publications

2026

1. 

   Faster, Cheaper, More Accurate: Specialised Knowledge Tracing Models Outperform LLMs

   Prarthana Bhattacharyya^*, Joshua Mitton^*, Ralph Abboud, and 1 more author

   Pre-print, Under review, 2026

   TL;DR arXiv Blog

   Investigates whether large language models can effectively predict student responses compared to specialised knowledge tracing models. Demonstrates that KT models outperform LLMs in accuracy and F1 scores while being orders of magnitude faster and cheaper to deploy.

2. 

   Misconception Diagnosis From Student-Tutor Dialogue: Generate, Retrieve, Rerank

   Joshua Mitton^*, Prarthana Bhattacharyya^*, Digory Smith, and 3 more authors

   Pre-print, Under review, 2026

   TL;DR arXiv

   Presents a method for identifying student misconceptions in educational dialogues using a generate-retrieve-rerank pipeline. Fine-tuned models enhance quality and enable smaller models to match larger closed-source alternatives.

2025

1. 

   Uncertainty-Aware Knowledge Tracing Models

   Joshua Mitton^*, Prarthana Bhattacharyya^*, Ralph Abboud, and 1 more author

   Pre-print, Under review, 2025

   TL;DR arXiv Blog

   New uncertainty-based approach that identifies struggling students where traditional metrics fail, enabling EdTech platforms to intervene at critical moments.

2. 

   Helios 2.0: A Robust, Ultra-Low Power Gesture Recognition System Optimised for Event-Sensor based Wearables

   Prarthana Bhattacharyya^*, Joshua Mitton^*, Ryan Page^*, and 8 more authors

   Pre-print, Under review, 2025

   TL;DR arXiv PDF Video

   Architected mobile-optimized models for smart glasses achieving 25x power reduction with 20% accuracy gain (F1>80%). Profiled inference bottlenecks, implemented quantization and efficient architectures. Deployed at CES-2025 in commercial development kit.

2024

1. 

   Helios: An extremely low power event-based gesture recognition for always-on smart eyewear

   Prarthana Bhattacharyya^*, Joshua Mitton^*, Ryan Page^*, and 8 more authors

   Best Paper, Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2024

   TL;DR arXiv PDF Video

   Pioneered low-power low-latency ML models, representations and custom loss functions enabling micro-gesture detection for Meta-Raybans. Rapidly delivered a prototype within 6 months, publicly demonstrated at AWE-USA 2024 and featured by CNET. Won Best Paper award at ECCV-2024 Workshop.

2. 

   SSL-Interactions: Pretext Tasks for Interactive Trajectory Prediction

   Prarthana Bhattacharyya, Chengjie Huang, and Krzysztof Czarnecki

   In 2024 IEEE Intelligent Vehicles Symposium (IV), 2024

   TL;DR arXiv PDF

   Self-supervised pretext tasks boost multi-agent interaction prediction by 8% in complex driving scenarios.

2023

1. 

   Perception and Prediction in Multi-Agent Urban Traffic Scenarios for Autonomous Driving

   Prarthana Bhattacharyya

   University of Waterloo, 2023

   HTML PDF Video

2022

1. 

   SSL-Lanes: Self-Supervised Learning for Motion Forecasting in Autonomous Driving

   Prarthana Bhattacharyya, Chengjie Huang, and Krzysztof Czarnecki

   In Conference on Robot Learning (CoRL), 2022

   TL;DR arXiv PDF Code

   Pioneered SSL methods for autonomous vehicle motion forecasting (80+ citations, 120+ GitHub stars, presented to Gatik AI). Achieves state-of-the-art prediction accuracy while maintaining lower computational complexity than transformer-based approaches.

2. 

   Visual Representation Learning with Self-Supervised Attention for Low-Label High-data Regime

   Prarthana Bhattacharyya, Chenge Li, Xiaonan Zhao, and 2 more authors

   In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022

   TL;DR arXiv PDF Code

   Developed self-supervised transformer framework reducing manual data annotation requirements. Achieved 4-11% improvement over SOTA on five benchmarks (ICASSP-2022).

2021

1. 

   Sa-Det3D: Self-Attention Based Context-Aware 3D Object Detection

   Prarthana Bhattacharyya, Chengjie Huang, and Krzysztof Czarnecki

   In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops , 2021

   TL;DR arXiv PDF Video Code

   First method to combine self-attention with CNNs for 3D point cloud detection (90+ citations, 150+ GitHub stars). 1.5 AP accuracy gain while slashing model size by 15-80% and compute by 30-50%. Proven across KITTI, nuScenes, and Waymo datasets.

2020

1. 

   Deformable PV-RCNN: Improving 3D Object Detection with Learned Deformations

   Prarthana Bhattacharyya and Krzysztof Czarnecki

   In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, 2020

   TL;DR arXiv Code

   First method to integrate deformable convolutions in 3D object detection (40+ citations, 100+ GitHub stars).

2019

1. 

   Fantrack: 3d multi-object tracking with feature association network

   Erkan Baser, Venkateshwaran Balasubramanian^*, Prarthana Bhattacharyya^*, and 1 more author

   In 2019 IEEE Intelligent Vehicles Symposium (IV), 2019

   TL;DR arXiv PDF Code

   NeurIPS 2019 3D tracking competition winner (150+ citations). Replaces hand-crafted cost functions and complex optimization with a CNN that learns end-to-end. Competitive multi-object tracking by automatically learning optimal assignments.

2017

1. 

   3D scene understanding at urban intersection using stereo vision and digital map

   Prarthana Bhattacharyya, Yanlei Gu, Jiali Bao, and 2 more authors

   In IEEE 85th Vehicular Technology Conference (VTC Spring), 2017

   TL;DR arXiv PDF

   Proposed a novel method for 3D scene understanding at urban intersections by integrating stereo vision with digital map data. Tested on real-data collected from urban intersections in Tokyo.