Graduate Students Seminar

Speaker 1: Weixin Wang

Title

Adaptive Matrix Constrained Entropy Bound Minimization Algorithm for Independent Component Analysis

Abstract

Independent component analysis (ICA) is a blind source separation technique used to discover hidden components in observed data.  ICA model is a completely data driven method, relying on the assumption of independence among hidden components to extract latent signals, without making strong assumptions about the source signals or the mixing process.  ICA has been widely applied to medical imaging technologies like functional magnetic resonance imaging (fMRI) to identify and separate distinct brain activity patterns. In contrast, constrained independent component analysis is a semi-blind approach, which utilizes appropriate prior information about the data.

In this work, we propose a novel  contrained ICA algorithm called constrained adaptive mixing matrix  entropy bound minimization (acr-A-EBM) algorithm. Compared to the existed constrained ICA algorithms, which mostly applied constraints to the hidden component,  acr-A-EBM imposes constraints on selected columns of the mixing matrix. This enables the algorithm to link the mixing matrix with prior information, like behavioral variables from psychiatric disorders, aiding studies of brain activity under different stimuli.  Moreover, compared with other constrained approaches that rely on the choice of constraint parameters, acr-A-EBM adaptively selects the constraint parameters for each constrained source.  We evaluate the performance of acr-A-EBM with other unconstrained ICA methods on two sets of face images. The results show that introducing this constrained structure improves ICA performance, even with non-independent sources, and acr-A-EBM performs better with more prior information and accurate constraints.

Speaker 2: Ehsan Shakeri

Title

Exploring Image Classification: From CNN Basics to AlexNet with CIFAR-10

Abstract

This talk will introduce the fundamentals of convolutional neural networks (CNNs), drawing inspiration from the human visual cortex, and their application in image classification tasks. We will start by examining the core components of CNNs, such as feature extraction, convolutional layers, and pooling layers. The focus will then shift to AlexNet, a pioneering deep learning model for image classification, and its application to the CIFAR-10 dataset. We will delve into the architecture of AlexNet and assess its performance, providing insights into the strengths and challenges of using deep learning for image classification.