Edge AI, Efficient LLMs, and ML for Health
My group develops adaptive, efficient, and trustworthy AI methods for edge and embodied systems, efficient LLMs and scaling, and digital health.
Themes
Three aligned research themes spanning edge AI, efficient LLMs and scaling, and digital health.
Full lists: Publications and
Google Scholar.
Edge and Embodied AI Systems
On-device, uncertainty-aware, and human-centred AI for trustworthy interaction, decision-making, and autonomy.
Edge & embodied AI
Efficient LLMs, Reasoning, and Scaling
Test-time scaling, reasoning and scaling strategies for efficient LLMs.
Efficient LLMs · Test-time scaling · reasoning
Machine Learning for Health
Design machine learning algorithms and systems for digital health systems, including mobile and wearable sensing to enable personalised health monitoring, risk prediction, and timely intervention in real-world settings.
mobile health · healthcare monitoring
Publications by theme
Edge and Embodied AI Systems
Efficient ML, Reasoning, and Scaling
Machine Learning for Health
Edge and Embodied AI Systems
On-device, uncertainty-aware, and human-centred AI for trustworthy interaction, decision-making, and autonomy:
-
TinyTTA: Efficient Test-time Adaptation via Early-exit Ensembles on Edge Devices
NeurIPS 2024
Introduces efficient test-time adaptation tailored to resource-constrained edge deployment.
-
HoRD: Robust Humanoid Control via History-Conditioned Reinforcement Learning and Online Distillation
ArXiv 2026
Develops robust control strategies for humanoid systems under real deployment challenges.
-
Edge-Cloud Collaborative Speech Emotion Captioning via Token-Level Speculative Decoding in Audio-Language Models
Interspeech 2026
Combines edge and cloud inference for efficient speech-emotion captioning with audio-language models.
-
E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models
NeurIPS 2025
Extends efficient adaptation to foundation models while removing the dependence on backpropagation at test time.
-
Position: Human-Robot Interaction Demands a Shift From Static Privacy Controls to Dynamic Learning
NeurIPS 2025, LAW
Argues for adaptive privacy mechanisms in embodied human-robot interaction systems.
-
XTransfer: Modality-Agnostic Few-Shot Model Transfer for Human Sensing at the Edge
ICML 2026
Enables few-shot cross-modality transfer for edge-deployed human sensing models.
-
AdaNODEs: Test Time Adaptation for Time Series Forecasting Using Neural ODEs
ICASSP 2026
Extends test-time adaptation to time-series forecasting using neural ODE dynamics.
-
LQA: A Lightweight Quantized-Adaptive Framework for Vision-Language Models on the Edge
ArXiv Preprint 2026
Introduces lightweight quantized-adaptive techniques for practical edge deployment of vision-language models.
-
LiteFat: Lightweight Spatio-Temporal Graph Learning for Real-Time Driver Fatigue Detection
IROS 2025
Targets real-time human-state understanding in safety-critical scenarios with efficient edge-ready modelling.
-
LeanTTA: A Backpropagation-Free and Stateless Approach to Quantized Test-Time Adaptation on Edge Devices
ArXiv Preprint 2025
Proposes a lightweight, backpropagation-free test-time adaptation method for quantized models on edge hardware.
-
LifeLearner: Hardware-Aware Meta Continual Learning System for Embedded Computing Platforms
ACM SenSys 2023
Builds a hardware-aware continual learning system designed for embedded and resource-constrained platforms.
-
PROS: an efficient pattern-driven compressive sensing framework for low-power biopotential-based wearables with on-chip intelligence
ACM MobiCom 2022
Enables efficient on-chip biopotential sensing for low-power wearable health monitoring devices.
-
LightLLM: A Versatile Large Language Model for Predictive Light Sensing
ACM SenSys 2025
Adapts language-model techniques for efficient predictive light sensing on embedded platforms.
Efficient ML, Reasoning, and Scaling
Test-time scaling, reasoning and adaptive scaling strategies for small language models:
-
Decoding Ambiguous Emotions with Test-Time Scaling in Audio-Language Models
TASLP 2026
Studies how additional test-time compute improves prediction for ambiguous emotion understanding in audio-language models.
-
Menta: A Small Language Model for On-Device Mental Health Prediction
ArXiv Preprint 2025
Introduces a compact model for on-device mental-health prediction with practical deployment focus.
-
Test Time Scaling for Auditory Cognition in Audio Language Models
ICASSP 2026
Demonstrates how test-time compute scaling improves auditory cognition in audio-language models.
-
Test Time Adaptation for Speech Emotion Recognition
ICASSP 2026
Applies test-time adaptation to improve speech emotion recognition under distribution shift.
-
FlowerTune: A Cross-Domain Benchmark for Federated Fine-Tuning of Large Language Models
NeurIPS 2025
Benchmarks federated fine-tuning of language models across heterogeneous domains for practical deployment.
-
CoCoEmo: Composable and Controllable Human-Like Emotional TTS via Activation Steering
ICML 2026
Enables composable and controllable emotional speech synthesis via activation steering in language models.
-
Beyond Scale: Small Language Models are Comparable to GPT-4 in Mental Health Understanding
ACII 2025
Provides evidence that compact models can be highly competitive for mental-health understanding tasks.
-
Disentangling Reasoning in Large Audio-Language Models for Ambiguous Emotion Prediction
Interspeech 2026
Analyzes and separates reasoning behavior in multimodal models for robust prediction under ambiguous inputs.
-
Enabling On-Device LLMs Personalization with Smartphone Sensing
ACM IMWUT/UbiComp 2024
Personalises compact on-device language models using smartphone sensing for user-adaptive behaviour.
Machine Learning for Health
Machine learning algorithms and systems for digital health, including mobile and wearable sensing:
-
UR2M: Uncertainty and Resource-Aware Event Detection on Microcontrollers
IEEE PerCom 2024
Combines uncertainty modeling with resource awareness for efficient event detection on embedded devices.
-
HealthSLM-Bench: Benchmarking Small Language Models for Mobile and Wearable Healthcare Monitoring
NeurIPS GenAI4Health 2025
Benchmarks compact language models for realistic mobile and wearable healthcare tasks.
-
Efficient and Personalized Mobile Health Event Prediction via Small Language Models
ACM MobiCom 2024
Shows how compact language models can support personalised mobile health prediction efficiently.
-
Multimodal Large Language Models in Human-centered Health: Practical Insights
IEEE Pervasive Computing 2025
Provides practical guidance for applying multimodal LLMs in human-centered health settings.
-
UDAMA: Unsupervised Domain Adaptation through Multi-discriminator Adversarial Training with Noisy Labels Improves Cardio-fitness Prediction
NeurIPS ML4H 2023
Improves wearable cardio-fitness prediction under domain shift using adversarial domain adaptation with noisy labels.
-
Turning Silver into Gold: Domain Adaptation with Noisy Labels for Wearable Cardio-Respiratory Fitness Prediction
NeurIPS ML4Health 2022
Addresses label noise and domain shift for reliable cardio-respiratory fitness prediction from wearables.
-
NarrativeSense: Predicting Affective States in University Students through Smartphone Sensing and Contextual Narratives
ACM Transactions on Computing for Healthcare 2026
Combines sensing and contextual narratives to understand human affect in everyday life.
-
Token-Level Logits Matter: A Closer Look at Speech Foundation Models for Ambiguous Emotion Recognition
Interspeech 2025
Analyzes token-level model behaviour for more reliable ambiguous-emotion recognition in speech.
-
Subject Adaptive Loose Fitting Smart Garment Platform For Human Activity Recognition
ACM Transactions on Sensor Networks 2023
Develops a wearable garment platform for subject-adaptive activity recognition in everyday health monitoring.
-
A Framework for On-Device Uncertainty-Aware Event Detection on Microcontrollers
Pervasive and Mobile Computing 2026
Develops a practical on-device framework for reliable event detection under severe microcontroller constraints.