Deep Neural Networks (DNNs) have significantly enhanced the baseline performance of Human Activity Recognition (HAR) models by learning patterns directly from raw sensor signals. For wearables, on-device inference is crucial for preserving personally identifiable information and ensuring long battery life, particularly in HAR-based health and wellness applications. However, existing DNN models for HAR are often either too large for wearable devices, designed for simple binary activity classification, or rely on manually crafted features. To address these challenges, we propose a scalable solution comprising a tunable tiny attention condenser-based architecture and a Deep Q-Learning-based tuner. Together, these components enable the generation of compact, wearable-friendly DNN models for complex activity recognition tasks. Experimental results demonstrate that the proposed methodology achieves state-of-the-art accuracy on the HAR-Box and UCI-HAR datasets, with model sizes under 1 MB, making it suitable for resource-constrained devices.

Deep learning models for edge deployments must be small, efficient, and robust. Knowledge distillation from large foundation models (FMs) can help, as they capture rich, transferable representations from large, multimodal datasets. However, two key challenges hinder this process: (1) extreme model compression based on a test dataset often leads to brittleness under distribution shifts, and (2) the distribution gap between an FM’s training data and a small model’s target dataset makes standard distillation methods ineffective.We propose a tunable dynamic loss curriculum for knowledge distillation to address these issues. Experiments show that small encoders trained with our approach achieve balanced transfer learning performance across both primary and out-of-distribution tasks. For instance, a tiny GPT-like model effectively transfers to sentiment classification and language modeling. Likewise, a tiny ResNet trained on CIFAR-10 achieves 10% higher accuracy on corrupted CIFAR-10 than state-of-the-art baselines for tiny models. While it trails dedicated robustness training by 4%, our method ensures superior adaptability across diverse datasets and tasks.

Large language models (LLMs) with retrieval augmented-generation (RAG) have been the optimal choice for scalable generative AI solutions in the recent past. Although RAG implemented with AI agents (agentic-RAG) has been recently popularized, its suffers from unstable cost and unreliable performances for Enterprise-level data-practices. Most existing use-cases that incorporate RAG with LLMs have been either generic or extremely domain specific, thereby questioning the scalability and generalizability of RAG-LLM approaches. In this work, we propose a unique LLM-based system where multiple LLMs can be invoked to enable data authentication, user-query routing, data-retrieval and custom prompting for question-answering capabilities from Enterprise-level data tables on sustainability. The source tables here are highly fluctuating and large in size storing carbon footprint, energy and water usage at buildings in regional levels globally and the proposed framework enables structured responses in under 10 seconds per query. Additionally, we propose a five metric scoring module that detects and reports hallucinations in the LLM responses. Our proposed system and scoring metrics achieve >90% confidence scores across hundreds of user queries in the sustainability, financial health and social media domains. Extensions to the proposed extreme RAG architectures can enable heterogeneous source querying using LLMs.

Large Language Models (LLMs) have been applied to build several automation and personalized question-answering prototypes so far. However, scaling such prototypes to robust products with minimized hallucinations or fake responses still remains an open challenge, especially in niche data-table heavy domains such as financial decision making. In this work, we present a novel Langchain-based framework that transforms data tables into hierarchical textual "data chunks" to enable a wide variety of actionable question answering. First, the user-queries are classified by intention followed by automated retrieval of the most relevant data chunks to generate customized LLM prompts per query. Next, the custom prompts and their responses undergo multi-metric scoring to assess for hallucinations and response confidence. The proposed system is optimized with user-query intention classification, advanced prompting, data scaling capabilities and it achieves over $ 90\%$ confidence scores for a variety of user-queries responses ranging from {What, Where, Why, How, predict, trend, anomalies, exceptions} that are crucial for financial decision making applications. The proposed data to answers framework can be extended to other analytical domains such as sales and payroll to ensure optimal hallucination control guardrails.

Emergence of small satellites for earth observation missions has opened up new horizons for space research but at the same time posed newer challenges of limited power and compute resource arising out of the size & weight constraints imposed by these satellites. The currently evolving neuromorphic computing paradigm shows promise in terms of energy efficiency and may possibly be exploited here. In this paper, we try to prove the applicability of neuromorphic computing for on-board data processing in satellites by creating a 2-stage hierarchical cloud cover detection application for multi-spectral earth observation images. We design and train a CNN and convert it into SNN using the CNN2SNN conversion toolkit of Brainchip Akida neuromorphic platform. We achieve 95.46% accuracy while power consumption and latency are at least 35x and 3.4x more efficient respectively in stage-1 (and 230x & 7x in stage-2) compared to the equivalent CNN running on Jetson TX2.

The emerging trend of small satellites for earth observation missions has enabled commercial organisations to exploit the horizon for various business applications related to weather forecasting/monitoring, Land Use Land Cover (LULC) classifications, disaster (such as oil spill or forest fire) monitoring etc. However, the limited power and computational capacity of these small satellites arising out of the size and weight restrictions have posed newer challenges, primarily related to low-power on-board data processing and transmission. One possible approach is to harness the capabilities of the evolving neuromorphic computing paradigm for such low-power computing requirements. One possible application can be lossless compression of high-resolution earth observation images before sending those downstream to ground stations for further analysis. In this paper, we propose a novel method of lossless image compression based on classical Arithmetic Encoding that exploits the low power computing capability of Spiking Neural Networks and neuromorphic platforms. We experimentally prove that our SNN approach achieves compression ratio at-par with state of the art ANN methods with an estimated 2.5× power efficiency and 50% lower latency with a much smaller model - thereby enabling on-board image compression and at the same time, saving on a corresponding amount of energy during transmission.

In recent times, in many industrial domains, the need for in-situ processing of time series data from sensors have grown extensively, to ensure low-latency real-time responses, making embedded edge AI an important area of work. Most current techniques, being computationally intensive, are not suited for edge implementation; and in real world scenarios, connectivity to the cloud for processing sensed data leads to higher latency and often less reliability. Neuromorphic systems, coupled with spiking neural networks (SNN) offer a solution to such problems. This paper explores this evolving paradigm to address the need for low-footprint efficient time series classifiers implementable on edge, and targeted for predictive maintenance scenarios. A reservoir-based SNN architecture is designed and tried for classification of different vibration time series datasets. While the system is found to obtain at par classification accuracy for each of the datasets compared to prior arts, it is also observed to be more efficient in terms of synaptic operations per timestep (13% to 38%) using Gaussian temporal spike encoding scheme compared to Poisson rate encoding. Moreover, the system is found to be robust with respect to learning with reduction in training data (upto 20%).

Fire-related incidents continue to be reported as a leading cause of life and property destruction. Automated fire detection and localization (AFDL) systems have grown in importance with the evolution of applied robotics, especially because use of robots in disaster situations can lead to avoidance of human fatality. The importance of AFDL on resource-constrained devices has further grown, as most unmanned vehicles (drones or ground vehicles) are battery operated with limited computational capacity, the disaster situations cannot guarantee uninterrupted communication with high-end resources in the cloud, and yet faster response time is a prime necessity. Traditional computer vision–based techniques require hand-engineered features on a case-by-case basis. Deep Learning –based classifiers perform well for fire/no-fire classification due to the availability of large datasets for training; however, a dearth of good fire localization datasets renders the localization performance below par. We have tried to address both problems with a multi-task learned cascaded model that triggers localization workflow only if the presence of fire is detected, through a strong classifier trained on available large fire datasets. This presents only fire images to a relatively weaker localization model, reducing false positives, false negatives, and thereby improving overall AFDL accuracy. The multi-task learning (MTL) approach for end-to-end training of a stitched classifier and object localizer model on diverse datasets enabled us to build a strong fire classifier and feature extractor. It also resulted in a single unified model, capable of running on “on-board” compute infrastructure without compromising on accuracy. To achieve the target inference rate for the AFDL deployment, we have investigated the effect of quantization and compression due to hardware acceleration on an MTL model. This article presents an approach to automate the hardware-software co-design to find the optimum parameter partitioning for a given MTL problem, especially when some parts of the model are hardware accelerated. We present combined evaluation results showing that our methodology and the corresponding AFDL model strikes a balance between the frames inferred per second and several accuracy metrics. We report fire localization accuracy in terms of mean average precision (object detection), which has not been done earlier for embedded AFDL systems.

In the past two decades, radar-based human sensing has become a topic of intense research. Unlike vision-based techniques which require the use of camera, radars are unobtrusive and privacy preserving in nature. Further, radars are agnostic of the lighting conditions and can be used for through-the-wall imaging thereby making them hugely effective in many situations. Compact, affordable radars have been designed that can be easily integrated with remote monitoring systems. However, the classical machine learning techniques currently used for learning and inferring human actions from radar images are compute intensive, and require large volume of training data, making them unsuitable for deployment on the network edge. In this paper, we propose to use the concepts of neuromorphic computing and Spiking Neural Networks (SNN) to learn human actions from data captured by the radar. To the best our knowledge, this is the first attempt of using SNNs on micro-Doppler data from radars. Our SNN model is capable of learning spatial as well as temporal features from the data and our experiments have resulted in 85% accuracy which is comparable with the classical machine learning approaches that are typically used on similar data. Further, the use of neuromorphic and SNN concepts make our model deployable over evolving neuromorphic edge devices thereby making the entire approach more efficient in terms of data, computation and energy consumption.

Mammalian neural circuits respond to different sensory stimuli by firing spikes at particular times. Closely mimicking this phenomenon, the evolving 3rd generation neural networks, known as Spiking Neural Networks (SNNs), are found to be capable of memorizing and learning from the spatio-temporal spike patterns. This makes SNN applicable in identification of human actions and gestures, especially in the robotics domain. The paradigm is also suited for Neuromorphic Systems leading to less energy intensive applications. In this work, we present a novel spiking neural network constituting multiple convolutional layers and a reservoir layer to extract spatial and temporal features respectively from human gesture videos captured with DVS camera. We achieved more than 95% Top-3 accuracy on IBM DVS dataset and we claim that the performance of our network is better in terms of accuracy vs. learning parameters ratio when compared to other networks.

Automatic question generation (QG) is a challenging problem in natural language understanding.QG systems are typically built assuming access to a large number of training instances where each instance is a question and its corresponding answer.For a new language, such training instances are hard to obtain making the QG problem even more challenging.Using this as our motivation, we study the reuse of an available large QG dataset in a secondary language (e.g.English) to learn a QG model for a primary language (e.g.Hindi) of interest.For the primary language, we assume access to a large amount of monolingual text but only a small QG dataset.We propose a cross-lingual QG model which uses the following training regime: (i) Unsupervised pretraining of language models in both primary and secondary languages and (ii) joint supervised training for QG in both languages.We demonstrate the efficacy of our proposed approach using two different primary languages, Hindi and Chinese.We also create and release a new question answering dataset for Hindi consisting of 6555 sentences.

Running deep learning algorithms at the edge is a necessity in many industrial use-cases, especially in applications that use robots and drones in disaster recovery, surveillance, oil & gas operations etc. Current state of the art deep learning algorithms are extremely efficient in analysing image, audio, video and other time-series signals. However, their performance degrades considerably on constrained edge devices. In this demo, we show how standard pretrained CNN (Convolutional Neural Network) models can be partitioned for efficient parallel execution between constrained devices and also achieve real-time response.

The current trend in industry to augment lightweight and inexpensive robots with cloud-based distributed intelligence for executing complex, collaborative tasks has led to the emergence of fields like Internet of Robotic Things (IoRT) and Cloud Robotics (CR). However, reliable execution of orchestrated jobs in a networked setup with chances of infrastructural failures remains a concern. The only perceivable solution is resource redundancy, where the traditional approaches are not cost effective for tightly budgeted IoRT/CR deployments. In this work, an agile redundancy model, At most M, is proposed, which provides a handle to tune the trade-off between resource cost and reliability. The model is evaluated in a simulated warehouse environment with robots, drones, AGVs and private cloud servers deployed to accomplish multiple pickup and delivery tasks. The benefits w.r.t. resource usage cost by using our optimized redundancy model is illustrated and the trade-off between cost and reliability is demonstrated.

Industry 4.0 deployments involve machines, robots, Internet of Things, business processes and human participants coordinating in time constrained and safety critical environments. As these deployments make use of complex workflow patterns, accurate formal modeling, verification and timing analysis of such industrial systems are needed. In this paper, we model the workflow interactions of Industry 4.0 warehouse operations using the concurrent programming language Orc. Complex deployments involving multiple robotic agents and business processes further require analysis of correctness, liveness and safety properties. In order to verify the workflows, the Orc specifications are translated into Workflow net representations, with verification done using the TAPAAL model checker. Additional composition of timing constraints are analyzed to enable hard robotic deadlines to interact with best effort Service Level Agreement (SLA) requirements. We demonstrate these aspects over a realistic use case in automated warehouse management with pick and delivery robots.

Dynamic or demand-driven service deployment in a Grid or Cloud environment is an important issue considering the varying nature of demand. Most distributed frameworks either offer static service deployment which results in resource allocation problems, or, are job-based where for each invocation, the job along with the data has to be transferred for remote execution resulting in increased communication cost. An alternative approach is dynamic demand-driven provisioning of services as proposed in earlier literature, but the proposed methods fail to account for the volatility of resources in a Grid environment. In this paper, we propose a unique peer-to-peer based approach for dynamic service provisioning which incorporates a Bit-Torrent like protocol for provisioning the service on a remote node. Being built around a P2P model, the proposed framework caters to resource volatility and also incurs lower provisioning cost.

Developing analytical applications for IoT based on sensor signal processing tends to be complicated as applications are executed as sequence of steps comprising of multiple alternative algorithms, including suitable feature extraction modules depending on the goal of the application. Experience shows that developers spend considerable time and effort in performing feature extraction and dimensionality reduction. In this paper we propose a framework based on a relevant case study which allows developers to drag and drop algorithms to create a workflow chain, automatically select the most relevant signal features for the particular analytic application using a training data set to generate a model and deploy the model for use. The method reduces the effort and cost of development which is deemed highly important for the analytics industry.

In a mobile grid computing framework where mobile devices are used as computing resources, minimizing the task offloading time remains an important issue. A task is an independent unit of execution consisting of a input data volume for execution and optionally a target-specific executable. We consider a mobile grid infrastructure where mobile devices are connected via Wi-Fi network and the grid infrastructure has a set of tasks (i.e. a set of data volumes) to be transferred to a subset of the mobile devices. In a Wi-Fi network, mobile devices usually associate themselves to the access points (APs) having the strongest radio signal. In this paper, we address the problem of AP activation (by frequency assignment) and association of AP with devices in the context of minimizing the overall data-transfer completion time. We present a constraint based formulation and also a heuristic as solutions. Simulations results are presented which contrast our proposed methods with some of the earlier works.

In recent times, many human-centric applications are being developed to leverage the diverse range of sensors present in Android smartphones. Smartphones are also being used as edge network gateways for fusing data from multiple sensors. These applications often run as background services and usuall

Smartphones are currently being used to develop diverse range of applications (apps) involving sensors. These apps generally acquire and analyze sensor data and are usually implemented as background services. The importance values of Android processes are in a hierarchy of foreground, visible, background etc. in decreasing order of importance. Whenever a new process arrives, it may necessitate removal of old and less important processes for reclaiming memory. Current smartphones do not provide any options through which user's idea of priority can override that of the system defaults. In this work we present an implementation that enables the user to obtain alerts on system load and recommendations to proactively kill a set of processes to reclaim system memory. This enables user selected background process to be spared from the standard android policy of process termination, in lieu of foreground apps, relatively unimportant from user perspective, during that period. We show that manual reclaiming of memory based on recommendations from our app, reduces the automatic killing and measurement lag experienced by a sensor analytics app under test. This work is redundant if processing power and main memory of a smartphone is always surplus than required for its normal usage.

Internet of Things (IoT) has the potential to become a technology revolution with a vision of creating very large scale network, comprising of unprecedented number of connected devices. These devices, often referred to as smart items or intelligent things can be home appliances, healthcare devices, vehicles, buildings, factories and almost anything networked and fitted with sensors, actuators, embedded computers. There has been sustained research work and standardization effort from different IoT perspectives like integration of sensor and RFID devices to the Internet. With the increasing trend of gathering business insights from unstructured data, the high volume of data generated by such devices is also of interest. Cloud based data mining platforms are suitable for analyses of such data and researchers have proposed architectures where personal mobile phones can act as Edge Gateway between the sensor network and cloud analytics platform. It seems that the surge in the volume of data generated by huge number of Smart Items can only be matched if a large percentage of mobile users start sharing the computation capability of their personal devices and work together towards true Participatory Computing in the IoT systems. In this work we try to understand the challenges associated with running computation jobs on the mobile devices using different types of workload often observed in IoT applications. Based on the insights gained from experiments performed by us, we propose a scheme where mobile phones, residential gateways and other edge devices offer free slots to servers in a cloud based data analytics system. Based on the free time slots offered by the mobile phones, if commensurately sized computational jobs can be scheduled, the unpredictability associated with using mobile phones as grid resources can be solved.

The current emphasis on sensor-based intelligent and ubiquitous systems, more commonly known as “cyber-physical systems”, has the potential to give rise to a new generation of systems and services encompassing several domains such as e-Governance, healthcare, transportation, waste management, energy & utilities, insurance, etc., resulting in the metamorphosis of the Internet as we see it, into the Internet of Things (IoT). One probable commonality in each of these services will be the abundance of different types of data from different sources with the success of the systems depending on real-time or near real-time analysis of data. Such analyses are normally performed via well-known algorithms with a time-constraint on the execution, thus creating a requirement for parallel execution techniques. Some of these analyses may have a higher frequency of execution on a relatively small set of data, in which case the current big-data frameworks may actually add an overhead. Further, the frameworks like Hadoop demand the algorithms to be mapped onto a particular paradigm, which may not always be a suitable option. This paper, which is a work-in-progress, provides an experience report on the use of Condor, a well known Grid framework, for data-parallel “black-box” style execution of analysis algorithms in the context of Internet of Things. We concentrate on algorithms which are already in use, and can be partitioned into data-parallel subtasks without any modification and use Condor, which has traditionally been used for high-performance or high-throughput computing, as the execution framework.

Dynasoar is an infrastructure for dynamically deploying Web services over a grid or the Internet. It enables an approach to grid computing in which distributed applications are built around services instead of jobs. Dynasoar automatically deploys a service on an available host if no existing deployments exist, or if performance requirements cannot be met by existing deployments. This is analogous to remote job scheduling, but offers the opportunity for improved performance as the cost of moving and deploying the service can be shared across the processing of many messages. A key feature of the architecture is that it makes a clear separation between Web service providers, who offer services to consumers, and host providers, who offer computational resources on which services can be deployed, and messages sent to them processed. Separating these two components and defining their interactions, opens up the opportunity for interesting new organisational/business models