Pythonic Vision With DenseNets

Manan Singh (~manan36)


16

Votes

Description:

Vision has been one of the hottest topics in the domain of deep learning and researchers have, over time, discovered some really great neural architectures to make progress with machine vision tasks.

In this talk, I'll be giving a quick interview of past remarking visual object recognition neural nets, with main focus on this year's CVPR Best Paper Award Winner DenseNets, it's comparison to previous networks (how actually the neural nets evolved), it's remarking features and it's applications, along with the code.

The talk will cover the following things:

  • A quick overview of CNNs
  • Hierarchy of greatest visual object recognition neural networks in history (LeNet < AlexNet < VggNet < GoogLeNet(Inception) < ResNet < DENSENET)
  • Comparison of architectures with pros / cons
  • Understanding DenseNet in detail (code implementation along side)
  • Memory efficient implementation for DenseNet
  • Remarking features of DenseNet
    • Feature Reuse
    • Tackling vanishing gradient problem
  • Wide-DenseNets (accuracy trade-off for time/memory efficiency)
  • Applications of DenseNet explained along with the code implementations
    • Object Classification
    • Object Detection (Deep Supervised Object Detection)
    • Segmentation with Fully-Convolutional DenseNets

The aim of talk is to

  • Explain the evolution of neural networks in the field of Vision and how DenseNets have brought a revolution and may serve to be a standard architecture in future.

Prerequisites:

  • Familiarity with python and a deep learning framework.
  • Knowledge of basics of neural networks and related terms.

Content URLs:

slides jupyter gist

Speaker Info:

  • I'm a pythonista and a deep learning enthusiast.
    • I'm a CS undergrad at NSIT, Delhi.
    • Interest domains: Computer Vision, Web Security, Cryptography, Mathematics.
    • Past-time hobby: Cracking math problems for fun, reading AI and security blogs.
    • My motive is to contribute to the progress in the AI field, and working on some missing theories in the field to develop a better understanding.

Speaker Links:

Manan Singh

Section: Others
Type: Talks
Target Audience: Intermediate
Last Updated: