# CS 395T - Deep learning seminar - Fall 2018

**meets** TTh 2:00 - 3:30pm in GDC 4.304

**instructor** Philipp Krähenbühl

email philkr (at) utexas.edu

office hours by appointment (send email)

**TA** TBD

email TBD

office TBD

TA hours TBD

Please use canvas for assignment questions.

### Prerequisites

- 391L - Intro Machine learning (or equivalent)
- 311 or 311H - Discrete math for computer science (or equivalent)
**proficiency in Python**, high level C++ understanding- All projects are in Python with Tensorflow, PyTorch or Theano as recommended deep learning backends. It is also recommended to familiarize yourself with numpy, scipy, scikit-learn and matplotlib as additional libraries.
- NOTE: It is possible to use other languages, but the course staff cannot provide support.

- Basic
**deep learning background**- You should be familiar with at least one deep learning package (Caffe, Tensorflow, Torch, Matconvnet, …). You should have
**trained at least one deep network**with one of these packages. I’d encourage the use of Tensorflow for projects (if you want help from the course staff), but it is not required.

- You should be familiar with at least one deep learning package (Caffe, Tensorflow, Torch, Matconvnet, …). You should have

### Class overview

- The class reads (rates and reviews) two recent research papers per class
- Each paper is presented by two students, one student defends the paper, one student criticizes it
- Each student submits a short written 4 point review for each paper (before class).
- Two projects in teams up to 3 students (teams can change between projects)
- Auditing allowed if there is space (no homework or presentation, but participation required)

This class covers advanced topics in deep learning, ranging from optimization to computer vision, computer graphics and unsupervised feature learning, and touches on deep language models, as well as deep learning for games. This is meant to be a very interactive class for upper level students (MS or PhD). For every class we read two recent research papers (most no older than two years), which we will discuss in class.

#### Goals of the class

After this class you should be able to

- Read, rate and review deep learning papers
- Be able to present, defend and criticize a research paper
- Devise and execute a research project in deep learning (at the level of a top tier workshop publication: CVPR, ICCV, ICML, NIPS, ACL, SIGGRAPH or equivalent)

### Grading

- 5% written paper review
- 20% paper presentation
- 10% paper discussion (critique)
- 25% project 1 (10% presentation, 15% project)
- 40% project 2 (10% presentation, 30% project)

To map percentages to letter grades we use the following python script

```
def grade(p):
from math import floor
if p < 60: return 'F'
v = (100-p) / (10 + 1e-5)
return chr(ord('A')+int(v)) + ['+','','','-'][int((v-floor(v))*4)]
```

### Schedule

### Paper assignment

All papers are assigned by a Deep Networks based on student preferences. If you don’t like your paper, please blame deep learning, tensorflow or your preference list (instead of the instructor). The code for the assignment problem will be make available.

### Expected workload

Estimates of required effort to pass the class are:

- 2-4 hours per week reading / reviewing papers
- 7 hours per semester (1/2 hour per week) preparing paper presentations
- 2-10 hours per week of programming

### General tips

- Start the projects early
- most deep neural networks take 1 day to train on a GPU
- let us know early if you don’t have GPU access (first or second week)

- read your assigned papers early and prepare the slides/discussion
- a bad presentation/discussions will waste 30 min of your fellow students lives
- you have the option to get feedback on your slides ahead of time from the instructor

#### Tips for reading/reviewing a paper

- Just reading the paper is
**not**sufficient- Do
**more**than merely summarizing the paper - No paper is
*trivial*

- Do
- Question any decision and claim made by the authors
- It is the authors responsibility to convince you that their approach works better than prior (or simpler) alternatives
- If a claim is not backed by experiments or a citation (or backed by a wrong citation), you may assume it’s wrong

- Think about how this fits with other peoples findings
- Is there a larger theme across a series of papers?
- Does it contradict other papers you know?

- Use colored markers
- Mark important things in one color
- Mark things you disagree with (you think are wrong) in a different color

#### Tips for presentations

- Have a story
- Motivate well
- Provide context

- Make the presentation
**visual**- Use at least one picture/figure/graph per slide
- Walls of bulleted text are unacceptable
- Most mathematical concepts can first be expressed in a figure

- Feel free to make things interactive
- as long as it fits into the time budget

- Show your slides to the instructor a week before the presentation

### Notes

Syllabus subject to change.