CS 294-131 Special Topics in Deep Learning Fall 2016

Instructor and co-Instructors:

Trevor Darrell, Sergey Levine, Dawn Song

Teaching Assistant:

Lectures:

Wed 10:00am-noon (First class starts on Aug 31)
location: SDH250

Course mailing list:

cs294-dl-f16@googlegroups.com
Please sign up for the course mailing list for future updates.

If you do not plan to take the class, but are interested in getting announcements about guest speakers in class, and more generally, deep learning talks at Berkeley, please sign up for the talk announcement mailing list for future announcements.

Please also sign up for our Piazza.

Syllabus:

Lecture 1 8/31/16	Kaiming He: Deep Learning Gets Way Deeper [slides, introductory slides] Reading: Deep Learning, by Yann LeCun, Yoshua Bengio, and Geoffrey Hinton.
Lecture 2 9/7/16	Ilya Sutskever: Why does deep learning actually work? Main readings: Explaining and Harnessing Adversarial Examples, by Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Qualitatively Characterizing Neural Network Optimization Problems, by Ian J. Goodfellow, Oriol Vinyals, and Andrew M. Saxe. Background readings: Algorithmic probability, Wikipedia Does Algorithmic Probability Solve the Problem of Induction?, by Ray Solomonoff A Theory of the Learnable, by Leslie Valiant On Small Depth Threshold Circuits, by Alexander A. Razborov (read table on last page)
Lecture 3 9/14/16	Wojciech Zaremba: Turing-complete neural-network based models [slides] Main readings: Extensions and Limitations of the Neural GPU, by Eric Price, Wojciech Zaremba, and Ilya Sutskever Learning Simple Algorithms from Examples, by Wojciech Zaremba, Tomas Mikolov, Armand Joulin, and Rob Fergus Background readings: Learning Efficient Algorithms with Hierarchical Attentive Memory, by Marcin Andrychowicz and Karol Kurach Neural Random-Access Machines, by Karol Kurach, Marcin Andrychowicz, and Ilya Sutskever Neural Programmer-Interpreters, by Scott Reed and Nando de Freitas
Lecture 4 9/21/16	Kunal Talwar: Deep Learning with Differential Privacy Main readings: Deep Learning with Differential Privacy by Abadi, Chu, Goofellow, McMahan, Mironov, Talwar and Zhang Chapters 1 and 2 of The Algorithmic Foundations of Differential Privacy by Cynthia Dwork and Aaron Roth Background readings: Chapter 3 of The Algorithmic Foundations of Differential Privacy Differentially Private Empirical Risk Minimization by Bassily, Smith and Thakurta (Sections 1 and 2) Analyze Gauss: Optimal bounds for privacy-preserving PCA by Dwork, Talwar, Thakurta and Zhang
Lecture 5 9/27/16 (Special date)	Yoshua Bengio Location: Wozniak Lounge (430-8 Soda Hall) Time: 12:30 pm
Lecture 6 9/28/16	Samy Bengio: Sequence models [slides] Main readings: Order Matters: Sequence to sequence for sets, by Oriol Vinyals, Samy Bengio, and Manjunath Kudlur Reward Augmented Maximum Likelihood for Neural Structured Prediction, by Mohammad Norouzi, Samy Bengio, Zhifeng Chen, Navdeep Jaitly, Mike Schuster, Yonghui Wu, and Dale Schuurmans Background readings: Sequence to Sequence Learning with Neural Networks, by Ilya Sutskever, Oriol Vinyals, and Quoc V. Le Pointer Networks, by Oriol Vinyals, Meire Fortunato, and Navdeep Jaitly Sequence Level Training with Recurrent Neural Networks, by Marc'Aurelio Ranzato, Sumit Chopra, Michael Auli, and Wojciech Zaremba
Lecture 7 10/5/16	Ian Goodfellow: Adversarial Examples and Adversarial Training [slides] Main readings: Explaining and Harnessing Adversarial Examples, by Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy Transferability in Machine Learning: from Phenomena to Black-Box Attacks using Adversarial Samples, by Nicolas Papernot, Patrick McDaniel, and Ian Goodfellow Background readings: Intriguing properties of neural networks, by Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus
Lecture 8 10/12/16	Oriol Vinyals: Sequences and one-shot learning Main readings: Matching Networks for One Shot Learning, by Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra One-shot Learning with Memory-Augmented Neural Networks, by Adam Santoro, Sergey Bartunov, Matthew Botvinick, Daan Wierstra, and Timothy Lillicrap
Lecture 9 10/19/16	Bill Dally: Hardware for Deep Learning [slides] Main readings: EIE: Efficient Inference Engine on Compressed Deep Neural Network, by Song Han, Xingyu Liu, Huizi Mao, Jing Pu, Ardavan Pedram, Mark A. Horowitz, and William J. Dally Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding, by Song Han, Huizi Mao, and William J. Dally Background readings: Deep learning with COTS HPC systems, by Adam Coates, Brody Huval, Tao Wang, David J. Wu, Andrew Y. Ng, and Bryan Cantanzaro
Lecture 10 10/26/16	Hugo Larochelle: Autoregressive Generative Models with Deep Learning Main readings: Neural Autoregressive Distribution Estimation, by Benigno Uria, Marc-Alexandre Côté, Karol Gregor, Iain Murray, and Hugo Larochelle Background readings: Deep Learning, Chapter 20 (Generative Models) Sections 20.1 to 20.4 (background on undirected graphical models) Sections 20.9 and 20.10.1 to 20.10.5 (background on directed graphical models)
Lecture 11 11/2/16	Jitendra Malik: Embodied Cognition: Phylogeny and Ontogeny Main readings: The Development of Embodied Cognition: Six Lessons from Babies, by Linda Smith and Michael Gasser Background readings: Embodied Cognition: Phylogeny and Ontogeny, by Claes von Hofsten and Kerstin Rosander
Lecture 12 11/9/16	Surya Ganguli: Deep learning theory and physics-inspired generative models Main readings: Exact solutions to the nonlinear dynamics of learning in deep linear neural networks, by Andrew M. Saxe, James L. McClelland, and Surya Ganguli Deep Unsupervised Learning using Nonequilibrium Thermodynamics, by Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, Surya Ganguli Background readings: Identifying and attacking the saddle point problem in high-dimensional non-convex optimization, by Yann Dauphin, Razvan Pascanu, Caglar Gulcehre, Kyunghyun Cho, Surya Ganguli, Yoshua Bengio Exponential expressivity in deep neural networks through transient chaos, by Ben Poole, Subhaneil Lahiri, Maithra Raghu, Jascha Sohl-Dickstein, and Surya Ganguli
Lecture 13 11/16/16	Timothy Lillicrap: Recent advances in deep reinforcement learning Main readings: Human-level control through deep reinforcement learning, by Volodymyr Mnih et al. Continuous control with deep reinforcement learning, by Timothy P. Lillicrap, Jonathan J. Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra Background readings: Chapter 3 of Reinforcement learning: An Introduction, by Richard G. Sutton and Andrew G. Barto High-dimensional continuous control using generalized advantage estimation, by John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel
Lecture 14 11/23/16	CANCELLED! No class 11/23 Diedrik Kingma: Generative models Main readings: Auto-Encoding Variational Bayes, by Diedrik P. Kingma and Max Welling Improved Variational Inference with Inverse Autoregressive Flow, by Diedrik P. Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling
Lecture 15 11/30/16	In-class project presentations

Course Description:

In recent years, deep learning has enabled huge progress in many domains including computer vision, speech, NLP, and robotics. It has become the leading solution for many tasks, from winning the ImageNet competition to winning at Go against a world champion. This class is designed to help students develop a deeper understanding of deep learning and explore new research directions and applications of deep learning. It assumes that students already have a basic understanding of deep learning. In particular, we will explore a selected list of new, cutting-edge topics in deep learning:

Security and privacy issues in deep learning. First, we will explore attack methods and defenses in the area of adversarial deep learning, where attackers can purposefully generate adversarial examples to fool state-of-the-art deep learning systems. Second, we will explore the area of privacy-preserving deep learning. A deep learning system trained over private data could memorize and leak private information undesirably. We will explore areas including model-inversion attacks and how to provide differential privacy guarantees for deep learning algorithms. Finally, we will explore the use of deep learning in security applications such as malware and fraud detection.
Novel application domains of deep learning, beyond the mainstays of computer vision and speech recognition. First, we will explore new techniques in deep reinforcement learning, involving both applications of reinforcement learning to traditionally supervised learning problems and applications of deep learning to tasks that involve decision making and control. Second, we will explore new domains at the intersection of deep learning and program synthesis and formal verification. We will also explore other new application domains such as using deep learning for graph analysis.
Recent advances in the theoretical and systems aspects of deep learning. First, we will cover the recent advances in generative models, including variational autoencoders and generative adversarial networks. Second, we will explore new theoretical advances in understanding deep learning such as the Deep Rendering Model. Third, we will explore new system and architectural advances in scaling up deep learning and new architectural designs.

Class Format and Project:

This is a lecture, discussion, and project oriented class. Each lecture will focus on one of the topics, including a survey of the state-of-the-art in the area and an in-depth discussion of the topic. Each week, students are expected to complete reading assignments before class and participate actively in class discussion.

Students will also form project groups and complete a research-quality class project. Groups will consist of one to three students.

Course #: COMPSCI 294-131 (Class #34939)

Deadlines:

Reading assignment deadlines:
- For students,
  - Submit questions by Sunday noon
  - Vote on the poll of discussion questions by Monday 11:59 pm
- For discussion leads,
  - Send form to collect questions from students by Friday 11:59 pm
  - Summarize questions proposed by students to form the poll and send it by Sunday 11:59 pm
  - Summarize the poll to generate a ranked & categorized discussion question list and send the list to teaching staff by Tuesday 7pm
Project deadlines:
- 9/21: project groups due
- 9/28: project proposal due
- 10/26: project milestone report due
- 11/30: in-class presentation
- 12/16: final project report due

Background reading:

Deep Learning, by Ian Goodfellow, Yoshua Bengio, and Aaron Courville

Grading:

20% class participation
35% weekly reading assignment
- 10% discussion leads
- 25% individual reading assignments
45% project

All information is tentative and subject to change.

CS 294-131: Special Topics in Deep Learning Fall, 2016