Lectures on Deep Learning

[sarthaksharma13]

Hi, Coming Thursday we would start with lectures on Deep Learning. The expectations is that theory uploaded here would be read before coming to the lectures. We would be covering 3 assignments (minimally) during the sessions which would last 5-7 days. For the Thursday (23 May 2019) session, please read the following before the class : Intro : http://neuralnetworksanddeeplearning.com/ Chap 1: http://neuralnetworksanddeeplearning.com/chap1.html Chap 2: http://neuralnetworksanddeeplearning.com/chap2.html

[talsperre]

For tomorrow's lecture, bring your own laptops with working internet connection (3G or 4G net). All tutorials will be done on Google Colab so internet connection is a must.

[sarthaksharma13]

Here are a few reads to be done for different optimization methods :

1. https://towardsdatascience.com/neural-network-optimization-algorithms-1a44c282f61d
2. http://fa.bianp.net/teaching/2018/eecs227at/

[sarthaksharma13]

Code for today's class (27/5/19) :

1. CIFAR10 : https://colab.research.google.com/drive/18dgQudsJr3t0NPdzaW0CUyLzDoTlc4ls
2. Finetune CIFAR 100 : https://colab.research.google.com/drive/10hmDESmHoxDl3bCnfjFTHHr766pbgeu8

[sarthaksharma13]

Task 1 : Aim : To get familiar with different network designs and hyperparameter search. Try methods to normalize the data. Have a look at different architectures like VGGNet, AlexNet, ResNet etc.

• Try min 2 network definitions with min 5 hyperparameters each. Try to push the test accuracy above 80%. Do this for both CIFAR10 and CIFAR100.

Task 2: Aim : To get familiar with performance of the network when variation in the dataset occurs. Try to reason about that using plots.

• Train a network for the standar MNIST dataset. Test it on the test split.
• Test the same network on the rotated MNIST dataset (w/o training on it) Report the performance.
• After this, report graph(s) that adds a certain percentage(0%,10%,20%..) of rotated dataset to the original dataset (X axis) and report performance on a test split consisting of only rotated, normal and a mixure of both normal and rotated dataset(Y axis).

Deadline for both the tasks : 3 June 11:59 PM.

[talsperre]

Read this lesson once before coming for tomorrow's lecture:

1. https://colah.github.io/posts/2015-08-Understanding-LSTMs/

[talsperre]

Bring your own laptops with working internet connection (3G or 4G net). All tutorials will be done on Google Colab so internet connection is a must.

[talsperre]

The notebook for todays class:

1. https://colab.research.google.com/drive/1HHAM10nxpxA4u0uKrVGZJl9EEPWpRydv

[AnirudhaRamesh]

Could we be given access to ADA anytime soon? Training our models for this assignment and beyond is taking very long with our local machines + colab, floydhub etc free tier platforms. Thanks in advance

[talsperre]

Send a mail to hpc@iiit.ac.in in the following format.

Please make an ADA account for me as I need it for pursuing my research. The details are 
given below:

Name: X
Research Center: RRC (Robotics Research Center)
Faculty Advisor: Professor Madhava Krishna
Preferred Login Id: X