Soumyajit Chatterjee

Fan Mo, Mohammad Malekzadeh, Soumyajit Chatterjee et al.

Ubiquitous wearable and mobile devices provide access to a diverse set of data. However, the mobility demand for our devices naturally imposes constraints on their computational and communication capabilities. A solution is to locally learn knowledge from data captured by ubiquitous devices, rather than to store and transmit the data in its original form. In this paper, we develop a federated learning framework, called Centaur, to incorporate on-device data selection at the edge, which allows partition-based training of a deep neural nets through collaboration between constrained and resourceful devices within the multidevice ecosystem of the same user. We benchmark on five neural net architecture and six datasets that include image data and wearable sensor time series. On average, Centaur achieves ~19% higher classification accuracy and ~58% lower federated training latency, compared to the baseline. We also evaluate Centaur when dealing with imbalanced non-iid data, client participation heterogeneity, and different mobility patterns. To encourage further research in this area, we release our code at https://github.com/nokia-bell-labs/data-centric-federated-learning

Prithviraj Pramanik, Prasenjit Karmakar, Praveen Kumar Sharma et al.

Efficient air quality sensing serves as one of the essential services provided in any recent smart city. Mostly facilitated by sparsely deployed Air Quality Monitoring Stations (AQMSs) that are difficult to install and maintain, the overall spatial variation heavily impacts air quality monitoring for locations far enough from these pre-deployed public infrastructures. To mitigate this, we in this paper propose a framework named AQuaMoHo that can annotate data obtained from a low-cost thermo-hygrometer (as the sole physical sensing device) with the AQI labels, with the help of additional publicly crawled Spatio-temporal information of that locality. At its core, AQuaMoHo exploits the temporal patterns from a set of readily available spatial features using an LSTM-based model and further enhances the overall quality of the annotation using temporal attention. From a thorough study of two different cities, we observe that AQuaMoHo can significantly help annotate the air quality data on a personal scale.

Josh Millar, Ashok Samraj Thangarajan, Soumyajit Chatterjee et al.

The miniaturisation of neural processing units (NPUs) and other low-power accelerators has enabled their integration into microcontroller-scale wearable hardware, supporting near-real-time, offline, and privacy-preserving inference. Yet physiological signal analysis has remained infeasible on such hardware; recent Transformer-based models show state-of-the-art performance but are prohibitively large for resource- and power-constrained hardware and incompatible with $μ$NPUs due to their dynamic attention operations. We introduce PhysioLite, a lightweight, NPU-compatible model architecture and training framework for ECG/EMG signal analysis. Using learnable wavelet filter banks, CPU-offloaded positional encoding, and hardware-aware layer design, PhysioLite reaches performance comparable to state-of-the-art Transformer-based foundation models on ECG and EMG benchmarks, while being <10% of the size ($\sim$370KB with 8-bit quantization). We also profile its component-wise latency and resource consumption on both the MAX78000 and HX6538 WE2 $μ$NPUs, demonstrating its viability for signal analysis on constrained, battery-powered hardware. We release our model(s) and training framework at: https://github.com/j0shmillar/physiolite.

Soumyajit Chatterjee, Bivas Mitra, Sandip Chakraborty

Complex activities of daily living (ADLs) often consist of multiple micro-activities. When performed sequentially, these micro-activities help the user accomplish the broad macro-activity. Naturally, a deeper understanding of these micro-activities can help develop more sophisticated human activity recognition (HAR) models and add explainability to their inferred conclusions. Previous research has attempted to achieve this by utilizing fine-grained annotated data that provided the required supervision and rules for associating the micro-activities to identify the macro-activity. However, this ``bottom-up'' approach is unrealistic as getting such high-quality, fine-grained annotated sensor datasets is challenging, costly, and time-consuming. Understanding this, in this paper, we develop AmicroN, which adapts a ``top-down'' approach by exploiting coarse-grained annotated data to expand the macro-activities into their constituent micro-activities without any external supervision. In the backend, AmicroN uses \textit{unsupervised} change-point detection to search for the micro-activity boundaries across a complex ADL. Then, it applies a \textit{generalized zero-shot} approach to characterize it. We evaluate AmicroN on two real-life publicly available datasets and observe that AmicroN can identify the micro-activities with micro F\textsubscript{1}-score $>0.75$ for both datasets. Additionally, we also perform an initial proof-of-concept on leveraging the state-of-the-art (SOTA) large language models (LLMs) with attribute embeddings predicted by AmicroN to enhance further the explainability surrounding the detection of micro-activities.

Aritra Hota, Soumyajit Chatterjee, Sandip Chakraborty

Traditional human-in-the-loop-based annotation for time-series data like inertial data often requires access to alternate modalities like video or audio from the environment. These alternate sources provide the necessary information to the human annotator, as the raw numeric data is often too obfuscated even for an expert. However, this traditional approach has many concerns surrounding overall cost, efficiency, storage of additional modalities, time, scalability, and privacy. Interestingly, recent large language models (LLMs) are also trained with vast amounts of publicly available alphanumeric data, which allows them to comprehend and perform well on tasks beyond natural language processing. Naturally, this opens up a potential avenue to explore LLMs as virtual annotators where the LLMs will be directly provided the raw sensor data for annotation instead of relying on any alternate modality. Naturally, this could mitigate the problems of the traditional human-in-the-loop approach. Motivated by this observation, we perform a detailed study in this paper to assess whether the state-of-the-art (SOTA) LLMs can be used as virtual annotators for labeling time-series physical sensing data. To perform this in a principled manner, we segregate the study into two major phases. In the first phase, we investigate the challenges an LLM like GPT-4 faces in comprehending raw sensor data. Considering the observations from phase 1, in the next phase, we investigate the possibility of encoding the raw sensor data using SOTA SSL approaches and utilizing the projected time-series data to get annotations from the LLM. Detailed evaluation with four benchmark HAR datasets shows that SSL-based encoding and metric-based guidance allow the LLM to make more reasonable decisions and provide accurate annotations without requiring computationally expensive fine-tuning or sophisticated prompt engineering.

Jiaheng Dong, Hong Jia, Soumyajit Chatterjee et al.

Speech Foundation Models encounter significant performance degradation when deployed in real-world scenarios involving acoustic domain shifts, such as background noise and speaker accents. Test-time adaptation (TTA) has recently emerged as a viable strategy to address such domain shifts at inference time without requiring access to source data or labels. However, existing TTA approaches, particularly those relying on backpropagation, are memory-intensive, limiting their applicability in speech tasks and resource-constrained settings. Although backpropagation-free methods offer improved efficiency, existing ones exhibit poor accuracy. This is because they are predominantly developed for vision tasks, which fundamentally differ from speech task formulations, noise characteristics, and model architecture, posing unique transferability challenges. In this paper, we introduce E-BATS, the first Efficient BAckpropagation-free TTA framework designed explicitly for speech foundation models. E-BATS achieves a balance between adaptation effectiveness and memory efficiency through three key components: (i) lightweight prompt adaptation for a forward-pass-based feature alignment, (ii) a multi-scale loss to capture both global (utterance-level) and local distribution shifts (token-level) and (iii) a test-time exponential moving average mechanism for stable adaptation across utterances. Experiments conducted on four noisy speech datasets spanning sixteen acoustic conditions demonstrate consistent improvements, with 4.1%-13.5% accuracy gains over backpropagation-free baselines and 2.0-6.4 times GPU memory savings compared to backpropagation-based methods. By enabling scalable and robust adaptation under acoustic variability, this work paves the way for developing more efficient adaptation approaches for practical speech processing systems in real-world environments.

Michal Danilowski, Soumyajit Chatterjee, Abhirup Ghosh

The performance of deep learning models depends heavily on test samples at runtime, and shifts from the training data distribution can significantly reduce accuracy. Test-time adaptation (TTA) addresses this by adapting models during inference without requiring labeled test data or access to the original training set. While research has explored TTA from various perspectives like algorithmic complexity, data and class distribution shifts, model architectures, and offline versus continuous learning, constraints specific to mobile and edge devices remain underexplored. We propose BoTTA, a benchmark designed to evaluate TTA methods under practical constraints on mobile and edge devices. Our evaluation targets four key challenges caused by limited resources and usage conditions: (i) limited test samples, (ii) limited exposure to categories, (iii) diverse distribution shifts, and (iv) overlapping shifts within a sample. We assess state-of-the-art TTA methods under these scenarios using benchmark datasets and report system-level metrics on a real testbed. Furthermore, unlike prior work, we align with on-device requirements by advocating periodic adaptation instead of continuous inference-time adaptation. Experiments reveal key insights: many recent TTA algorithms struggle with small datasets, fail to generalize to unseen categories, and depend on the diversity and complexity of distribution shifts. BoTTA also reports device-specific resource use. For example, while SHOT improves accuracy by $2.25\times$ with $512$ adaptation samples, it uses $1.08\times$ peak memory on Raspberry Pi versus the base model. BoTTA offers actionable guidance for TTA in real-world, resource-constrained deployments.

Ting Dang, Soumyajit Chatterjee, Hong Jia et al.

Test time adaptation (TTA) has emerged as a promising solution to adapt pre-trained models to new, unseen data distributions using unlabeled target domain data. However, most TTA methods are designed for independent data, often overlooking the time series data and rarely addressing forecasting tasks. This paper presents AdaNODEs, an innovative source-free TTA method tailored explicitly for time series forecasting. By leveraging Neural Ordinary Differential Equations (NODEs), we propose a novel adaptation framework that accommodates the unique characteristics of distribution shifts in time series data. Moreover, we innovatively propose a new loss function to tackle TTA for forecasting tasks. AdaNODEs only requires updating limited model parameters, showing effectiveness in capturing temporal dependencies while avoiding significant memory usage. Extensive experiments with one- and high-dimensional data demonstrate that AdaNODEs offer relative improvements of 5.88\% and 28.4\% over the SOTA baselines, especially demonstrating robustness across higher severity distribution shifts.

Alexander Murphy, Michal Danilowski, Soumyajit Chatterjee et al.

Test-Time Adaptation (TTA) methods are often computationally expensive, require a large amount of data for effective adaptation, or are brittle to hyperparameters. Based on a theoretical foundation of the geometry of the latent space, we are able to significantly improve the alignment between source and distribution-shifted samples by re-centering target data embeddings at the origin. This insight motivates NEO -- a hyperparameter-free fully TTA method, that adds no significant compute compared to vanilla inference. NEO is able to improve the classification accuracy of ViT-Base on ImageNet-C from 55.6% to 59.2% after adapting on just one batch of 64 samples. When adapting on 512 samples NEO beats all 7 TTA methods we compare against on ImageNet-C, ImageNet-R and ImageNet-S and beats 6/7 on CIFAR-10-C, while using the least amount of compute. NEO performs well on model calibration metrics and additionally is able to adapt from 1 class to improve accuracy on 999 other classes in ImageNet-C. On Raspberry Pi and Jetson Orin Nano devices, NEO reduces inference time by 63% and memory usage by 9% compared to baselines. Our results based on 3 ViT architectures and 4 datasets show that NEO can be used efficiently and effectively for TTA.

Irene Tenison, Soumyajit Chatterjee, Fahim Kawsar et al.

To utilize pre-trained neural networks on edge and mobile devices, we often require efficient adaptation to user-specific runtime data distributions while operating under limited compute and memory resources. On-device retraining with a target dataset can facilitate such adaptations; however, it remains impractical due to the increasing depth of modern neural nets, as well as the computational overhead associated with gradient-based optimization across all layers. Current approaches reduce training cost by selecting a subset of layers for retraining, however, they rely on labeled data, at least one full-model backpropagation, or server-side meta-training; limiting their suitability for constrained devices. We introduce AdaBet, a gradient-free layer selection approach to rank important layers by analyzing topological features of their activation spaces through Betti Numbers and using forward passes alone. AdaBet allows selecting layers with high learning capacity, which are important for retraining and adaptation, without requiring labels or gradients. Evaluating AdaBet on sixteen pairs of benchmark models and datasets, shows AdaBet achieves an average gain of 5% more classification accuracy over gradient-based baselines while reducing average peak memory consumption by 40%.

Yuvraj Dutta, Soumyajit Chatterjee, Sandip Chakraborty et al.

Accurate and adaptive network throughput prediction is essential for latency-sensitive and bandwidth-intensive applications in 5G and emerging 6G networks. However, most existing methods rely on centralized training with uniformly collected data, limiting their applicability in heterogeneous mobile environments with non-IID data distributions. This paper presents the first comprehensive benchmarking of federated learning (FL) strategies for throughput prediction in realistic 5G edge scenarios. We evaluate three aggregation algorithms - FedAvg, FedProx, and FedBN - across four time-series architectures: LSTM, CNN, CNN+LSTM, and Transformer, using five diverse real-world datasets. We systematically analyze the effects of client heterogeneity, cohort size, and history window length on prediction performance. Our results reveal key trade-offs among model complexities, convergence rates, and generalization. It is found that FedBN consistently delivers robust performance under non-IID conditions. On the other hand, LSTM and Transformer models outperform CNN-based baselines by up to 80% in R2 scores. Moreover, although Transformers converge in half the rounds of LSTM, they require longer history windows to achieve a high R2, indicating higher context dependence. LSTM is, therefore, found to achieve a favorable balance between accuracy, rounds, and temporal footprint. To validate the end-to-end applicability of the framework, we have integrated our FL-based predictors into a live adaptive streaming pipeline. It is seen that FedBN-based LSTM and Transformer models improve mean QoE scores by 11.7% and 11.4%, respectively, over FedAvg, while also reducing the variance. These findings offer actionable insights for building scalable, privacy-preserving, and edge-aware throughput prediction systems in next-generation wireless networks.

Soumyajit Chatterjee, Arun Singh, Bivas Mitra et al.

Human activities within smart infrastructures generate a vast amount of IMU data from the wearables worn by individuals. Many existing studies rely on such sensory data for human activity recognition (HAR); however, one of the major bottlenecks is their reliance on pre-annotated or labeled data. Manual human-driven annotations are neither scalable nor efficient, whereas existing auto-annotation techniques heavily depend on video signatures. Still, video-based auto-annotation needs high computation resources and has privacy concerns when the data from a personal space, like a smart-home, is transferred to the cloud. This paper exploits the acoustic signatures generated from human activities to label the wearables' IMU data at the edge, thus mitigating resource requirement and data privacy concerns. We utilize acoustic-based pre-trained HAR models for cross-modal labeling of the IMU data even when two individuals perform simultaneous but different activities under the same environmental context. We observe that non-overlapping acoustic gaps exist with a high probability during the simultaneous activities performed by two individuals in the environment's acoustic context, which helps us resolve the overlapping activity signatures to label them individually. A principled evaluation of the proposed approach on two real-life in-house datasets further augmented to create a dual occupant setup, shows that the framework can correctly annotate a significant volume of unlabeled IMU data from both individuals with an accuracy of $\mathbf{82.59\%}$ ($\mathbf{\pm 17.94\%}$) and $\mathbf{98.32\%}$ ($\mathbf{\pm 3.68\%}$), respectively, for a workshop and a kitchen environment.

Ratna Mandal, Prasenjit Karmakar, Soumyajit Chatterjee et al.

Intelligent city transportation systems are one of the core infrastructures of a smart city. The true ingenuity of such an infrastructure lies in providing the commuters with real-time information about citywide transports like public buses, allowing her to pre-plan the travel. However, providing prior information for transportation systems like public buses in real-time is inherently challenging because of the diverse nature of different stay-locations that a public bus stops. Although straightforward factors stay duration, extracted from unimodal sources like GPS, at these locations look erratic, a thorough analysis of public bus GPS trails for 720km of bus travels at the city of Durgapur, a semi-urban city in India, reveals that several other fine-grained contextual features can characterize these locations accurately. Accordingly, we develop BuStop, a system for extracting and characterizing the stay locations from multi-modal sensing using commuters' smartphones. Using this multi-modal information BuStop extracts a set of granular contextual features that allow the system to differentiate among the different stay-location types. A thorough analysis of BuStop using the collected dataset indicates that the system works with high accuracy in identifying different stay locations like regular bus stops, random ad-hoc stops, stops due to traffic congestion stops at traffic signals, and stops at sharp turns. Additionally, we also develop a proof-of-concept setup on top of BuStop to analyze the potential of the framework in predicting expected arrival time, a critical piece of information required to pre-plan travel, at any given bus stop. Subsequent analysis of the PoC framework, through simulation over the test dataset, shows that characterizing the stay-locations indeed helps make more accurate arrival time predictions with deviations less than 60s from the ground-truth arrival time.

Snigdha Das, Soumyajit Chatterjee, Sandip Chakraborty et al.

In an organization, individuals prefer to form various formal and informal groups for mutual interactions. Therefore, ubiquitous identification of such groups and understanding their dynamics are important to monitor activities, behaviours and well-being of the individuals. In this paper, we develop a lightweight, yet near-accurate, methodology, called MeetSense, to identify various interacting groups based on collective sensing through users' smartphones. Group detection from sensor signals is not straightforward because users in proximity may not always be under the same group. Therefore, we use acoustic context extracted from audio signals to infer interaction pattern among the subjects in proximity. We have developed an unsupervised and lightweight mechanism for user group detection by taking cues from network science and measuring the cohesivity of the detected groups in terms of modularity. Taking modularity into consideration, MeetSense can efficiently eliminate incorrect groups, as well as adapt the mechanism depending on the role played by the proximity and the acoustic context in a specific scenario. The proposed method has been implemented and tested under many real-life scenarios in an academic institute environment, and we observe that MeetSense can identify user groups with close to 90% accuracy even in a noisy environment.