Abstract: Conventionally, applying analytics on dataset is the scarcity of labelled data. With increase of data there is cost fact effecting nature of servicing required for data (e.g., cost in terms of resource and time and effort is high for data annotation). Though data is analysed, it may be prone to error. Present disclosure provides systems/methods for reducing volume of data to be annotated for time series data thereby reducing time and effort of resources, thus resulting in effective utilization of system's resources (e.g., memory, processor, etc.). More specifically, the method of the present disclosure adaptively modifies the volume of the data to be annotated based on the performance of the unsupervised learning method applied in the system. Moreover, in the absence of an annotation mechanism for clusters of time series data, meta data associated with the time series data is utilized for annotation and validation of dataset.

Abstract: Current approaches for atrial fibrillation (AF) detection use deep learning models which remain opaque. In particular, they lack in providing explanation of why this particular decision (around existence of AF) has been made, thereby making it unacceptable to clinical domain experts. Present disclosure provides method and system for explaining decision-making process of deep learning models used for detecting AF in ECG waves. The system receives ECG signal which is converted into two-dimensional (2D) representation which further helps in classification of diagnosis condition from ECG signal using classifier model. Thereafter, system generates class activation maps (CAM) to find attention scores and finally uses these attention scores, to identify top R-R intervals where classifier model is placing greater emphasis. Further, system converts ECG image into ECG signal which also converts CAM into attention wave.

Abstract: The present invention generally relates to the field of deep learning, and, more particularly, to a method and system for task agnostic distillation in foundation models. Conventional distillation methods are not scalable and also does not handle mismatches between embedding sizes of teacher and student models. Thus, embodiments of present disclosure first transforms the teacher model in such a way that its embedding size matches that of the student model. This is done by augmenting a linear layer having dimensions equal to the embedding size of the student model and a projector network to the teacher model. Then the augmented layers are trained using a self-supervised learning technique by freezing the teacher model. Finally, the projector network is discarded to obtain a transformed teacher model. The student model is then trained using transformed teacher model by performing knowledge distillation based on similarity loss.

Abstract: This disclosure relates generally to method and system for an adaptive filter based learning model for time series sensor signal classification on edge devices. The adaptive filter based learning model for time series sensor signal classification enables automated-computationally lightweight learning (significant reduction in computational resources) and inferring/classification in real-time or near-real-time on CPU/memory/battery life constrained edge devices. The disclosed techniques for time series sensor signal classification on edge devices characterizes the intrinsic signal processing properties of the input time series sensor signals using linear adaptive filtering and derivative spectrum to efficiently construct the adaptive filter based learning model based on standard classification algorithms for time series sensor signal classification.

Abstract: Small and compact Deep Learning models are required for embedded AI in several domains. In many industrial use-cases, there are requirements to transform already trained models to ensemble embedded systems or re-train those for a given deployment scenario, with limited data for transfer learning. Moreover, the hardware platforms used in embedded application include FPGAs, AI hardware accelerators, System-on-Chips and on-premises computing elements (Fog/Network Edge). These are interconnected through heterogenous bus/network with different capacities. Method of the present disclosure finds how to automatically partition a given DNN into ensemble devices, considering the effect of accuracy—latency power—tradeoff, due to intermediate compression and effect of quantization due to conversion to AI accelerator SDKs.

Abstract: The present invention generally relates to the field of deep learning, and, more particularly, to a method and system for deep neural network model size reduction by layer and filter elimination. Conventional methods result in zero weight parameters but does not effectively reduce the size of the models. Thus, the method of present disclosure imbibes capability of effective discovery of sparsified layers and systematically generates smaller size models close to the performance of base model. Further, the disclosed method gains regularization effect due to the removal of unnecessary parameters. The smaller size model reduces computational burden, energy consumption, and latency along with satisfying smaller memory requirements and thus, it leads to the possibility of deploying at different edge devices that provides the opportunity to develop niche and important applications in different domains including automobiles, robotics, healthcare.

Abstract: This disclosure relates generally to a system and a method for mitigating generalization loss in deep neural network for time series classification. In an embodiment, the disclosed method includes compute an entropy of a timeseries training dataset, and a mean and a variance of the entropy and a regularization factor is computed. A plurality of iterations are performed to dynamically adjust the learning rate of the deep Neural Network (DNN) using a Mod-Adam optimization, and obtain a network parameter, and based on the network parameter, the regularization factor is updated to obtain an updated regularized factor. The learning rate is adjusted in the plurality of iterations by repeatedly updating the network parameter based on a variation of a generalization loss during the plurality of iterations. The updated regularized factor of the current iteration is used for adjusting the learning rate in a subsequent iteration of the plurality of iterations.

Abstract: This disclosure relates generally to multi-class multi-label classification and more particularly to contradiction avoided learning for multi-class multi-label classification. Conventional classification methods do not consider contradictory outcomes in multi-label classification tasks wherein contradictory outcomes have significant negative impact in the classification problem solution. The present disclosure provides a contradiction avoided learning multi-class multi-label classification. The disclosed method utilizes a binary contradiction matrix constructed using domain knowledge. Based on the binary contradiction matrix the training dataset is divided into two parts, one comprising contradictions and the second without contradictions. The classification model is trained using the divided datasets using a contradiction loss and a binary cross entropy loss to avoid contradictions during learning of the classification model.

Abstract: This disclosure relates generally to multi-class multi-label classification and more particularly to contradiction avoided learning for multi-class multi-label classification. Conventional classification methods do not consider contradictory outcomes in multi-label classification tasks wherein contradictory outcomes have significant negative impact in the classification problem solution. The present disclosure provides a contradiction avoided learning multi-class multi-label classification. The disclosed method utilizes a binary contradiction matrix constructed using domain knowledge. Based on the binary contradiction matrix the training dataset is divided into two parts, one comprising contradictions and the second without contradictions. The classification model is trained using the divided datasets using a contradiction loss and a binary cross entropy loss to avoid contradictions during learning of the classification model.

Abstract: The present invention relates to the field of evaluating clinical diagnostic models. Conventional metrics does not consider context dependent clinical principles and is unable to capture critically important features that ought to be present in a diagnostic model. Thus, present disclosure provides a method and system for evaluating clinical efficacy of multi-label multi-class computational diagnostic models. Diagnosis for a given dataset of diagnostic samples is obtained from the diagnostic model which is then classified as wrong, missed, over or right diagnosis, based on which a first penalty is calculated. A second penalty is calculated for each diagnostic sample using a contradiction matrix. The first and second penalties are summed up to compute a pre-score for each diagnostic sample. Finally, the diagnostic model is evaluated using a metric that is based on sum of pre-scores, and scores from a perfect and a null multi-label multi-class computational diagnostic model.

Abstract: Portable ECG monitors available in market have the disadvantage that the ECG data they provide as input aren't directly interpretable and requires medical knowledge for the users. The disclosure herein generally relates to Electrocardiogram (ECG), and, more particularly, to a method and system for generating 2d representation of electrocardiogram (ECG) signals. The system provides a mechanism for determining variability between a plurality of segments of an ECG data measured, and uses the information on the determined variability to generate the 2D representation corresponding to the ECG signal. The system further provides means to generate a data model that can be further used for processing real-time ECG data for generating corresponding interpretations. This allows a user to obtain the interpretations as output.

Abstract: This disclosure relates generally to methods and systems for unobtrusive digital health assessment of high risk subjects, wherein bio-markers pertaining to a disease are identified automatically using physical activity and physiology monitoring on a continuous basis. Identification of bio-markers in the medical domain is conventionally dependent on insights derived from medical tests which are obtrusive in nature. Systems and methods of the present disclosure integrate physical characteristics, lifestyle habits and prevailing medical conditions with monitored physical activities and physiological measurements to assess health of high risk subjects. Systems and methods of the present disclosure also enable automatic generation of control class and treatment class that may be effectively used for health assessment.

Abstract: This disclosure relates generally to physiological monitoring, and more particularly to feature set optimization for classification of physiological signal. In one embodiment, a method for physiological monitoring includes identifying clean physiological signal training set from an input physiological signal based on a Dynamic Time Warping (DTW) of segments associated with the physiological signal. An optimal features set is extracted from a clean physiological signal training set based on a Maximum Consistency and Maximum Dominance (MCMD) property associated with the optimal feature set that strictly optimizes on the objective function, the conditional likelihood maximization over different selection criteria such that diverse properties of different selection parameters are captured and achieves Pareto-optimality. The input physiological signal is classified into normal signal components and abnormal signal components using the optimal features set.

Abstract: In many real-life applications, ample amount of examples from one class are present while examples from other classes are rare for training and learning purposes leading to class imbalance problem and misclassification. Methods and systems of the present disclosure facilitate generation of an extended synthetic rare class super dataset that is further pruned to obtain a synthetic rare class dataset by maximizing similarity and diversity in the synthetic rare class dataset while preserving morphological identity with labeled rare class training dataset. Oversampling methods used in the art result in cloning of datasets and do not provide the needed diversity. The methods of the present disclosure can be applied to classification of noisy phonocardiogram (PCG) signals among other applications.

Abstract: A method and system for a feature subset-classifier pair for a classification task. The classification task corresponds to automatically classifying data associated with a subject(s) or object(s) of interest into an appropriate class based on a feature subset selected among a plurality of features extracted from the data and a classifier selected from a set of classifier types. The method proposed includes simultaneously determining the feature subset-classifier pair based on a relax-greedy {feature subset, classifier} approach utilizing sub-greedy search process based on a patience function, wherein the feature subset-classifier pair provides an optimal combination for more accurate classification. The automatic joint selection is time efficient solution, effectively speeding up the classification task.

Abstract: Small and compact Deep Learning models are required for embedded Al in several domains. In many industrial use-cases, there are requirements to transform already trained models to ensemble embedded systems or re-train those for a given deployment scenario, with limited data for transfer learning. Moreover, the hardware platforms used in embedded application include FPGAs, AI hardware accelerators, System-on-Chips and on-premises computing elements (Fog/Network Edge). These are interconnected through heterogenous bus/network with different capacities. Method of the present disclosure finds how to automatically partition a given DNN into ensemble devices, considering the effect of accuracy—latency power—tradeoff, due to intermediate compression and effect of quantization due to conversion to AI accelerator SDKs.

Abstract: Systems and methods for detecting an anomaly in a cardiovascular signal using hierarchical extremas and repetitions. The traditional systems and methods provide for some anomaly detection in the cardiovascular signal but do not consider the discrete nature and strict rising and falling patterns of the cardiovascular signal and frequency in terms of hierarchical maxima points and minima points. Embodiments of the present disclosure provide for detecting the anomaly in the cardiovascular signal using hierarchical extremas and repetitions by smoothening the cardiovascular signal, deriving sets of hierarchical extremas using window detection, identifying signal patterns based upon the sets of hierarchical extremas, identifying repetitions in the signal patterns based upon occurrences and randomness of occurrences of the signal patterns and classifying the cardiovascular signal as anomalous and non-anomalous for detecting the anomaly in the cardiovascular signal.

Abstract: Conventionally, applying analytics on dataset is the scarcity of labelled data. With increase of data there is cost fact effecting nature of servicing required for data (e.g., cost in terms of resource and time and effort is high for data annotation). Though data is analysed, it may be prone to error. Present disclosure provides systems/methods for reducing volume of data to be annotated for time series data thereby reducing time and effort of resources, thus resulting in effective utilization of system's resources (e.g., memory, processor, etc.). More specifically, the method of the present disclosure adaptively modifies the volume of the data to be annotated based on the performance of the unsupervised learning method applied in the system. Moreover, in the absence of an annotation mechanism for clusters of time series data, meta data associated with the time series data is utilized for annotation and validation of dataset.

Abstract: The present disclosure addresses the technical problem of information loss while representing a physiological signal in the form of symbols and for recognizing patterns inside the signal. Thus making it difficult to retain or extract any relevant information which can be used to detect anomalies in the signal. A system and method for anomaly detection and discovering pattern in a signal using morphology aware symbolic representation has been provided. The system discovers pattern atoms based on the strictly increasing and strictly decreasing characteristics of the time series physiological signal, and generate symbolic representation in terms of these pattern atoms. Additionally the method possess more generalization capability in terms of granularity. This detects discord/abnormal phenomena with consistency.

Abstract: This disclosure relates generally to a system and a method for mitigating generalization loss in deep neural network for time series classification. In an embodiment, the disclosed method includes compute an entropy of a timeseries training dataset, and a mean and a variance of the entropy and a regularization factor is computed. A plurality of iterations are performed to dynamically adjust the learning rate of the deep Neural Network (DNN) using a Mod-Adam optimization, and obtain a network parameter, and based on the network parameter, the regularization factor is updated to obtain an updated regularized factor. The learning rate is adjusted in the plurality of iterations by repeatedly updating the network parameter based on a variation of a generalization loss during the plurality of iterations. The updated regularized factor of the current iteration is used for adjusting the learning rate in a subsequent iteration of the plurality of iterations.