular but relevant items are recommended. System described herein analyses popularity bias in session-based RS obtained via deep learning (DL) models. DL models trained on historical user-item interactions in session logs (having long-tailed item-click distributions) tend to amplify popularity bias. To understand source of this bias amplification, potential sources of bias at data-generation stage (user-item interactions captured as session logs) and model training stage are considered by the system for recommendation wherein popularity of item has causal effect on user-item interactions via conformity bias, and item ranking from models via biased training process due to class imbalance. While most existing approaches address only one of these effects, a comprehensive causal inference framework is implemented by present disclosure that identifies and mitigates effects at both stages.

System, method, and non-transitory machine-readable information storage medium for recommendations of items and controlling an associated bias thereof

Session-based Recommendation (SR) is the task of recommending the next item based on previously recorded user interactions. However, most existing approaches for SR either rely on costly online interactions with real users (model-free approaches) or rely on potentially biased rule-based or data-driven user-behavior models (model-based approaches) for learning. This disclosure relates to a system and method for selecting session-based recommendation policies using historical recommendations and user feedback. Herein, the learning of recommendation policies given offline or batch data from old recommendation policies based on a Distributional Reinforcement Learning (DRL) based recommender system in the offline or batch-constrained setting without requiring access to a user-behavior model or real-interactions with the users.

System and method for training recommendation policies

Keypoint extraction is done for extracting keypoints from images of documents. Based on different keypoint extraction approaches used by existing keypoint extraction mechanisms, number of keypoints extracted and related parameters vary. Disclosed herein is a method and system for keypoint extraction from images of one or more documents. In this method, a reference image and a test image of a document are collected as input. During the keypoint extraction, based on types of characters present in words extracted from the document images, a plurality of words are extracted. Further, all connected components in each of the extracted words are identified. Further, it is decided whether keypoints are to be searched in a first component or in a last component of all the identified connected components, and accordingly searches and extracts at least four of the keypoints from the test image and the corresponding four keypoints from the reference image.

Method and system for keypoint extraction from images of documents

Several applications capture data from sensors resulting in multi-sensor time series. Existing neural networks-based approaches for such multi-sensor/multivariate time series modeling assume fixed input-dimension/number of sensors. Such approaches can struggle in practical setting where different instances of same device/equipment come with different combinations of installed sensors. In the present disclosure, neural network models are trained from such multi-sensor time series having varying input dimensionality, owing to availability/installation of different sensors subset at each source of time series. Neural network (NN) architecture is provided for zero-shot transfer learning allowing robust inference for multivariate time series with previously unseen combination of available dimensions/sensors at test time. Such combinatorial generalization is achieved by conditioning layers of core NN-based time series model with 'conditioning vector' carrying information of available sensors combination for each time series and is obtained by summarizing learned “sensor embedding vectors set” corresponding to available sensors in time series.

Neural networks for handling variable-dimensional time series data

State of the art deep network based Universal Lesion Detection (ULD) techniques inherently depend on large number of datasets for training the systems. Moreover, these system are specifically trained for lesion detection in organs of a Region of interest (RoI) of a body. Thus, requires retraining when the RoI varies. Embodiments herein disclose a method and system for domain knowledge augmented multi-head attention based robust universal lesion detection. The method utilizes minimal number of Computer Tomography (CT) scan slices to extract maximum information using organ agnostic HU windows and a convolution augmented attention module for a computationally efficient ULD with enhanced prediction performance.

Method and system for domain knowledge augmented multi-head attention based robust universal lesion detection

Existing techniques assume that all time varying covariates are confounding and thus attempts to balance a full state representation of a plurality of historical observants. The present disclosure processes a plurality of historical observants and treatment at a timestep t specific to each patient using an encoder network to a obtain a state representation s. A first set of disentangled representations comprising an outcome, a confounding and a treatment representation is learnt to predict an outcome. The first set of disentangled representations are concatenated to obtain a unified representation and the decoder network is initialized using the unified representation to obtain a state representation s. A second set of disentangled representations is learnt and concatenated to predict outcomem+1 timesteps ahead of the timestep t and proceeding iteratively until m=τ−1.

System and method for learning disentangled representations for temporal casual inference

This disclosure relates generally to method and system for handling popularity bias in item recommendations. In an embodiment the method includes initializing an item embedding look-up matrix corresponding to items in a sequence of item-clicks with respect to a training data. L2 norm is applied to the item embedding look-up matrix to learn a normalized item embeddings. Using a neural network, a session embeddings corresponding to the sequences of item-clicks is modeled and L2 norm is applied to the session embeddings to obtain a normalized session embeddings. Relevance scores corresponding to each of the plurality of items are obtained based on similarity between the normalized item embeddings and the normalized session embeddings. A multi-dimensional probability vector corresponding to the relevance scores for the items to be clicked in the sequence is obtained. A list of the items ordered based on the multi-dimensional probability vector is provided as recommendation.

System, method, and non-transitory machine readable information storage medium for handling popularity bias in item recommendations

This disclosure relates generally to intelligent visual reasoning over graphical illustrations using a MAC unit. Prior arts use visual attention to map particular words in a question to specific areas in an image to memorize the corresponding answers, thereby resulting in a limited capability to answer questions of a specific type. The present disclosure incorporates the MAC unit to enable reasoning capabilities and accordingly attend to an area in the image to find the answer. The present disclosure therefore allows generalizing over a possible set of questions with varying complexities so that an unseen question can also be answered correctly based on the reasoning methods that it has learned. The system and method of the present disclosure can be used for understanding of visual information when processing documents like business reports, research papers, consensus reports etc. containing charts and reduce the time spent in manual analysis.

Intelligent visual reasoning over graphical illustrations using a MAC unit

Traditional systems that enable extracting information from Piping and Instrumentation Diagrams (P&IDs) lack accuracy due to existing noise in the images or require a significant volume of annotated symbols for training if deep learning models that provide good accuracy are utilized. Conventional few-shot/one-shot learning approaches require a significant number of training tasks for meta-training prior. The present disclosure provides a method and system that utilizes the one-shot learning approach that enables symbol recognition using a single instance per symbol class which is represented as a graph with points (pixels) sampled along the boundaries of different symbols present in the P&ID and subsequently, utilizes a Graph Convolutional Neural Network (GCNN) or a GCNN appended to a Convolutional Neural Network (CNN) for symbol classification. Accordingly, given a clean symbol image for each symbol class, all instances of the symbol class may be recognized from noisy and crowded P&IDs.

Symbol recognition from raster images of PandIDs using a single instance per symbol class

This disclosure relates to optimizing an operation of an equipment by a neural network based optimizer is provided. The method include receiving, information associated with at least one equipment instance (j) as an input at a predefined sequence of timestamps; training, a plurality of simulation models for each equipment instance to obtain a function (f); processing, the external input parameters (e) to obtain a fixed-dimensional vector and passed as an input to obtain an vector (i); generating, a modified (i) from the output vector (i) based on a domain constraint value; computing, a reward (r) based on (i) the function (f), (ii) the modified (i), (iii) the external input parameters (e), and (iv) a reward function (R); and iteratively performing the steps of processing, generating, and computing reward (r) for a series of subsequent equipment instances after expiry of the predefined sequence of timestamps associated with a first equipment instance.

System and method for operation optimization of an equipment

In the world of digital advertising, optimally allocating an advertisement campaign within a fixed pre-defined budget for an advertising duration aimed at maximizing number of conversions is very important for an advertiser. Embodiments of present disclosure provides a robust and easily generalizable method of optimal allocation of advertisement campaign by formulating it as a constrained Markov Decision Process (MDP) defined by agent state comprising user state and advertiser state, action space comprising a plurality of ad campaigns, state transition routine and a cumulative reward model which rewards maximum total conversions in an advertising duration. The cumulative reward model is trained in conjunction with a deep Q-network for solving the MDP to optimally allocate advertisement campaign for an advertising duration within a constrained budget.

Budget constrained deep Q-network for dynamic campaign allocation in computational advertising

State of art techniques that utilize spatial association based Table structure Recognition (TSR) have limitation in selecting minimal but most informative word pairs to generate digital table representation. Embodiments herein provide a method and system for TSR from an table image via deep spatial association of words using optimal number of word pairs, analyzed by a single classifier to determine word association. The optimal number of word pairs are identified by utilizing immediate left neighbors and immediate top neighbors approach followed redundant word pair elimination, thus enabling accurate capture of structural feature of even complex table images via minimal word pairs. The reduced number of word pairs in combination with the single classifier trained to determine the word associations into classes comprising as same cell, same row, same column and unrelated, provides TSR pipeline with reduced computational complexity, consuming less resources still generating more accurate digital representation of complex tables.

Method and system for table structure recognition via deep spatial association of words

Organizations are constantly flooded with questions, ranging from mundane to the unanswerable. It is therefore respective department that actively looks for automated assistance, especially to alleviate the burden of routine, but time-consuming tasks. The embodiments of the present disclosure provide BiLSTM-Siamese Network based Classifier for identifying target class of queries and providing responses to queries pertaining to the identified target class, which acts as an automated assistant that alleviates burden of answering queries in well-defined domains. Siamese Model (SM) is trained for a epochs, and then the same Base-Network is used to train Classification Model (CM) for b epochs iteratively until best accuracy is observed on validation test, wherein SM ensures it learns which sentences are similar/dissimilar semantically while CM learns to predict target class of every user query. Here a and b are assumed to be hyper parameters and are tuned for best performance on the validation set.

BiLSTM-siamese network based classifier for identifying target class of queries and providing responses thereof

Growing community of inventors

Gurgaon, India

Lovekesh Vig