CS 677: Deep learning
Summer 2019

Instructor: Usman Roshan
Office: GITC 4214B
Ph: 973-596-2872
Grader: Kalp Dalal, kdd32@njit.edu
Email: usman@njit.edu

Textbook: Not required
Grading: 40% programming assignments, 25% mid-term, 35% final exam
Course Overview: This course will cover deep learning and current topics in data science. We will begin with machine learning background and then move to CUDA and OpenCL languages for parallel programming on Graphics Processing Units (GPUs) followed by OpenMP for multi-core programming. We will then proceed to neural networks, machine learning for image recognition, convolutional filters for image recognition, convolutional neural networks, optimization algorithms to train such networks, adversarial attacks, and deep learning for text.

Course plan:

-->

Topic	Date	Notes
Basic machine learning and Python scikit-learn		Introduction Basic Unix command sheet Instructions for AFS login Basic machine learning background with Python scikit-learn Datasets Linear data Non-linear data Breast cancer (bc.train.0) (bc.test.0) Ionosphere (ion.train.0) (ion.test.0) Climate simulation (climate.train.0) (climate.test.0) Qsar (qsar.train.0) (qsar.test.0) Fifty two datasets selected from the UCI machine learning repository
Introduction to GPU computing		GPU coding (also see Cuda by Example by Kandrot and Sanders) Numpy tutorials Official Numpy tutorial External Numpy tutorial CUDA in Python Numba CUDAJIT in Anaconda PyCUDA (PyCUDA slides)
CUDA programming		Parallel chi-square 2-df test Chi-square 2-df test in parallel on a GPU Simulated GWAS Class labels for above data
CUDA programming		Parallel Chi-square 2-df test Assignment 1
OpenCL and OpenMP programming		CUDA to OpenCL slides libOpenCL.so (NVIDIA library file for OpenCL code) Chi2 opencl implementation OpenCL files CUDA to OpenMP slides OpenMP reference Chi2 openmp implementation Assignment 2
Neural networks		Multi-layer perceptrons Scikit-learn MLP code
Image classification		Image classification code Python Image Library
Convolutional kernels for images		Convolutional kernels
Convolutional neural networks		Convolutional neural network (Additional slides by Yunzhe Xue) Convolution and single layer neural networks objective and optimization Flower image classification with CNNs code
Convolutional neural networks		Convolution and single layer neural networks objective and optimization
Back propagation for single layer network with numpy		Assignment 3
Online lecture for backpropagation weight updates
Back propagation for single layer network with numpy, stochastic gradient descent		Optimization in neural networks Assignment 4
Back propagation for convolutional network with one convolutional layer followed by global average pooling		Assignment 5
Introduction to Keras
Keras and mid-term review		Mid-term review sheet Assignment 6
	07/11/2019	Mid-term exam
Convolutional neural networks in Keras: convolutional blocks and layering		Image classification code v2
Convolutional neural networks in Keras: normalization types and their effect		Batch normalization Batch normalization paper Group normalization paper How does batch normalization help optimization
Convolutional neural networks: gradient descent optimization. Large-scale image classification		Gradient descent optimization An overview of gradient descent optimization algorithms Curriculum learning CIFAR 10 CIFAR 100 STL 10 Mini ImageNet
More on training convolutional neural networks		On training deep networks The Loss Surfaces of Multilayer Networks Ant colony optimization for deep networks Simulated Annealing Algorithm for Deep Learning
Convolutional neural networks in Keras: pertrained models and transfer learning Convolutions and deconvolutions, temporal convolutional kernels Common architectures: ResNet, DenseNet, VGG		A guide to convolution arithmetic for deep learning Common architectures Transfer learning by Yunzhe Xue Pre-trained models in Keras
Continuation of above plus cross-entropy loss function vs. least squares		Assignment 7 Mini ImageNet in original form. Data are 256x256 images spread across different directories
Multi-path networks, data augmentation, time-series and sequence networks		An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling
Image classification on Kaggle datasets, self-designed networks, self-supervised learning		Assignment 8 Kaggle datasets: (a) Fruits (b) Flowers (c) Chest X-rays Exploring Randomly Wired Neural Networks for Image Recognition Automatically Designing CNN Architectures Using Genetic Algorithm for Image Classification UNSUPERVISED REPRESENTATION LEARNING BY PREDICTING IMAGE ROTATIONS Unsupervised Visual Representation Learning by Context Prediction Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
Data augmentation with generative adversarial networks		Data Augmentation Using GANs Transferring GANs: generating images from limited data Large Scale GAN Training for High Fidelity Natural Image Synthesis
U-networks		U-Net: Convolutional Networks for Biomedical Image Segmentation GitHub UNet in Keras
Image retrieval		Deep Learning for Image Retrieval: What Works and What Doesn't CNN Features off-the-shelf: an Astounding Baseline for Recognition Deep Learning for Content-Based Image Retrieval: A Comprehensive Study
Adversarial attacks, final exam review		Final exam review sheet
Deep learning with Pytorch		Flower classification in Pytorch
Text data		CNNs for text Word2Vec for word representations Liar Liar dataset Basic spam filtering