11-785 Deep Learning

11-785 Introduction to Deep Learning

Fall 2019

Summer Homework to Prepare for Fall

The summer homeworks are optional and all can be done on a local machine, without any fancy equiptment. The first parts of each homework have an autograder, while the second part is meant to be evaluated by yourself. Only summer homework 2 part 2 has a Kaggle option.

Number	Part	Topics	Links
HW1	1 & 2	A Basic Introduction to Neural Networks	Write-up (PDF) Part 1, Code and Data Files (ZIP) Part 2, Code and Data Files (ZIP)
HW2	1	Feedforward Convolutional Neural Networks (CNNs)	Write-up (PDF) Code and Data Files (TAR)
	2	Image Recognition with CIFAR-10	Write-up, Code, and Data Imports (PDF) Kaggle Classification (LINK)
HW3	1	TBD	Write-up (PDF) Code and Data Files (TAR)
	2	TBD	Write-up (PDF) Code and Data Files (TAR)
HW4	1	TBD	Write-up (PDF) Code and Data Files (TAR)
	2	TBD	Write-up (PDF) Code and Data Files (TAR)

“Deep Learning” systems, typified by deep neural networks, are increasingly taking over all AI tasks, ranging from language understanding, and speech and image recognition, to machine translation, planning, and even game playing and autonomous driving. As a result, expertise in deep learning is fast changing from an esoteric desirable to a mandatory prerequisite in many advanced academic settings, and a large advantage in the industrial job market.

In this course we will learn about the basics of deep neural networks, and their applications to various AI tasks. By the end of the course, it is expected that students will have significant familiarity with the subject, and be able to apply Deep Learning to a variety of tasks. They will also be positioned to understand much of the current literature on the topic and extend their knowledge through further study.

You don't have to be a CMU student to follow the course

Petr Ermakov and Artem Trunov are mirroring the course at OpenDataScience (ODS.ai). The mirrored course follows the CMU course in its entirety, quizzes, homeworks, piazza, discussion boards and all, and runs roughtly 3 weeks behind the CMU schedule. . There are currently about 1300 students signed up for it. If you are interested in the full course experience, you too can sign up for it at this site.

If you are only interested in the lectures, you can watch them on the YouTube channel listed below.

Course description from student point of view

The course is well rounded in terms of concepts. It helps us understand the fundamentals of Deep Learning. The course starts off gradually with MLPs and it progresses into the more complicated concepts such as attention and sequence-to-sequence models. We get a complete hands on with PyTorch which is very important to implement Deep Learning models. As a student, you will learn the tools required for building Deep Learning models. The homeworks usually have 2 components which is Autolab and Kaggle. The Kaggle components allow us to explore multiple architectures and understand how to fine-tune and continuously improve models. The task for all the homeworks were similar and it was interesting to learn how the same task can be solved using multiple Deep Learning approaches. Overall, at the end of this course you will be confident enough to build and tune Deep Learning models.

Click here to read what students say about the previous edition of the course

Instructor: Bhiksha Raj

bhiksha@cs.cmu.edu

TAs:

Aime Cesaire Musangamfura : amusanga@andrew
Aishwarya Reganti : areganti@andrew
Chris Xia : chenxia@andrew
Ethan Xuanyue Yang : xuanyuey@andrew
Hanna Moazam :hmoazam@andrew
Hariharan Muralidharan : hmuralid@andrew
Joseph Konan : jkonan@andrew
Kangrui Ruan : kangruir@andrew
Liwei Cai : liweicai@andrew
Natnael Daba : ndaba@andrew
Parth Shah : pvshah@andrew
Ryan Brigden : rbrigden@andrew
Stephen Tsou : stsou@andrew
Shefali Garg : shefalig@andrew
Wenting Ye : wye2@andrew

* -- contingent on registration

Lecture: Monday and Wednesday, 9.00am-10.20am

Location: TBD

Recitation: Friday, 9.00am-10.20am

Location: TBD

Office hours:

Aime Cesaire Musangamfura : TBD
Aishwarya Reganti : TBD
Chris Xia : TBD
Hanna Moazam : TBD
Hariharan Muralidharan : TBD
Joseph Konan : TBD
Kangrui Ruan : TBD
Liwei Cai : TBD
Natnael Daba : TBD
Parth Shah : TBD
Ryan Brigden : TBD
Stephen Tsou : TBD
Shefali Garg : TBD
Wenting Ye : TBD
TA hours are also available here.

Prerequisites

We will be using one of several toolkits (the primary toolkit for recitations/instruction is PyTorch). The toolkits are largely programmed in Python. You will need to be able to program in at least one of these languages. Alternately, you will be responsible for finding and learning a toolkit that requires programming in a language you are comfortable with,
You will need familiarity with basic calculus (differentiation, chain rule), linear algebra and basic probability.

Units

This course is worth 12 units.

Course Work

Grading

Grading will be based on weekly quizzes (24%), homeworks (51%) and a course project (25%).

		Policy
Quizzes		There will be weekly quizzes. There are 14 quizzes in all. We will retain your best 12 scores. Quizzes will generally (but not always) be released on Friday and due 48 hours later. Quizzes are scored by the number of correct answers. Quizzes will be worth 24% of your overall score.
Assignments		There will be five assignments in all. Assignments will include autolab components, where you must complete designated tasks, and a kaggle component where you compete with your colleagues. Autolab components are scored according to the number of correctly completed parts. We will post performance cutoffs for A, B, C, D and F for Kaggle competitions. These will translate to scores of 100, 80, 60, 40 and 0 respectively. Scores will be interpolated linearly between these cutoffs. Assignments will have a “preliminary submission deadline”, an “on-time submission deadline” and a “late-submission deadline.” Early submission deadline: You are required to make at least one submission to Kaggle by this deadline. People who miss this deadline will automatically lose 10% of subsequent marks they may get on the homework. This is intended to encourage students to begin working on their assignments early. On-time deadline: People who submit by this deadline are eligible for up to five bonus points. These points will be computed by interpolation between the A cutoff and the highest performance obtained for the HW. The highest performance will get 105. Late deadline: People who submit after the on-time deadline can still submit until the late deadline. There is a 10% penalty applied to your final score, for submitting late. Slack days: Everyone gets up to 7 slack days, which they can distribute across all their homeworks. Once you use up your slack days you will fall into the late-submission category by default. Slack days are accumulated over all parts of all homeworks, except HW0, to which no slack applies. Kaggle scoring: We will use max(max(on-time score), max(slack-day score), .0.9max(late-submission score))* as your final score for the HW. If this happens to be a slack-days submission, slack days corresponding to the selected submission will be counted. Assignments carry 51% of your total score. HW0 is worth 1%, while each of the subsequent four are worth 12.5%.
Project		All students are required to do a course project. The project is worth 25% of your grade
Final grade		The end-of-term grade is curved. Your overall grade will depend on your performance relative to your classmates.
Pass/Fail		Students registered for pass/fail must complete all quizzes, HWs and the project. A grade equivalent to B- is required to pass the course.
Auditing		Auditors are not required to complete the course project, but must complete all quizzes and homeworks. We encourage doing a course project regardless.

Books

The course will not follow a specific book, but will draw from a number of sources. We list relevant books at the end of this page. We will also put up links to relevant reading material for each class. Students are expected to familiarize themselves with the material before the class. The readings will sometimes be arcane and difficult to understand; if so, do not worry, we will present simpler explanations in class.

Discussion board: Piazza

We will use Piazza for discussions. Here is the link. You should be automatically signed up if you're enrolled at the start of the semester. If not, please sign up.

You can also find a nice catalog of models that are current in the literature here. We expect that you will be in a position to interpret, if not fully understand many of the architectures on the wiki and the catalog by the end of the course.

Kaggle

Kaggle is a popular data science platform where visitors compete to produce the best model for learning or analyzing a data set.

For assignments you will be submitting your evaluation results to a Kaggle leaderboard.

Videos

All recitations and lectures will be recorded and uploaded to Youtube. Here is a link to the Youtube channel. Links to individual lectures and recitations will also be posted below as they are uploaded. All videos for the Fall 2019 edition are tagged “F19”. CMU students can also access the videos on Panopto from this link.

Academic Integrity

You are expected to comply with the University Policy on Academic Integrity and Plagiarism.

You are allowed to talk with / work with other students on homework assignments
You can share ideas but not code, you should submit your own code

Your course instructor reserves the right to determine an appropriate penalty based on the violation of academic dishonesty that occurs. Violations of the university policy can result in severe penalties including failing this course and possible expulsion from Carnegie Mellon University. If you have any questions about this policy and any work you are doing in the course, please feel free to contact your instructor for help.

Tentative Schedule

Lecture	Date	Topics	Lecture notes/Slides	Additional readings, if any	Quizzes/Assignments
0	-	Course logistics	Slides
1	August 26	Introduction to deep learning Course logistics History and cognitive basis of neural computation. The perceptron / multi-layer perceptron	Slides
2	August 28	The neural net as a universal approximator	Slides	Approximation capabilities of multilayer feedforward networks Multilayer Feedforward Networks are Universal Approximators
3	September 2	Training a neural network Perceptron learning rule Empirical Risk Minimization Optimization by gradient descent	Slides	Understanding the difficulty of training deep feedforward neural networks	..
4	September 4	Back propagation Calculus of back propagation	Slides
5	September 9	Convergence in neural networks Rates of convergence Loss surfaces Learning rates, and optimization methods RMSProp, Adagrad, Momentum	Slides	Back propagation fails to separate where perceptrons succeed
6	September 11	Stochastic gradient descent Optimization Acceleration Overfitting and regularization Tricks of the trade: Choosing a divergence (loss) function Batch normalization Dropout	Slides	Decoupled Weight Decay Regularization On the Convergence of Adam and Beyond SGD Learns Over-parameterized Networks that Provably Generalize on Linearly Separable Data
7	September 16	Optimization Generalization	Slides
8	September 18	Convolutional Neural Networks (CNNs) Weights as templates Translation invariance Training with shared parameters Arriving at the convolutional model	Slides	MobileNetV2: Inverted Residuals and Linear Bottlenecks Deep Residual Learning for Image Recognition Densely Connected Convolutional Networks
9	September 23	Convolutional Neural Networks Models of vision Neocognitron Mathematical details of CNNs	Slides		.
10	September 25	Backpropagation through CNNs Increasing output map size Transform invariance Alexnet, Inception, VGG	Slides	ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices MobileNetV2: Inverted Residuals and Linear Bottlenecks
11	September 30	Recurrent Neural Networks (RNNs) Modeling series Back propagation through time (BPTT) Bidirectional RNNs	Slides
12	October 2	Stability Exploding/vanishing gradients Long Short-Term Memory Units (LSTMs) and variants	Slides	How to compute a derivative
13	October 7	Cascade Correlation Nets by Scott Fahlman	Slides	Cascade-Correlation
14	October 9	Continual Learning by Pulkit Agarwal	Slides	Superposition of many models into one
15	October 14	Sequence To Sequence Modeling Connectionist Temporal Classification (CTC)	Slides
16	October 16	Connectionist Temporal Classification (CTC)	Slides
17	October 21	Attention Models	Slides		.
18	October 23	What do networks learn Autoencoders and dimensionality reduction	Slides
19	October 28	Hopfield Networks	Slides
20	October 30	Boltzmann Machines	Slides
21	November 4	Restricted Boltzman Machines (RBMs) Deep Boltzman Machines (DBMs)	Slides		.
22	November 6	Linear Generative Models Factor analysis EM
23	November 11	Generative Adversarial Networks (GANs) Part 1 Non-linear generators	Slides
24	November 13	Generative Adversarial Networks (GANs) Part 2	Slides
25	November 18	Variational autoencoders
26	November 20	Reinforcement Learning Part 1 Markov Process
27	November 25	Reinforcement Learning Part 2 Value and Policy Iterations
28	November 27	Reinforcement Learning Part 3 TD Learning SARSA
29	Decmeber 2	Q Learning Deep Q Learning			.
30	December 4	Guest lecture on biological models of cognition by Mike Tarr

Tentative Schedule of Recitations (Note: dates may shift)

Recitation	Date	Topics	Lecture notes/Slides	Notebook	Videos	Instructor
0 - Part 1	August 15	Python coding for the deep learning student			TBD
0 - Part 2	August 15	Python coding for the deep learning student		Notebook		TBD
1	August 30	Amazon Web Services (AWS)	Slides			Parth Shah, Kangrui Ruan
2	September 6	Your First Deep Learning Code	Slides			TBD
3	September 13	Efficient Deep Learning/Optimization Methods	Slides	Notebook		TBD
4	September 20	Debugging and Visualization	Slides	Notebook		TBD
5	September 27	Convolutional Neural Networks	Slides	Notebook		TBD
6	October 4	CNNs: HW2	Slides	Notebook		TBD
7	October 11	Recurrent Neural Networks	Slides	Notebook		TBD
8	October 18	RNN: CTC	Slides	Notebook		TBD
9	October 25	Attention	Slides	Notebook		TBD
10	November 1	Variation Auto Encoders	Slides			TBD
11	November 8	Attention	Slides		Video	TBD
12	November 15	GANs				TBD
13	November 29	Reinforcement Learning				TBD

Homeworks

Most homeworks require submissions to autolab. If you are an autolab novice here is an “autolab for dummies” document to help you.

Number	Part	Topics	Early-submission deadline	Links
HW0	-	Python coding for DL	none	pdf html
HW1	1	An Introduction to Neural Networks	-	pdf html pdf (summer version) zip (summer version)
	2	Frame level classification of speech		Kaggle html zip (summer version)
HW2	1	CNN	-	pdf pdf (summer version) tar (summer version)
	2	Face Classification/Verification via CNN		pdf Classification Kaggle Verification Kaggle pdf (summer version) Kaggle (summer version)
HW3	1	GRU	-	pdf
	2	Utterance to Phoneme Mapping		pdf Kaggle Slack Kaggle Code Submission Form
Hw4	1	Language Modeling using RNNs	-	pdf
	2	Attention		pdf Kaggle
Project				Template