Curriculum Learning for Image Classification CNNs

Project Overview

As principal researcher in the ULL DADL Lab (Data Analysis and Deep Learning Lab) under Dr. Aminul Islam, I investigated curriculum learning approaches for training convolutional neural networks on image classification tasks. This research was presented at IBM Thomas J. Watson Laboratory during AICS’24, providing valuable experience with academic research methodologies and scientific presentation.

Research Focus

The project explored how the complexity and ordering of training examples affects CNN learning performance. A key finding was that sorting images by wavelet entropy alone produced poor results, leading to investigation of more sophisticated complexity-based curriculum learning approaches that combine multiple indicators.

Skills Developed

Through this research experience, I gained:

• Technical Skills: Experience with Python, CNN architectures, and wavelet analysis techniques
• Research Methods: Understanding of experimental design, statistical modeling, and negative result interpretation
• Problem Solving: Approach to debugging complex machine learning systems and analyzing unexpected results
• Academic Communication: Experience presenting technical research at IBM Thomas J. Watson Laboratory

Key Findings

• Negative Results: Wavelet entropy alone as a complexity measure produced poor curriculum learning results
• Methodology Insights: Importance of rigorous experimental design in curriculum learning research
• Technical Learning: Hands-on experience with CNN training methods and statistical analysis
• Research Skills: Understanding of academic presentation and peer review processes

Learning Outcomes

This research experience demonstrated the value of negative results in advancing scientific understanding and reinforced my interest in pursuing graduate study in computational learning systems. The presentation at IBM validated the quality of undergraduate research contributions to the field.

Future Directions

This experience has motivated me to pursue advanced study in machine learning and educational technology, particularly exploring how AI systems can better support human learning while avoiding the pitfalls of poorly designed educational technology.

Academic Impact

Conference Presentation: Presented research findings at IBM Thomas J. Watson Laboratory during the AICS’24 conference, demonstrating the caliber of research recognized by industry leaders.

Research Insights: The discovery that simple complexity measures like wavelet entropy alone are insufficient for effective curriculum learning opened new research directions into multi-modal complexity assessment.

Future Directions

This research established the foundation for more sophisticated curriculum learning approaches, highlighting the need for:

• Multi-factor complexity measures
• Adaptive curriculum strategies
• Domain-specific curriculum design principles

The work demonstrates both technical rigor and the kind of critical thinking that leads to breakthrough insights through careful investigation of negative results.