Tata Consultancy Services Limited

Swarnava Dey

Arijit Mukherjee

Arpan Pal

Himadri Sekhar Paul

Swagata Biswas

Balamuralidhar Purushothaman

Arijit Ukil

Ansuman Banerjee

Pubali Datta

Jeet Dutta

Jayeeta Mondal

Avik Ghose

Soma Bandyopadhyay

Chirabrata Bhaumik

Abhijan Bhattacharyya

Sanjay Madhukar Kimbahune

Snehasis Banerjee

Amit Swain

Devraj Goulikar

Ashish Omprakash Indani

Ishan Sahu

Gitesh Kulkarni

Aakash Tyagi

Sujit Raghunath Shinde

Ja Yeeta Mondal

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. Method of the present disclosure is an iterative approach to obtain a set of partitions by repeatedly refining the partitions and generating a cascaded model for inference and training on ensemble hardware.

Combining compression, partitioning and quantization of DL models for fitment in hardware processors

Embodiments of the present disclosure provide a method and system for co-operative and cascaded inference on the edge device using an integrated Deep Learning (DL) model for object detection and localization, which comprises a strong classifier trained on largely available datasets and a weak localizer trained on scarcely available datasets, and work in coordination to first detect object (fire) in every input frame using the classifier, and then trigger a localizer only for the frames that are classified as fire frames. The classifier and the localizer of the integrated DL model are jointly trained using Multitask Learning approach. Works in literature hardly address the technical challenge of embedding such integrated DL model to be deployed on edge devices. The method provides an optimal hardware software partitioning approach for components or segments of the integrated DL model which achieves a tradeoff between latency and accuracy in object classification and localization.

Integrated deep learning model for co-operative and cascaded inference on edge utilizing classifiers for input images

Automatic speech recognition techniques are implemented in resource constrained devices such as edge devices in internet of things where on-device speech recognition is required for low latency and privacy preservation. Existing neural network models for speech recognition have a large size and are not suitable for deployment in such devices. The present disclosure provides an architecture of a size constrained neural network and a method of training the size constrained neural network. The architecture of the size constrained neural network provides a way of increasing or decreasing number of feature blocks to achieve an accuracy-model size trade off. The method of training the size constrained neural network comprises creating a training dataset with short utterances and training the size constrained neural network with the training dataset to learn short term dependencies in the utterances. The trained size constrained neural network model is suitable for deployment in resource constrained devices.

Method and system for automatic speech recognition in resource constrained devices

A growing need for inferencing to be run on fog devices exists, in order to reduce the upstream network traffic. However, being computationally constrained in nature, executing complex deep inferencing models on such devices has been proved difficult. A system and method for partitioning of deep convolution neural network for execution of computationally constraint devices at a network edge has been provided. The system is configured to use depth wise input partitioning of convolutional operations in deep convolutional neural network (DCNN). The convolution operation is performed based on an input filter depth and number of filters for determining the appropriate parameters for partitioning based on an inference speedup method. The system uses a master-slave network for partitioning the input. The system is configured to address these problems by depth wise partitioning of input which ensures speedup inference of convolution operations by reducing pixel overlaps.

Method and system for partitioning of deep convolution network for executing on computationally constraint devices

A system and method for offloading scalable robotic tasks in a mobile robotics framework. The system comprises a cluster of mobile robots and they are connected with a back-end cluster infrastructure. It receives scalable robotic tasks at a mobile robot of the cluster. The scalable robotics tasks include building a map of an unknown environment by using the mobile robot, navigating the environment using the map and localizing the mobile robot on the map. Therefore, the system estimate the map of an unknown environment and at the same time it localizes the mobile robot on the map. Further, the system analyzes the scalable robotics tasks based on computation, communication load and energy usage of each scalable robotic task. And finally the system priorities the scalable robotic tasks to minimize the execution time of the tasks and partitioning the SLAM with computation offloading in edge network and mobile cloud server setup.

System and method for offloading robotic functions to network edge augmented clouds

A data driven approach for fault detection in robotic actuation is disclosed. Here, a set of robotic tasks are received and analyzed by a Deep Learning (DL) analytics. The DL analytics includes a stateful (Long Short Term Memory) LSTM. Initially, the stateful LSTM is trained to match a set of activities associated with the robots based on a set of tasks gathered from the robots in a multi robot environment. Here, the stateful LSTM utilizes a master slave framework based load distribution technique and a probabilistic trellis approach to predict a next activity associated with the robot with minimum latency and increased accuracy. Further, the predicted next activity is compared with an actual activity of the robot to identify any faults associated robotic actuation.

System and method for fault detection in robotic actuation

This disclosure relates generally to the use of distributed system for computation, and more particularly, relates to a method and system for optimizing computation and communication resource while preserving security in the distributed device for computation. In one embodiment, a system and method of utilizing plurality of constrained edge devices for distributed computation is disclosed. The system enables integration of the edge devices like residential gateways and smart phone into a grid of distributed computation. The edged devices with constrained bandwidth, energy, computation capabilities and combination thereof are optimized dynamically based on condition of communication network. The system further enables scheduling and segregation of data, to be analyzed, between the edge devices. The system may further be configured to preserve privacy associated with the data while sharing the data between the plurality of devices during computation.

System and method for distributed computation using heterogeneous computing nodes

Cloud robotics infrastructures generally support heterogeneous services that are offered by heterogeneous resources whose reliability or availability also varies widely with varying lifetime. For such systems, defining a static redundancy configuration for all services is difficult and often biased. Also, it is not feasible to define a redundancy configuration separately for each unique service. Therefore, in the present disclosure a trade-off between the two is ensured by providing At-most M-Modular Flexible Redundancy Model wherein an exact degree of redundancy is defined and is given to each service in a heterogeneous service environment and monitoring each task and subtask status to ensure that each subtask gets accomplished thereby enabling the tuning of the tradeoff between redundancy and cost and determining efficiency of the system by estimating number of resources utilized to complete specific subtask and comparing the resources utilization with the exact degree of redundancy defined.

Resources management in internet of robotic things (IoRT) environments

Signal analysis is applied in various industries and medical field. In signal analysis, wavelet analysis plays an important role. The wavelet analysis needs to identify a mother wavelet associated with an input signal. However, identifying the mother wavelet associated with the input signal in an automatic way is challenging. Systems and methods of the present disclosure provides signal analysis with automatic selection of wavelets associated with the input signal. The method provided in the present disclosure receives the input signal and a set of parameters associated with the signal. Further, the input signal is analyzed converted into waveform. The waveforms are analyzed to provide image units. Further, the image units are processed by a plurality of deep architectures. The deep architectures provides a set of comparison scores and a matching wavelet family is determined by utilizing the set of comparison scores.

System and method for signal analysis

In current distributed simultaneous localization and mapping (SLAM) implementations on multiple robots in a robotic cluster, failure of a leader robot terminates a map building process between multiple robots. Therefore, a technique for fault-tolerant SLAM in robotic clusters is disclosed. In this technique, robotic localization and mapping SLAM is executed in a resource constrained robotic cluster such that the distributed SLAM is executed in a reliable fashion and self-healed in case of failure of the leader robot. To ensure fault tolerance, the robots are enabled, by time series analysis, to find their individual failure probabilities and use that to enhance cluster reliability in a distributed manner.

System and method for executing fault-tolerant simultaneous localization and mapping in robotic clusters

A method for data partitioning in an internet-of-things (IoT) network is described. The method includes determining number of computing nodes in the IoT network capable of contributing in processing of a data set. At least one capacity parameter associated with each computing node in the IoT network and each communication link between a computing node and a data analytics system can be ascertained. The capacity parameter can indicate a computational capacity for each computing node and communication capacity for each communication link. An availability status, indicating temporal availability, of each of computing nodes and each communication link is determined. The data set is partitioned into subsets, based on the number of computing nodes, the capacity parameter and the availability status, for parallel processing of the subsets.

Data partitioning in internet-of-things (IOT) network

Methods and devices for controlling execution of a data analytics application on a computing device are described. The devices include an alert app to prompt a user on system load and to recommend the user for proactively controlling the execution of a set of processes to reclaim computational resources required for execution of the data analytics application on the devices.

User driven smartphone scheduling enhancement for running data analytics application

A method comprises, receiving, at each of a plurality of computing devices, a task execution estimation request message from a central server, the task execution estimation request message comprising a worst-case execution time (WCET) corresponding to the computing device. The method further comprises, computing, by each of the plurality of computing devices, an estimate task execution time for the task based on the WCET and a state transition model corresponding to the computing device, wherein the state transition model indicates available processing resources corresponding to the computing device. Further, the method comprises transmitting, by each of the plurality of computing devices, the estimate task execution time to the central server for allocation of the task to a computing device from amongst the plurality of computing devices based on the estimate task execution time corresponding to the computing device.

Task allocation in a computing environment

Described herein, are methods and devices for execution of a task in a grid computing system. According to an implementation, free time-slots are identified and durations of the free time-slots are estimated, by an edge device, for execution of a sub-task. The free time-slots are indicative of an idle state of the edge device. At least one computation capability parameter of the edge device is determined by the edge device for execution of a sub-task during the free time-slots. An advertisement profile having at least one free time-slot, and the duration and the at least one computation capability parameter associated with the at least one free time-slot is created by the edge device. The advertisement profile is provided by the edge device to grid servers in the grid computing system for partitioning a main task to create a sub-task executable by the edge device.

Growing community of inventors

Kolkata, India

Swarnava Dey