Schedule

Introduction to Deep Learning, Applications of DL in vision, natural language processing (NLP), transportphenomena,fluidmechanics, hemicalengineering,materialscienceandhealth. CourseLogistics,LearningObjectives, Grading and Deadlines.

History and cognitive basis of neural computation, Perceptrons and Multi-layer Perceptrons.

Intro to Backpropagation, Comparisons of Activation function, Why ReLU?, Famous Cost functions in NN, NN terminology and definitions, NN Implementation Algorithm.

Forward Propagation, Partial Derivatives in Backpropagation, Examples of Back‐propagation, Advice onnumber of hidden layer selection.

Intro to SGD, Overfitting and regularization, Choosing Loss Function, Batch normalization, Adagrad,Adadelta, RMSProp, ADAM optimization, Downsides of SGD, Momentum SGD, Nesterov AcceleratedGradient (NAG), AdaGrad and adaptive learning rate

Motivations for Convolutional Neural Networks (CNN), Convolution Operation, Convolution Filter, LocalFeature Extraction, Volume Convolution, Strides and Filter size, Optimization of CNN

Recap on Convolution, Convolution Layers, RGB Channels, Padding, NFSP rule, Pooling (max, average, sum), Why Pooling?, Pooling Layers, Downsides of Pooling, Capsule Network, Architecture of Conv Layers, Conv and FC Layers Connection, Parameter sharing, Parameter (weight) Compression in ConvNet, CNN in vision, Challenges in vision (illumination, deflection, occlusions), Applications of CNNs (image classification, Segmentation, Labeling, Captioning, Counting, Pose Prediction, etc.), What if data is not image like?, Introduction to Graph conv, Zachary Karate Club Network, Social Network applications of Graph Conv, Fluid mechanic Example and idea of channels.

Training CNN/NN, Saturated Non-linearity, Internal Covariate Shift, Batch Normalization (BN), Benefits of BN, BN algorithm and shift/scale, Why BN works?, Hessian-Covariance accumulation, Xavier initialization, derivation of Xavier init., cost and accuracy per initialization scheme, Training with shared parameters Coding up CNN, MNIST CNN in Keras.

Node Representation Learning: Embedding Nodes, Shallow Encoding and its Limitations, Random Walk, Graph Embedding, Deep Encoders.

Motivation for Recurrent Neural Network (RNN), sequential models, simple RNN, hidden states, RNN examples, Cross-entropy loss for RNNs, Training RNNs, Vanishing and exploding gradient problem, Gradient clipping

Recap of RNN, Long Short Term Memory (LSTM): Gated Recurrent Unit (GRU), reset gate, update gate, Long-short term Memory (LSTM), input, forget, new and output gates, Sequence Prediction

The Simplest Autoencoder, Nonlinearity, Learning Dictionaries, NN used to perform linear or nonlinear PCA

Introduction to Generative Adversarial Networks (GANs), Generator and Discriminator Networks, Minmax game theory, GAN loss, Conditional GANs, Steady state, time-dependent heat transfer with GANs, Phase segregation and turbulence with GAN, L1 regularization, Image to image translation.

GAN Optimization Issues, Improved Techniques for Training GANs: Feature Matching, Minibatch Discrimination, Historical Averaging, Current Research.

Network Compression (pruning, knowledge distillation, parameter quantization, architecture design), Choosing a loss function, Regularization: Data Augmentation, Early Stopping, Dropout, Visualizing Neural Nets

Examples of Deep Learning Projects: Face Recognition, Learning Process: Feature Learning, Deep Learning Frameworks: Tensorflow, Keras, Pytorch.

Scaling Up, Orthogonalization, Hyperparameter Tuning, Batch Normalization, Softmax Regression, Error Analysis, Bias and Variance, Numerical approximation of gradients.

Density Estimation, Denoising, Missing Value Imputation, Sampling (synthesize), Markov Random Fields, Energy Representation, Restricted Boltzmann Machines, Deep Belief Networks

Basics of Reinforcement Learning (RL): Environment-agent-action-reward, Applications of RL, RL vs Un/Supervised Learning, Markov Decision Process (MDP), MDP tuples, transition probabilities, Policy, Value function, Bellman’s Equation. Bellman maximum value, Value iteration, Synchronous and asynchronous update, Policy iteration, Learning transition probabilities, Helicopter control and Inverted Pendulum dynamics, State-Action Learning.

Temporal-Difference Learning, Greedy Policy, Exploration vs. Exploitation, SARSA, On-policy vs.Off-policy, Instability,Q-Learning with NN

Actor Critic, Deep Q Network (DQN), Loss function in DQN, Deep Deterministic policy gradients (DDPG), Experience Replay, Hindsight Experience Replay in Robotics

Natural Policy Gradients, TRPO, PPO, ACKTR

Imitation Learning, GAN Imitation Learning, Transfer Learning

Model-based RL:Efficient Exploration (Billiards Example), Exploration by Planning, Maximum Entropy Inverse RL, DRL in Autonomous Vehicles and Controls