PhD.Dissertation Defense:Soner Özgün Pelvan

Knowledge Transfer with Fine Tuning for Visual and Visually Evoked EEG Signals

Soner Özgün Pelvan
Electronics Engineering, PhD Dissertation, 2024

Thesis Jury

Assoc. Prof. Hüseyin Özkan (Thesis Advisor),

Prof. Berrin Yanıkoğlu,

Prof. Özgür Gürbüz,

Prof. Ayşın Baytan Ertüzün,

Assoc. Prof. Erdem Akagündüz

Date & Time: 18th, December 2024 –  13:00 PM

Place: FENS 1040

Zoom Link: https://sabanciuniv.zoom.us/j/96422708783

Keywords : Knowledge Transfer, fine tuning, bias-variance tradeoff, context tree, domain adaptation, anomaly detection, Brain Computer Interface, SSVEP  

Abstract

Knowledge transfer in machine learning involves leveraging knowledge from one domain to improve performance in another, even when direct applicability is limited. This mirrors real-world scenarios where expertise is adapted across fields and is addressed through approaches like transfer learning, domain generalization/adaptation, and few-shot learning. Simple techniques, such as training local models for each domain or creating a global model using all data, often face limitations due to high variance or high bias, respectively. This thesis focuses on achieving a plausible balance in the bias-variance trade-off through careful fine-tuning of a global model to target domains, particularly for visual (video-based anomaly detection) and visually evoked EEG signals in brain-computer interface (BCI) spellers. The central idea introduced in this thesis is the transition from global to local expertise, starting with a low-variance global model trained on all available data and progressively fine-tuning it to capture domain-specific nuances as local data becomes available. We also consider the distinction between hierarchical and non-hierarchical structures in knowledge transfer applications. Hierarchical systems have natural distances or similarities between domains, enabling structured methods to effectively control variance whereas, non-hierarchical systems lack such a natural structure, necessitating alternative strategies to manage differences across domains.

To exploit hierarchical relationships, the thesis employs context tree partitioning to group similar domains, facilitating effective knowledge transfer and variance control through structured methods. Applying this approach to anomaly detection, we observe improved performance on the Street Scene and Shanghai datasets, achieving an Area Under Curve (AUC) of 0.87 with the context tree partitioning method β = 0.875, compared to 0.56 when using the entire dataset and 0.80 when using only the smallest partitions. For non-hierarchical data, such as those involving

EEG signals, where constructing a hierarchy is not feasible, direct similarity measures are employed to guide the transfer process. We enhance SSVEP BCI speller performance by adapting a DNN model for new users without calibration. The global model, initially trained on labeled data from previous users, undergoes unsupervised adaptation using pseudolabels generated from new user data. This iterative approach significantly improves the character identification accuracy on two publicly available large datasets (BENCH and BETA), particularly at short signal lengths. On the BENCH dataset, initial global model accuracy ranged from 21.75% to 71.32% for signal lengths of 0.2 to 1 second, improving after the first adaptation to 28.28%–88.34% and further to 29.85%–91.55% in subsequent iterations. Similarly, on the BETA dataset, initial accuracy ranged from 19.44% to 51.28%, increasing to 20.66%–66.90% and reaching 19.50%–75.53% after final adaptation. These results highlight the effectiveness of leveraging silhouette scores, normalized distances, and local regularity loss to refine pseudolabels and optimize model performance, particularly for short signals in new user adaptation scenarios.