COMP7250 Course Project
Submission deadline: 12 April 2024 / Oral presentation: 15 April 2024
In the Course Project, you are expected to use a machine learning approach to solve a research problem. You can use any programming language or machine learning platform.
Topics: choose either 1, 2 or 3
1. You can propose your own project topic. Please send the title and a brief description to bhanml@comp.hkbu.edu.hkno late than 31 Mar 2024. It cannot be your previous project.
2. You can reproduce a classical research paper from three specific areas (i.e., advanced knowledge in COMP7250), such as Deep Reinforcement Learning, Deep Learning with Graph-structured Data and Deep Learning in Imperfect Environments. Please send the title of the research paper tobhanml@comp.hkbu.edu.hkno late than 31 Mar 2024. You cannot use the existing source code. You need to develop the source code by yourself.
3. You can choose one from the following list of topics:
a. Data augmentation for image classification tasks: in this project, you shall design experiments to evaluate the effectiveness of data augmentation to improve the training performance and to reduce overfitting for image classification tasks. You can use existing CNN models and open data sets for this project.
Reference:https://towardsdatascience.com/data-augmentation-for-deep-learning-4fe21d1a4eb9
b. A comparison of SGD optimizers: in this project, you shall compare different SGD
optimizers in terms of their algorithms and convergence performance in different deep learning tasks. Please choose models and data sets by yourself.
Reference:https://keras.io/optimizers/
c. Bag of tricks for effective training: Training CNNs is typically very slow. In this project, you shall try different tricks to reduce the training time ofResNeton CIFAR10 data set using a single GPU.
Reference:https://myrtle.ai/how-to-train-your-resnet-8-bag-of-tricks/
Deliverables:
1. A project report that includes the introduction, problem definition, methodology, and experimental result, conclusion, and references.
2. Source codes and a readme file of how to reproduce your experiments and the plots.
Assessment Criteria:
1. Completeness of your project. (20%)
2. Quality and quantity of your implementation. We expect you to code by yourself. If you need to use some open-source codes, please clearly specify which parts of the codes are from open-source, and which parts are developed by yourself. (50%)
3. Quality of your report, with a focus on the presentation, analysis of experimental results, and visualization. (20%)
4. Quality of your oral presentation. (10%)