• Speaker: C. V. Jawahar
• Venue: Himalaya, Lecture Hall - 105
• Scope:
  
  1. Introduction to the summer school
  2. Introduction DL and CV
  3. Introduction to the topics and keywords;etc
• Keywords:machine learning, deep learning, big data, complex data, challenges, supervised learning, classifiers, support vector machines, kernel trick, neural networks, nonlinear, convex, non-convex, applications

• Speaker: Manish Gupta
• Venue: Himalaya, Lecture Hall - 105
• Objective:
  1. Overview of the topics
  2. Introduce many keywords
  3. Enables labs
• Keywords:biological inspiration, neurons, artificial neurons, nonlinearity, activation function, layers, hidden layers, loss function, back propagation, stochastic gradient descent, initialisation, regularization

• Speaker: Anoop M. Namboodiri
• Venue: Himalaya, Lecture Hall - 105
• Scope:
  Introduction to DL libraries
  1. Torch
  2. Theano
  3. Tensorflow
  4. Caffe

1. Getting Started
1. Basic Image Processing
2. Reading and Writing Data
3. Understanding basic data structures: Tensors, Blobs, Shared Variables
4. Running code on CPU/GPU
2. Implementation of multi layer perceptron
3. Walk through of basic CNN architecture.

• Speaker: R. Venkatesh Babu
• Venue: Himalaya, Lecture Hall - 105
• Objective:
  
  1. Introduction CNN, terminologies, popular architectures
  2. Features, Finetuning
  3. Loss functions
  4. Connect to lab
• Keywords:images, convolutions, pooling, fully connected layers, convolutional layers, invariances, gradients, hierarchical feature learning, deconvolution, weights, visualization, imagenet, lenet, alexnet, vggnet, googlenet, resnet

• Speaker:R. Venkatesh Babu
• Venue: Himalaya, Lecture Hall - 105
• Objective:
• Keywords: CNN Features and Fine tuning, Detection, Saliency, Visualization, Stereo, 3D and Optical Flow

• Speaker: C. V. Jawahar
• Venue: Himalaya, Lecture Hall - 105
• Objective: How to adapt the CNNs to solve a class of problems?
• Keywords: Alternate loss functions, Triplet Loss, Siamese Networks, Vanishing gradient problem

1. Train a network
2. Fine-tune a network
3. Loading weights in a network
4. Extracting features and classification
5. Extracting weights and visualize

• Speaker: Chetan Arora
• Venue: Himalaya, Lecture Hall - 105
• Objective:
  
  1. Intrdoduction to Reccurrent networks and learning
  2. LSTM and other popular architectures
  3. Some applications
• Keywords:sequences, RNN, back-propagation in time, unrolling, exploding and vanishing gradients, LSTM, GRU, BLSTM, applications

• Speaker: Vinay P Namboodiri
• Venue: Himalaya, Lecture Hall - 105
• Objective:
  1. Problem space
  2. Hybrid (CNN+RNN) solutions
  3. Captioning
  4. VQA
• Keywords:multimodal retrieval, cca, hybrid solutions, caption generation, visual question answering, attribute learning, word2vec, dcgan, applications

• Speaker: Chetan Arora
• Venue: Himalaya, Lecture Hall - 105
• Objective:

1. RNN
2. LSTM

• Speaker: Anoop M. Namboodiri
• Venue: Himalaya, Lecture Hall - 105
• Scope: unsupervised learning, autoencoders
• Keywords: autoencoders, unsupervised pre-training, denoising autoencoders, variatonal autoencoders, applications

• Speaker: Gaurav Sharma
• Venue: Himalaya, Lecture Hall - 105
• Topics
• Part I: Face
1. Siamese networks - Deep face, Deepid, VGG Face (Verification)
2. Networks for Age, Expressions etc.
3. Networks for fiducial points detection


• Part II: Human Actions and Activities
1. 3D Conv Nets for Human Actions
2. P-CNN for Actions
3. Two stream CNN
4. RNN variants for Actions


• Part III: Human Pose
1. Deep pose
2. Multi source deep learning for pose
3. DeepCur and DeeperCut (Multi-person pose)


• Part IV: Human Attributes and gestures
1. PANDA - Pose aligned networks for human attributes
2. Multimodal adaptive gestures (PAMI)

• Speaker: Gaurav Sharma
• Venue: Himalaya, Lecture Hall - 105
• Objective: Latent Ordinal Model for Video Face Analysis

1. Standard Auto-Encoders
2. Sparse Auto-Encoders
3. Denoising Auto-Encoders
4. Stacking of Auto-Encoders
5. Convolutional Auto-Encoders

• Speaker: Vineeth N Balasubramanian
• Venue: Himalaya, Lecture Hall - 105
• Objective:
  1. Associated mathematical and optimization issues
• Keywords: Gradient descent and variants, Backdrop review, Batch, stochastic, mini-batch gradient descent, Loss functions, Cross-entropy, Negative log-likelihood as the most general loss function, Overview of other loss functions (classification, regression, embedding), Challenges, Ill-conditioning, Multiple local optima, saddle points, plateaus, Vanishing/exploding gradients, Slow convergence, Choosing learning rate/parameters, Algorithms to address challenges, Momentum, Nesterov accelerated momentum, Adagrad, Adadelta, RMSProp, Adam, Which one to choose?, Advanced methods, Introduction to second-order methods: Newton method, Conjugate gradient, Natural gradient, Path-SGD, Performance guarantees,

• Speaker: Vineeth N Balasubramanian
• Venue: Himalaya, Lecture Hall - 105
• Scope:
  
  1. Normalization
  2. Non-linearities
• Keywords: Simple tricks, Early stopping, Data augmentation, Shuffling and Curriculum learning, Choosing activation functions and target values, Tanh, Softmax, ReLU, Leaky ReLU, Regularization, DropOut, DropConnect, MaxOut, Noise in Data, Label and Gradient, Weight initialization strategies, unsupervised pre-training, Transforming the inputs, Batch normalization

• Speaker: Shailesh Kumar
• Venue: Himalaya, Lecture Hall - 105
• Scope:
  

1. Batch Normalization
2. Dropout / MaxOut
3. Update rules
4. SGD, Adagrad, Adadelta, Adam, RMSProp
5. Weight Regularization
6. Weight Initialization
7. Data Augmentation

• Speaker: P. J. Narayanan
• Venue: Himalaya, Lecture Hall - 105
• Scope:
  
  1. GPU architectures
  2. GPU programming (CUDA)
  3. Practical issues in DL and CV
• Kegwords:GPU, parallel processing, throughput, Amadahl's law, graphics pipeline, graphics cards, CUDA, parallelizing neural networks, data parallelism, model parallelism, CUDNN, deep learning libraries, NVIDIA GPUs

• Speaker: Arjun Jain
• Venue: Himalaya, Lecture Hall - 105
• Objective:
1. State-of-the-art neural networks for solving vision tasks networks require millions of parameters and billions of arithmetic operations.
2. Look into optimizations which can enable running of such networks on low power, memory and compute capability devices such as mobile phones or cars.
• Keywords: Student-Teacher Networks, Deep Compression, SqueezeNet, HashedNets, BinaryConnect, Hardware Acceleration, Fixed-Point Optimization, Compute v/s Bandwidth Limitations, Power Considerations

1. Custom NN layers and functions
2. Introduction to CUDA
3. CUDA kernels
4. Interfacing CUDA kernels with DNN libraries

Abstract: Microsoft has been one of the early industry pioneers in the area of Deep Learning (DL) and has engaged with other pioneers from the academic world to create industry-scale successful products/applications in speech recognition as well as in speech translation, object recognition, automatic image captioning, natural language processing, multimodal processing, semantic modeling, web search, contextual entity search, ad selection, and big data analytics. Much of these successes are attributed to the availability of big datasets for training deep models, the powerful general-purpose GPU computing, and the innovations in deep learning architectures and algorithms. In this talk, I will provide an overview of the some of these exciting product areas at Microsoft which currently employ deep learning based technologies. I will also provide a peek into some start-of-the-art research areas, related to DL, currently being pursued at MS where we’re trying to push the envelope of this science.

Bio: Dr. Manoj Chinnakotla is a Senior Applied Scientist in the Relevance, Natural Language Understanding and Data Sciences group at Microsoft India, Hyderabad. He is also an adjunct faculty at the Language Technology Research Centre (LTRC), IIIT Hyderabad. He did his Ph.D in Information Retrieval (IR) and Natural Language Processing (NLP) from I.I.T. Bombay where he was an Infosys Research Fellow. He has published in several reputed conferences and journals in IR, NLP and ML such as SIGIR, ACL, IJCAI, WWW and ACM TALIP. He also serves as a reviewer and PC member for conferences - IJCAI, ICON, MIKE. His current research interests include Conversational Question Answering, Entity Mining and Machine Learning.

Speaker: Bharat Kaul

Bio: Bharat is an engineering manager and heads Intel Labs/Parallel Computing Lab-India. His research focus is on driving architectural performance leadership for Intel for multi-core/many architectures along with ensuring leadership in programmer efficiency. Bharat has ~20 years of industry experience and has been with Intel for 15+ years with experience in both technical and management leadership positions for product teams in several domains from data networking, communications, embedded systems and security architectures. Bharat received his B.E. in Electrical & Electronics from BITS Pilani.