The School of Computer Science at the University of Windsor is pleased to present …
Date: Wednesday, April 23rd, 2025
Time: 2:30 pm
Location: Essex Hall, Room 122
In recent years, attention mechanisms have become a cornerstone in advancing deep learning technologies, particularly within the realm of Convolutional Neural Networks (CNNs). Among various deep learning algorithms, CNNs have gained widespread acceptance and application in fields such as fault diagnosis due to their capacity to address data in different dimensions. However, the intrinsic properties of convolution operations mean that their receptive domain is limited, often failing to capture global feature information and easily overlooking key features of the image. To overcome these limitations, new methods and algorithms, including attention mechanisms, have been proposed. These mechanisms enable CNNs to extract multi-scale local feature information and employ parallel processing of channel and spatial attention mechanisms to capture comprehensive global features. This seminar explores the transformative role of attention mechanisms in enhancing CNNs, enabling them to focus selectively on crucial features of input data, thus significantly improving both the accuracy and efficiency of tasks such as image classification, object detection, and segmentation. Practical implementations in TensorFlow are reviewed, showcasing real-world applications and substantial improvements in model interpretability and performance. Empirical studies across diverse domains demonstrate that attention-enhanced CNNs consistently outperform their traditional counterparts, with accuracy improvements of 1-3% in general image classification and up to 5% in specialized domains like medical imaging. Beyond accuracy gains, attention mechanisms provide enhanced interpretability through visualizable attention maps that reveal which image regions influence classification decisions. The integration of attention mechanisms transforms CNNs from models that process images uniformly to intelligent systems that prioritize informative features, close to human visual perception. This advancement represents a significant step toward more accurate, efficient, and interpretable deep learning models for image classification tasks. Through a comprehensive guide from intuition to implementation, this seminar aims to equip participants with a deep understanding of how attention mechanisms can be tailored to optimize CNNs, pushing the boundaries of what's possible in visual data processing and analysis.
Internal Reader: Dr. Dima Alhadidi
Internal Reader: Dr. Imran Ahmad
External Reader: Dr. Mohamed Belalia
Advisor (s): Dr. Boubakeur Boufama
