CAP6614-Current Topics in Machine Learning (FALL 2019)

Assistant Professor  •  Imaging Scientist  •   •  Home  • 

CAP 6614:  Current Topics in Machine Learning

Fall 2019

Instructor:Dr. Ulas Bagci

Location: BA1 – 0213

Time: Monday / Wednesday 1.00 – 2.15 pm

Office Hour:  Monday / Wednesday 2.30 pm-3.30pm
Credit Hours: 3.0

Starts / Ends: 8/26/2019 – 12/11/2019

Course Description

This course will be a deep dive into current topics in machine learning, specifically deep learning. In this particular semester, instructor will focus on two important sub-field of deep learning, first: generalization, second: interpretability.

Overall, the course will include three core concepts:

1)individual presentation: papers selected from the cutting-edge machine learning conferences, including ICML (International Conference on Machine Learning), NeurIPS (Neural Information Processing Systems), and ICLR (International Conference on Learning Representations). The class will follow a discussion format.

2) mini projects (programming assignments): students will implement known deep network algorithms and will be asked to make further modifications to improve the state-of-the-art results (classification, segmentation, etc).

3) a course project, which will be the same for everyone, and chosen from available machine learning challenges. Students will work alone, solve the problem, run the experiments, and write 8-10 pages paper for their project. The paper/report style will be in NeurIPS format.

 

Selected Focus in Fall 2019

 

Student Learning Outcomes

 

Prerequisites

CAP 5610 (Machine Learning or C.I.)

Python

 

Textbook       

No textbook---students will read a selection of conference papers.

Background References:

 

The use of Newton Server at UCF --> See Paul's Lecture here

 

Syllabus & Grading Scheme

30% total

Paper Presentations (10%) + 1-page Report  (20%): Present one recent paper in class, chosen from 25 listed papers (I will distribute the papers myself). Presentation format will be discussed at the class (and below link gives the guideline too). It should include strengths and weaknesses of the method, as well as relevant work, comparisons, and the key ideas in discussion section to improve the paper with presenter’s own ideas. Paper presentation guideline is here.  

PAPER LIST IS HERE! 

Each student writes a report before the presentation (see guideline above), and upload it into the web course prior to paper discussion. At least 20 reports are expected (from 25 paper presentations). Each report is 1 point.

 

10% Total

- Mini-project #1 (10%)  Use Google’s deeplab v3++ to segment (semantic) MSCOCO database, obtain same/similar results with the state of the art, and modify the algorithm according to your own creativeness to improve the results via efficiency, accuracy, or both. Results should be written in a report NeurIPS paper style, but half in size (3-4 pages + additional pages for references).

 

20% Total

- Mini-project #2 (20%)  Implement the following paper

Title: "Self-Supervised Representation Learning by Rotation Feature Decoupling"
Authors: Zeyu Feng, Chang Xu, Dacheng Tao
Institution: UBTECH Sydney AI Centre, School of Computer Science, FEIT, University of Sydney, Australia
Code: 
https://github.com/philiptheother/FeatureDecoupling
Link: 
pdf and supp

After reproducing the paper with the available code, consider this method as baseline, now try to improve the method with your creativity.Contributions can be as little as making the network deeper to changing the architecture using dense modules etc. or it can be novelties in the method itself (how to deal with rotations etc which are mentioned in the paper).  Results should be written in a report NeurIPS paper style, but half in size (3-4 pages + additional pages for references).

 

40% Total

- Course Project (40%): Open images 2019 – visual relationship (detect pair of objects in particular relationship)

https://www.kaggle.com/c/open-images-2019-visual-relationship

Open Images is a collaborative release of ~9 million images annotated with image-level labels, object bounding boxes, object segmentation masks, and visual relationships. This uniquely large and diverse dataset is designed to spur state of the art advances in analyzing and understanding images. This year’s Open Images V5 release enabled the second Open Images Challenge to include the 3 tracks, you are supposed to work on Visual relationship detection track for detecting pairs of objects in particular relations, also relaunched from 2018. Please follow the guidelines, attend the contest (not obligatory), record your performance, write a paper comparable to NeurIPS style and quality, and include all details of the method you propose, and results in the paper. Discuss the preliminary version of the method with the professor and get some hints. Meet the professor at your earlier convenienceto discuss ideas, finalize the method and avoid any delays.

 

 

LETTER GRADES:  95-100 (A), 90-94 (A-), 85-89 (B+), 80-84 (B), 75-79 (B-), 70-74 (C+), 65-69 (C), 60-64 (C-), 50-59 (D), 0-49 (F), (based on agreements by students, we can change the grading into a curve when letter grades are not satisfactory (i.e., few reaches A)

 

COLLABORATION POLICY
Collaboration on assignments is encouraged at the level of sharing ideas and technical conversation only. Please write your own code. Students are expected to abide by UCF Golden Rule.

                                   

Contact

[Ulas Bagci in 2015]
Ulas Bagci in 2009
Name: Ulas Bagci
Email:
URL: http://www.cs.ucf.edu/~bagci
Work number: (+1) 407-823-1047
Fax number: (+1) 407-823-0594
CRCV Assistant: Cherry Place
Mailing address: Dr. Ulas Bagci
Center for Research in Computer Vision (CRCV)
4328 Scorpius Street, HEC 221, UCF

Orlando, Florida 32816, USA.

Last updated September, 2019 by Ulas Bagci.