Deep Learning

Columbia University - Fall 2017

Class is held in Mudd 1024, Mon and Wed 7:10-8:25pm

Office hours:

Monday 4:00-6:00pm, CEPSR 620: Lecturer, Iddo Drori

Tuesday 10:30-12:30, TA rm Mudd 1st floor: Course Assistant, Kartikeya Upasani

Wednesday 10:30-12:30, TA rm Mudd 1st floor: Course Assistant, Akshay Khatri

Thursday 10:30-12:30, TA rm Mudd 1st floor: Course Assistant, Ching-Hui Hsu

Friday 10:30-12:30, TA rm Mudd 1st floor: Course Assistant, Surabhi Bhargava

Lecture 1 (Wednesday, September 6): Introduction

Lecture 2 (Monday, September 11): Backpropagation
Forward and backward propagation, chain rule, computation graph, activation functions, loss functions, update, gradient descent, stochastic gradient descent.

Lecture 3 (Wednesday, September 13): Backpropagation
Transfer and multi-task learning, gradient checking, learning rate, minibatch.

Lecture 4 (Monday, September 18): Regularization
Bias and variance, norms, ridge regression, Lasso, regularized cost functions, dropout, data augmentation, input distortion, normalization.

Lecture 5 (Wednesday, September 20): Optimization
Gradient descent, SGD, momentum, Nesterov, Adagrad, RMSprop, Adam.

Lecture 6 (Monday, September 25): Game playing

Lecture 7 (Wednesday, September 27): Optimization
Newton's method, Conjugate Gradients, L-BFGS.

Lecture 8 (Monday, October 2): TensorFlow framework
Computation graph, session, placeholders, variables, estimators, TensorBoard.

Lecture 9 (Wednesday, October 4): Convolutional neural networks
Convolution, convolution layer, max pooling, ImageNet, AlexNet, VGG, Inception.

Lecture 10 (Monday, October 9): Information of deep neural networks
Markov matrices, Markov chain, mutual information, information of deep neural networks.

Lecture 11 (Wednesday, October 11): Convolutional neural networks
CNN architectures, ResNets, DenseNets, TensorFlow layers, applications: classification, detection, segmentation, pose estimation, completion, synthesis.

Lecture 12 (Monday, October 16): Recurrent neural networks
RNN's, architectures, deep, bidirectional, conditioned, recursive, backpropogation through time (BPTT).

Lecture 13 (Wednesday, October 18): Recurrent neural networks
Long short-term memory (LSTM), gated recurrent unit (GRU).

Lecture 14 (Monday, October 23): Reinforcement learning
Markov decision process (MDP), policy, value functions, model, Bellman equations, policy evaluation, policy iteration, value iteration.

Lecture 15 (Wednesday, October 25): Reinforcement learning
Dynamic programming, model-free prediction and control, Monte-Carlo methods, temporal-difference learning.

Lecture 16 (Monday, October 30): Deep reinforcement learning
Sarsa, Value-based deep RL, Q-Learning, Deep Q-Networks playing Atari.

Lecture 17 (Wednesday, November 1): Deep reinforcement learning
ResNet, Monte Carlo Tree Search, self-play, AlphaGo Zero.

Academic Holiday (Monday, November 6)

Lecture 18 (Wednesday, November 8): Unsupervised learning, deep generative models
Variational Auto-Encoders (VAE), generative adversarial networks (GAN), the GAN zoo, image synthesis and completion.

Lecture 19 (Monday, November 13): Visualizing deep neural networks, neural style transfer

Lecture 20 (Wednesday, November 15): Applications
Face recognition, self driving cars.

Lecture 21 (Monday, November 20): Adversarial examples, learning to learn and to write programs
Fast gradient sign method, universal perturbations, transfer attack, data driven discovery of models (D3M), DeepCoder.

Academic Holiday (Wednesday, November 22)

Thanksgiving Day (Thursday, November 23)

Lecture 22 (Monday, November 27): Deep probabilistic programming
Probabilistic programming, TensorFlow and BayesFlow, PyTorch and Pyro.

Lecture 23 (Wednesday, November 29): Projects fast forward

Lecture 24 (Monday, December 4): Project presentations
7:10-7:20 Wide and Deep Learning for Music Recommendation
7:20-7:30 Sarcoma Segmentation using Cascaded CNNs
7:30-7:40 KKBox Music Recommendation Challenge
7:40-7:50 Exploring Capsule Networks Applied to CIFAR Data
7:50-8:00 Learning to Optimize in the Information Plane
8:00-8:10 3D GAN for Novel Object Synthesis
8:10-8:20 Whale Recognition using Recurrent Attention CNNs
8:20-8:30 Drone Tracking and Navigation with Deep RL
8:30-8:40 KKBox Music Recommendation Challenge
8:40-8:50 Clustering of High Dimensional Mortgage-Backed Securities
8:50-9:00 Frame Rate Effects on Video Captioning of MSR-VTT Data

Lecture 25 (Wednesday, December 6): Project presentations
7:10-7:20 Playing Atari Games: Exploring the Dueling DQN Architecture
7:20-7:30 Transfer learning with LSTM on Amazon and Yelp Datasets
7:30-7:40 Scene Parsing using Semantic Labeling
7:40-7:50 Stock Performance Prediction Using News Sentiment Analysis
7:50-8:00 Question Answering and Reading Comprehension
8:00-8:10 Image Synthesis
8:10-8:20 Break, refreshments and pizza
8:20-8:30 Image Colorization
8:30-8:40 Image Captioning
8:40-8:50 Deep Cooperative Neural Network for Movie Rating
8:50-9:00 Learning to Optimize in the Information Plane
9:00-9:10 Activity Classification in Video
9:10-9:20 Overlapping Object Segmentation

Lecture 26 (Monday, December 11): Project presentations
7:10-7:20 Classification of Normal vs. Abnormal Heart Sounds
7:20-7:30 Music Genre Prediction
7:30-7:40 Automatic Image Colorization
7:40-7:50 3D GAN Object Synthesis
7:50-8:00 Learning to Learn Data Tasks
8:00-8:10 AskAway: Visual Question Answering with Attention
8:10-8:20 Break: 🥤 🍕 and αØ
8:20-8:30 Unpaired Image-to-Image Translation for Generating Faces
8:30-8:40 Study of Models for Image Classification
8:40-8:50 Image and Video Synthesis
8:50-9:00 Summarizing Documents using an RNN
9:00-9:10 Deep Autonomous Duckietown Navigation