Josef Bajada

Matthias Bartolo, Dylan Seychell, Josef Bajada

With the ever-growing variety of object detection approaches, this study explores a series of experiments that combine reinforcement learning (RL)-based visual attention methods with saliency ranking techniques to investigate transparent and sustainable solutions. By integrating saliency ranking for initial bounding box prediction and subsequently applying RL techniques to refine these predictions through a finite set of actions over multiple time steps, this study aims to enhance RL object detection accuracy. Presented as a series of experiments, this research investigates the use of various image feature extraction methods and explores diverse Deep Q-Network (DQN) architectural variations for deep reinforcement learning-based localisation agent training. Additionally, we focus on optimising the detection pipeline at every step by prioritising lightweight and faster models, while also incorporating the capability to classify detected objects, a feature absent in previous RL approaches. We show that by evaluating the performance of these trained agents using the Pascal VOC 2007 dataset, faster and more optimised models were developed. Notably, the best mean Average Precision (mAP) achieved in this study was 51.4, surpassing benchmarks set by RL-based single object detectors in the literature.

Josef Bajada, Francesco Borg Bonello

Real-time EEG-based Emotion Recognition (EEG-ER) with consumer-grade EEG devices involves classification of emotions using a reduced number of channels. These devices typically provide only four or five channels, unlike the high number of channels (32 or more) typically used in most current state-of-the-art research. In this work we propose to use Discrete Wavelet Transforms (DWT) to extract time-frequency domain features, and we use time-windows of a few seconds to perform EEG-ER classification. This technique can be used in real-time, as opposed to post-hoc on the full session data. We also apply baseline removal preprocessing, developed in prior research, to our proposed DWT Entropy and Energy features, which improves classification accuracy significantly. We consider two different classifier architectures, a 3D Convolutional Neural Network (3D CNN) and a Support Vector Machine (SVM). We evaluate both models on subject-independent and subject dependent setups to classify the Valence and Arousal dimensions of an individual's emotional state. We test them on both the full 32-channel data provided by the DEAP dataset, and also a reduced 5-channel extract of the same dataset. The SVM model performs best on all the presented scenarios, achieving an accuracy of 95.32% on Valence and 95.68% on Arousal for the full 32-channel subject-dependent case, beating prior real-time EEG-ER subject-dependent benchmarks. On the subject-independent case an accuracy of 80.70% on Valence and 81.41% on Arousal was also obtained. Reducing the input data to 5 channels only degrades the accuracy by an average of 3.54% across all scenarios, making this model appropriate for use with more accessible low-end EEG devices.

Josef Bajada, Maria Fox, Derek Long

Temporal planning often involves numeric effects that are directly proportional to their action's duration. These include continuous effects, where a numeric variable is subjected to a rate of change while the action is being executed, and discrete duration-dependent effects, where the variable is updated instantaneously but the magnitude of such change is computed from the action's duration. When these effects are linear, state--of--the--art temporal planners often make use of Linear Programming to ensure that these numeric updates are consistent with the chosen start times and durations of the plan's actions. This is typically done for each evaluated state as part of the search process. This exhaustive approach is not scalable to solve real-world problems that require long plans, because the linear program's size becomes larger and slower to solve. In this work we propose techniques that minimise this overhead by computing these checks more selectively and formulating linear programs that have a smaller footprint. The effectiveness of these techniques is demonstrated on domains that use a mix of discrete and continuous effects, which is typical of real-world planning problems. The resultant planner also outperforms most state-of-the-art temporal-numeric and hybrid planners, in terms of both coverage and scalability.