Capsule Networks - overcoming limitations of Convolutional Neural Networks
SWAPAN JAIN (~swapan) |
Description:
This talk will cover understanding limitations of Convolutional NN in detecting images. After understanding this limitation, I will introduce the concept of capsules. This talk is highly inspired from the paper from Geoff hinton- Dynamic routing betwen Capsules-https://arxiv.org/pdf/1710.09829.pdf I will try to explain the process of training a multi layer capsule system on MNIST and comparing it with a convolutional net at recognizing highly overlapping images. I will use Tensorflow or Keras to show my demo Jupyter notebook. I will also discuss the limitations of capsules.
Prerequisites:
1.Linear Algebra 2.Probability and Statistics 3.Any Deep Learning library (Tensorflow/pytorch/Keras) 4.Deep Learning layers- Fully connected and Convolutional layers
Content URLs:
1.Understanding Convolutional Neural Networks - CS231n Stanford-http://cs231n.stanford.edu/ 2.Any Deep Learning Library preferably Tensorflow
Speaker Info:
Hi, I am Swapan Jain. After graduating in Computer Science from Delhi Technological University, I self studied AI by reading books,papers and taking courses online. I am a prospective graduate student from fall 2019.
Speaker Links:
I currently do not have open source contributions, but I will begin the blog from August. my twitter handle is @swapanj162. I will update about my blog or any project on my twitter.