Lectures

You can download the lectures here. We will try to upload lectures prior to their corresponding classes.

• (dl4cv-1) Image Classification
  tl;dr: Image Classification, a fundamental task of computer vision and simple algorithms to solve it.
  [slides]
  
• (dl4cv-2) Optimization
  tl;dr: Process of finding the best parameters.
  [slides]
  
• (dl4cv-3) Neural Networks - Perceptron
  tl;dr: Basic Artificial Neuron: MP neuron and Perceptron.
  [perceptron-code] [slides]
  

  Suggested Readings:

  • Perceptron convergence proof
• (dl4cv-4) Neural Networks - MLP
  tl;dr: Multi-Layered Network of Perceptrons.
  [perceptron-code] [slides]
  
• (dl4cv-5) Backpropagation
  tl;dr: Elegant technique that implements the gradient descent algorithm for the neural network training.
  [Tensor-basics-code] [Autograd-example-code] [slides]
  

  Suggested Readings:

  • 5min-Tensor-basics
• (dl4cv-6) Building blocks of CNNs
  tl;dr: Modules that constitute a Convolutional Neural Network (CNN)
  [slides]
  

  Suggested Readings:

  • Simple CNN training
• (dl4cv-7) CNN Architectures
  tl;dr: Evolution of the design principles and the resulting CNN architectures.
  [slides]
  

  Suggested Readings:

  • Analysis of DNN models for practical applications
  • Aggregated Residual Transformations for Deep Neural Networks
  • Trimps-Soushen-1, Shao et al. ILSVRC 2016
  • Trimps-Soushen-2, Shao et al. ILSVRC 2016
  • Squeeze-and-Excitation Networks
  • Densely Connected Convolutional Networks
  • MobileNets
  • ShuffleNets
  • NAS
• (dl4cv-8) Training DNNs
  tl;dr: Important aspects of training deep neural networks.
  [slides]
  

  Suggested Readings:

  • Xavier Initialization
  • Fixup Initialization: Residual learning without Normalization
  • Dropout by Srivastave et al.
• (dl4cv-9) Training DNNs-II
  tl;dr: Some more important aspects of training deep neural networks.
  [sgd_update_rules.gif] [slides]
  

  Suggested Readings:

  • Optimization update rules by Sebastian Ruder
  • SGD with Momentum, I Sutskever et al. ICML 2013
  • Ada Grad, Duchi et al. JMLR, 2011
  • Adam, Kingma et al. ICLR 2015
  • Some insgihts about lr in dl
  • Random search for hyper-parameter selection, Bergstra, JMLR 2012
• (dl4cv-10) RNNs
  tl;dr: Beyond the feed-forward neural nets, processing sequential data!
  [Sample-sequential-task] [slides]
  

  Suggested Readings:

  • LSTM Hochreiter et al. 1997
  • Sequence to sequence learning by Sutskever et al. 2014
  • Neural Machine Translation with aligning by Bahdanau et al. ICLR 2015
• (dl4cv-11) Attention
  tl;dr: Attention in sequence-to-sequence tasks using RNNs.
  [slides]
  

  Suggested Readings:

  • Sequence to sequence learning by Sutskever et al. NeurIPS 2014
  • Neural Machine Translation with aligning by Bahdanau et al. ICLR 2015
  • Show Attend and Tell by Xu et al. 2015
• (dl4cv-12) Word Embeddings
  tl;dr: Representing words in the NLP reltaed tasks
  [slides]
  
• (dl4cv-13) Visualizing and Understanding CNNs
  tl;dr: What do the CNNs learn? Why the predict what they predict?
  [slides]
  

  Suggested Readings:

  • Rich Feature Hierarchies by Girshick et al. CVPR 2014
  • Visualizing and Understanding CNNs by Zeiler and Fergus ECCV 2014
  • Deep Inside CNNs by Simonyan et al. ICLRW 2014
  • Class Activation Maps (CAM) by B Zhou et al. CVPR 2016
  • Grad-CAM by Selvaraju et al. NIPSW 2016, ICCV 2017
  • CNN-Fixations by Mopuri et al. TIP 2018
  • Texture Synthesis using CNNs by Gatys et al. NeurIPS 2015
  • DeepDream
  • Neural style transfer by Gatys et al. CVPR 2016
  • Understanding Deep Representations by Inverting them by A Mahendran et al. CVPR 2015
  • Inverting Visual Representations with CNNs by A Dosovitskiy et al. CVPR 2016-
• (dl4cv-14) Object Detection
  tl;dr: Task of simultaneously locating and classifying multiple pbjects present in an image using CNNs.
  [slides]
  

  Suggested Readings:

  • R-CNN by Girshick et al. CVPR 2014
  • Fast R-CNN by Girshick ICCV 2015
  • Faster R-CNN by Ren et al. NeurIPS 2015
  • SSD by Liu et al. ECCV 2016
  • YOLO by Redmon et al. CVPR 2016
  • Comparison of CNN object detectors by Huang et al. CVPR 2017
  • CornerNet by Law et al. ECCV 2018
• (dl4cv-15) Semantic Segmentation
  tl;dr: Objective is to label each pixel present in an image using CNNs.
  [slides]
  

  Suggested Readings:

  • FCN by Long et al. CVPR 2015
  • Deconvolution for semantic segmentation by Noh et al. ICCV 2015
  • Transposed Convolutions
• (dl4cv-16) Video Classification
  tl;dr: Recognize the actions in videos using CNNs.
  [slides]
  

  Suggested Readings:

• (dl4cv-17) Generative Models
  tl;dr: ML models that understand and model the data.
  [slides]
  

  Suggested Readings:

  • Pixel RNN, Google DeepMind, ICML 2016
  • Pixel CNN, Google DeepMind, NeurIPS 2016
• (dl4cv-17a) Autoencoders
  tl;dr: Neural Networks that encode unlabeled data into lower dimensional subspaces driven by the reconstruction objective.
  [slides]
  
• (dl4cv-17b) Variational Autoencoder
  tl;dr: Stochastic modules of an Autoencoder makes it a generative model.
  [slides]
  

  Suggested Readings:

  • Auto-Encoding Variational Bayes, ICLR 2014
  • Blogpost on VAE
• (dl4cv-17c) Generative Adversarial Networks
  tl;dr: Generative models with implicit density modeling for generating samples that resemble real data.
  [example-code] [slides]
  

  Suggested Readings:

  • GANs, Goodfellow et al, NIPS 2014
  • GAN Zoo
• (dl4cv-18) Adversarial Images
  tl;dr: Inputs that are crafted (via adding a special noise) to fool the trained DL systems.
  [slides]
  
• (dl4cv-19) Learning Efficient DL Models (Compressing the Models)
  tl;dr: Making the power hungry and huge models light-weight.
  [slides]
  
• (dl4cv-20) Attetion++
  tl;dr: Attention is all that we need?!
  [slides]
  
• (dl4cv-0) Introduction
  tl;dr: Introduction to DL4CV and logistics of this course.
  [slides]
  

  Suggested Readings:

  • Brief history of CV by Rostyslav Demush