Summary: Our ongoing project focuses on exploring robust self-supervised learning algorithms. We're conducting detailed experiments to analyze the resilience of different learning methods, including contrastive and supervised contrastive learning, alongside standard supervised learning. Additionally, our research examines how adversarial training impacts representations in various learning schemes. We've observed significant differences between adversarial and clean representations across different models and tasks. However, after adversarial training, we notice a convergence between these representations, suggesting a universal set of features. Improving the similarity between adversarial and clean representations, especially towards the network's end, enhances overall robustness. These insights are crucial for designing and training more efficient and effective networks.

Summary: In this project, we introduced a novel algorithm aimed at constructing a family of decomposable models tailored for the classification process. Unlike traditional Bayesian networks that strictly adhere to directed acyclic graph structures, our algorithm enables the formation of cyclic graphs with limited size. This unique approach facilitates the effective extraction of feature interactions, enhancing the model's ability to capture complex relationships within the data. Additionally, by imposing constraints on the maximum cycle size, our algorithm mitigates the risk of overfitting, thereby ensuring more robust and generalized models. Furthermore, the project involved conducting comprehensive statistical analyses on internet traffic data. Leveraging a fusion of multiple classifiers and employing probabilistic models such as Hidden Naive Bayes, Averaged Tree Augmented Naïve Bayes, and the t-Cherry algorithm, we aimed to gain deeper insights into the underlying patterns and structures within the data.

Summary: In our project on supervised medical X-ray image analysis, we encountered a common challenge: the significant variations within each class and the similarities across different classes made categorization difficult. However, we found a solution in fuzzy set theory, which helped us determine each image's membership in different categories. This led us to develop a new fuzzy scheme for image classification, using Center Symmetric Local Binary Patterns (CCS-LBPs) extracted from multiscale image decompositions like Contourlet and Gabor representations after preprocessing.