Inkers is offering an external internship program where they teach the way neural networks work and based on the assignment solutions, surprise quiz and project, students would be taken into the internship offered by them.

This repository contains the materials and solutions to the assignment given for each session.

A brief overview of the sessions are as follows:

1. Session 1 - Basically the first class of the 6 sessions. Introduction to convolution operation, maxpooling, different use cases of deep learning.
2. Session 2 - It's an important session. In this session, topics such as similarity between algorithmic filters like gabor filter and filters learned by the network were discussed. Most importantly, backpropagation algorithmic was discussed for simple linear regression.
3. Session 3 - In Depth overview of convolution operation, different types of convolutions, dropouts, different architectural considerations for building a CNN were discussed.
4. Session 4 - More of a practical session on keras. Sequential and functional API of keras were discussed. Understand the writing of simple Deep Neural Network Code. Also discussed about DenseNet Architecture.
5. Session 5 - Session 5 and Session 6 are on advanced topics. There's no material on Session 5. Session 5 is combined with Session 6 material. This session discussed on YOLO and SSD algorithm for object recognition.
6. Session 6 - This session discussed on YOLO v2 and how to implement YOLO v2 by transfer learning.

Description for Assignments can be found out from Session files. All the Solutions for the assignments in this repository are my own.

In Order to run the ipython notebooks

1. Clone the Repository.
2. Use a google gmail account for google Colab, refer this link - How to setup Google Colab
3. Upload the notebooks to your google drive account
4. Double click on the notebooks, it automatically opens in colab or you have forgotten to add Colab as of one apps accessible to your drive.

Some of the Assignment Solutions take 8-10 hrs to run on the GPU. So, wait patiently. Also models with highest accuracy with the given number of epochs will be stored.

Session 6 assignment solution is taken from pavisj. Credits to pavisj

Update 1

Luckily I got selected for session 7, 8, 9 and 10.

7. Session 7 - Discussed in detail about Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM). The materials for this session are excellent videos and a blog- CS231n by Andrej Karpathy, Evolution: from vanilla RNN to GRU & LSTMs (How it works) [RU] and blog - Keras LSTM tutorial – How to easily build a powerful deep learning language model

Important Note - The 2nd video was inspired by this very good blog - Written Memories: Understanding, Deriving and Extending the LSTM. The slides to the video can be found here - link

8. Session 8 - Discussed in detail about GANs and intro to Reinforcement Learning (RL). The materials for this session are as follows:
   1. Progressive GANs - Link to video
   2. Tutorial on GANs by Jeremy Howard (Fast.ai) - Link
   3. Reinforcement Learning (Deep Q Learning) - Deep Q Learning for Video Games - The Math of Intelligence #9 by Siraj Raval
   4. Reinforcement Learning (Actor Critic Learning) - Actor Critic Algorithms by Siraj Raval

Additional Links and Tutorials

1. Notes on WGAN - Wasserstein GAN
2. Session 8 also discussed about TensorRT - Nvidia TensorRT
3. Intuitive RL: Intro to Advantage-Actor-Critic (A2C)
4. Simple Reinforcement Learning with Tensorflow Part 8: Asynchronous Actor-Critic Agents (A3C)

9.Session 9 - Unfortunately I couldn't attend session 9. Therefore, I collected all the possible materials and topics discussed from my batchmates. These were the topics discussed in session 9: 1. Embedding/Kernel Visualization and exploitation for different tasks - Watch this video ,Bolei Thesis Defense: Interpretable Representation Learning for Visual Intelligence. 2. Artistic Style Transfer for videos - 2 minute papers,Artistic Style Transfer For Videos | Two Minute Papers #68 3. Using few internal layers and embeddings can be used to do fast object tracking - Materials for this are yet to be updated.

10.Session 10 - The following concepts were covered during session 10.

1. One Shot Learning by Andrew NG
2. tSNE Visualizer
3. Reinforcement Learning Basics
4. Reinforcement Learning Survey