Projects.
Stochastic Sparse Sampling: A Framework for Variable-Length Medical Time Series Classification
Xavier Mootoo, Alan A. Díaz-Montiel, Milad Lankarany, Hina Tabassum
Full Version: Anonymous Machine Learning Conference (under review)
Workshop: NeurIPS 2024 Workshop on Time Series in the Age of Large Models
While the majority of time series classification research has focused on modeling fixed-length sequences, variable-length time series classification (VTSC) remains critical in healthcare, where sequence length may vary among patients and events. To address this challenge, we propose Stochastic Sparse Sampling (SSS), a novel VTSC framework developed for medical time series. SSS manages variable-length sequences by sparsely sampling fixed windows to compute local predictions, which are then aggregated and calibrated to form a global prediction. We apply SSS to the task of seizure onset zone (SOZ) localization, a critical VTSC problem requiring identification of seizure-inducing brain regions from variable-length electrophysiological time series. We evaluate our method on the Epilepsy iEEG Multicenter Dataset, a heterogeneous collection of intracranial electroencephalography (iEEG) recordings obtained from four independent medical centers. SSS demonstrates superior performance compared to state-of-the-art (SOTA) baselines across most medical centers, and superior performance on all out-of-distribution (OOD) unseen medical centers. Additionally, SSS naturally provides post-hoc insights into local signal characteristics related to the SOZ, by visualizing temporally averaged local predictions throughout the signal.
QoS-Aware Deep Unsupervised Learning for STAR-RIS Assisted Networks: A Novel Differentiable Projection Framework
Mehrazin Alizadeh, Xavier Mootoo, Hina Tabassum
IEEE Wireless Communication Letters 2024
We present a novel unsupervised learning framework for solving constrained optimization problems with non-convex Quality-of-Service (QoS) constraints. In particular, we propose a differentiable projection function that maps model outputs onto the feasible solution space, ensuring zero constraint violation while enabling efficient exploration. Our approach combines a custom neural network for generating candidate solutions, the custom projection function for constraint enforcement, and an unsupervised training procedure to optimize the primary objective. We evaluate our method on a challenging wireless communications application: joint beamforming and phase-shift optimization in reconfigurable intelligent surface (RIS) assisted networks. We consider a simultaneously transmitting and reflecting RIS (STAR-RIS) in a multi-user, multi-antenna system, maximizing downlink network sum-rate under user-specific QoS constraints. Results show that our framework outperforms genetic algorithms and existing projection-based methods in achieved sum-rate, computational efficiency, and convergence speed, while maintaining zero probability of constraint violation.
T-VICReg: Self-Supervised Learning for Time Series with Partial Temporal Invariance, Noncontrastive and Augmentation-Free
Xavier Mootoo, Alan A. Díaz-Montiel, Milad Lankarany
T-CAIREM Machine Learning Internship 2023, University of Toronto
We propose Temporal Variance-Invariance-Covariance Regularization (T-VICReg), a novel self-supervised learning method for time series enabling learned representations to be partially invariant to translations in time. Information from both past and future representations is integrated into the current-time representation, allowing for the detection of state transitions in a variety of downstream tasks. We validate T-VICReg on the OpenNeuro dataset ds003029, containing intracranial electroencephalography (IEEG) signals from epilepsy patients, on the task of binary seizure classification (ictal, nonictal) and multiclass seizure detection (preictal, ictal, postictal). Fine-tuning the encoder from T-VICReg resulted in Top-1 accuracies of 92.92% and 89.26%, compared to the supervised baseline with Top-1 accuracies of 89.23% and 84.07% for the binary and multiclass tasks, respectively. T-VICReg is noncontrastive,augmentation-free, and compatible with both continuous and discrete timeseries, allowing for flexible use in many contexts.