CS 342 - Neural networks - Fall 2018
meets
TTh 10am-11am UTC 3.124
Fr 10am-11am / 11pm-noon ECJ 1.204
instructor Philipp Krähenbühl ( philkr (at) utexas.edu )
office hours Tue 11am-12pm GDC 4.824
Please only visit during office hours or with appointment (or if I didn’t reply your email after the third try)
TA
Dian Chen ( dchen (at) cs.utexas.edu )
office hours Mon 11am-noon GDC 1.302 Desk 1
Ankur Garg ( ankgarg (at) cs.utexas.edu )
office hours Mon 4pm-5pm GDC 1.302 Desk 1
We will use piazza for questions and canvas homework.
Schedule
class section
Date | Topic | Concepts | Slides | Corresponding book chapters (optional; see below) | Notes and due dates |
---|---|---|---|---|---|
Aug 30 | Introduction | tensors layers | intro | Background: Ch 1-5 | |
Aug 31 | Installing pytorch and building the first compute graph | slides tutorial |
HW1 out | ||
Sep 4 | Simple networks | fully connected ReLU | slides exercise 1 solution | ||
Sep 6 | Output transformations and loss functions | sigmoid softmax log likelihood L1/L2 loss | slides | 5.9 | HW1 due 11:59pm solution |
Sep 7 | Logistic regression and MLP | slides Tutorial |
HW2 out | ||
Sep 11 | Back propagation | back prop | slides exercise 2 solution | Ch 6 | |
Sep 13 | Stochastic gradient descent | SGD momentum | slides | HW2 due 11:59pm solution | |
Sep 14 | Regression vs classification | slides | HW3 out | ||
Sep 18 | Activation functions + Initialization | LeakyReLU PReLU Gaussian Init Xavier Init | slides | Ch 6 | |
Sep 20 | Convolutions + Pooling | Conv2D Avg Pooling Max Pooling striding / padding | slides | 9.1-9.3 | HW3 due 11:59pm solution |
Sep 21 | ConvNets | slides tutorial |
HW4 out | ||
Sep 25 | Overfitting | early stopping data augmentation dropout parameter sharing ensembles | slides exercise 3 solution 3 | 7.4, 7.8, 7.9, 7.11, 7.12 | |
Sep 27 | Optimization tips and tricks | weight regularization batch normalization | slides exercise 4 solution 4 | 7.1-7.3 | HW4 due 11:59pm solution |
Sep 28 | Regularization and normalization | slides tutorial |
HW5 out | ||
Oct 2 | Popular Architectures I | LeNet AlexNet VGG auxillary losses | slides exercise 5 solution | ||
Oct 4 | Popular Architectures II | GoogLeNet ResNet DenseNet residual connections | slides | HW5 due 11:59pm solution | |
Oct 5 | ResNets | slides | HW6 out | ||
Oct 9 | Object detection | ROI Pooling R-CNN Fast R-CNN YOLO | slides exercise 6 solution | ||
Oct 11 | Segmentation | up-convolution FCN | slides | HW6 due 11:59pm solution | |
Oct 12 | Fully convolutional networks | slides tutorial |
HW7 out | ||
Oct 16 | Visualization and understanding | gradients and saliency | slides exercise 7 solution 7 | ||
Oct 18 | Generative models | reparametrization trick adversarial loss | slides | 20.10, 20.12, 20.13 | HW7 due 11:59pm solution |
Oct 19 | Upconvolution and deep rendering | slides tutorial |
HW8 out | ||
Oct 23 | Recurrent Network I | RNN LSTM GRU | slides exercise 8 solution 8 | 10.1-10.11 | |
Oct 25 | Recurrent Network II | Temporal CNNs | slides | HW8 due 11:59pm solution | |
Oct 26 | Future prediction | slides tutorial |
HW9 out | ||
Oct 30 | Deep Reinforcement Learning I | policy gradient | slides exercise 9 solution 9 | ||
Nov 1 | Deep Reinforcement Learning II | Q-learning actor-critic | slides | HW9 due 11:59pm solution1 solution2 | |
Nov 2 | Immitation learning | slides | HW10 out | ||
Nov 6 | Beyond RL I | gradient free optimization | slides exercise 10 solution 10 | ||
Nov 8 | Beyond RL II | imitation learning dagger direct future prediction | slides | HW10 due 11:59pm solution | |
Nov 9 | Gradient free optimization | slides | HW11 out | ||
Nov 13 | Embedding learning | triplet loss contrastive loss | slides exercise 11 solution 11 | ||
Nov 15 | Style transfer | gram statistic | slides | ||
Nov 16 | Driving in SuperTuxKart | HW11 due 11:59pm Final project out |
|||
Nov 20 | Language | slides | |||
Nov 22 | no class | ||||
Nov 23 | no class | ||||
Nov 27 | Adversarial examples | adversarial attack | slides | 7.13 | |
Nov 29 | Negative results | ||||
Nov 30 | Final project Q/A | ||||
Dec 4 | Final project flash presentations | Final project due 11:59pm | |||
Dec 6 | Final project competition | ||||
Dec 7 | Nagging TA for better grade | no class |
concepts: layers/architecture algorithm supervision applications
Prerequisites
- 311 or 311H - Discrete math for computer science (or equivalent)
- 343 or 363D - Artificial Intelligence or Statistical Learning and Data Mining
- proficiency in Python
- Bonus: numpy, scipy and matplotlib
Class overview
- The class meets three times a week, twice is a classroom setting, once in a lab setting
- In the classroom setting you’ll learn the theoretical and mathematical background on neural networks and deep learning
- In the lab you’ll learn to use tensorflow and implement the techniques learned in class
Goals of the class
After this class you should be able to
- Implement and train neural networks in tensorflow
- Have a basic understanding of the inner workings of neural networks
- Know several types of neural networks, including convolutional and recurrent neural networks
Grading
- 10% Quizzes (in class)
- 55% Homework
- 35% Final project
There will be in class quizzes throughout the semester (about one a week). You can work with your peers on those quizzes during class. Each quiz will be graded pass/fail.
There will be a homework assignment every week, but we will keep them small and manageable. Homework needs to be solved individually, if we detect duplicate solutions you will lose the full credit for the assignment. Most of the homework will be graded automatically, you’ll have access to a partial grader while you’re working on the assignment. We will use the same grader, but different test data for your final grade. You’ll submit your homework through canvas.
Late policy:
- 1 day: -25%
- 2 days: -50%
- 3+ days: -100%
Expected workload
Estimates of required effort to pass the class are:
- 3 hours per week: attend class
- 3 hours per week: background reading
- 3 hours per week: coding homework
Course material
The course should be self contained, but if you need additional reading material consult Deep learning, Goodfellow, Bengio and Courville 2016
www.deeplearningbook.org.
Notes
Syllabus subject to change.