Fine tuning of global models: localization ofanomaly detection for video surveillance anduser-adaptation for BCI spellers

Fine-tuning is a technique that leverages knowledge from a source domain to enhanceperformance in a smaller, more limited target domain. Simple approaches, such astraining separate local models for each domain or creating a single global model usingall available data, often suffer from inherent limitations. Local models are prone tohigh variance due to limited data, while global models may exhibit high bias, failingto capture domain-specific nuances. This thesis aims to strike a balance in the biasvariancetrade-off by carefully fine-tuning a global model to target domains, witha particular focus on applications in video-based anomaly detection and visuallyevoked EEG signals for brain-computer interface (BCI) spellers.The central idea introduced in this thesis is the transition from global to localexpertise, starting with a low-variance global model trained on all available dataand progressively fine-tuning it to capture domain-specific nuances as local databecomes available. Furthermore, the thesis examines two distinct types of domainstructures: hierarchical and non-hierarchical. Hierarchical domains exhibit naturalrelationships or similarities, allowing structured methods to identify and leveragerelated data effectively during fine-tuning. In contrast, non-hierarchical domainslack inherent structures, necessitating alternative strategies to manage inter-domaindifferences and select relevant data for fine-tuning. These strategies aim to optimizeperformance by addressing the unique challenges posed by each domain structure.To exploit hierarchical relationships, this thesis employs context tree partitioning togroup similar domains, enabling more effective fine-tuning of models. As new dataarrives, the transition to local models enhances the localization of anomaly detectionby refining both the anomaly labels and their corresponding spatial locations. Applying this approach to anomaly detection, we observe improved performance onthe Street Scene and Shanghai datasets, achieving an Area Under Curve (AUC) of0.87 with the context tree partitioning method compared to 0.56 when using the entiredataset and 0.80 when using only the smallest partitions. For non-hierarchicaldata, such as those involving EEG signals, where constructing a hierarchy is notfeasible,user adaptation is achieved through direct similarity measures to guide thefine-tuning process. We enhance SSVEP BCI speller performance by adapting aDNN model for each new user without calibration. Starting with a global modeltrained on labeled data from previous users, the adaptation process leverages unsupervisedfine-tuning using pseudolabels generated from the new user’s data. Thisiterative approach significantly improves the character identification accuracy on twopublicly available large datasets (BENCH and BETA), particularly at short signallengths. 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 to28.28%–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 to20.66%–66.90% and reaching 19.50%–75.53% after final adaptation. These resultshighlight the effectiveness of leveraging silhouette scores, normalized distances, andlocal regularity loss to refine pseudolabels and optimize model performance, particularlyfor short signals in new user adapta