Pushing the boundaries of Computer Vision with Flux.jl

Computer Vision is an ever expanding domain and recently there has been a huge rise in the number and size of datasets. Flux.jl offers a very simple API to design and train very deep Neural Networks. Flux allows us to exploit the speed of Julia and hence models are trained very fast.

This talk is meant to give an in-depth introduction to the usage of Flux.jl in the domain of Computer Vision. The main motivation of the talk is how we can define complicated models in Flux and get state of the art performance with minimal effort.

Flux.jl is a Deep Learning library with full support of Automatic Differentiation written in pure Julia for optimal performance. Currently, it is a fully fledged library with most functionalities of any other traditional Deep Learning Library.

Currently, Flux is being mainly used in the domain of natural language processing, even Deep Reinforcement Learning. However, the domain of Computer Vision is widely unexplored. So the major focus of my project is to design the existing state-of-the-art architectures and try them on newer datasets.

 
1 favorite thumb_down thumb_up 4 comments visibility_off  Remove from Watchlist visibility  Add to Watchlist
 

Outline/Structure of the Talk

  1. Basic Introduction to Flux.jl [5 min] - Flux.jl is a Deep Learning library with full support of Automatic Differentiation written in pure Julia for optimal performance. Currently, it is a fully fledged library with most functionalities of any other traditional Deep Learning Library. In this part a brief introduction as to how Flux works will be given.
  2. Using Flux.jl in Image Classification Problems [5 min] - Presently, the Flux model zoo consists of State of the Art Imagenet models like VGGNets, Resnets, XceptionNet and so on. Work is currently underway to train the Inception V3 on the MIT Places 2 Dataset. Also, DenseNets which form the baseline model for the MURA dataset are also in the process of being added to the model zoo.
  3. Incorporating the models in model-zoo and Metalhead.jl to perform Object Detection [10 min] - Metalhead.jl is a library developed on top of Flux.jl and provides a simple API to use models that are pre-trained on Imagenet. These models form the base of DeepDream.jl and FastStyleTransfer.jl. DeepDream.jl is based on google’s deepdream and allows users to generate fancy dreams. Also, the base models in Metalhead.jl allows easy incorporation into object detection models like Faster R-CNN and Mask R-CNN.

Learning Outcome

  1. Get to know about Flux.jl
  2. Usage of Flux.jl in a major Machine Learning Project
  3. Knowledge about object detection and localization

Target Audience

Anyone interested in using Julia for their next AI task

Prerequisites for Attendees

There are no formal prerequisites.

However, prior knowledge to the following might be useful.

  1. Know the basics of Julia
  2. Fundamentals of Computer Vision
schedule Submitted 1 year ago

Public Feedback

comment Suggest improvements to the Speaker
  • Vishal Gokhale
    By Vishal Gokhale  ~  1 year ago
    reply Reply

    Thanks for the proposal, Avik !

    Can you please share links to videos of your prior talks to help the program committee assess your ability to present on the topic?

    • Avik Pal
      By Avik Pal  ~  11 months ago
      reply Reply

      Sorry for the late response, for some reason I didn't receive a notification. This is my first talk proposal so I don't have any prior videos available. 
      If you are interested in seeing the work you can have a look at

      1. https://github.com/avik-pal/MURA.jl

      2. https://github.com/avik-pal/model-zoo

      3. https://github.com/avik-pal/CNNVisualize.jl

      4. https://github.com/avik-pal/FastStyleTransfer.jl

      • Naresh Jain
        By Naresh Jain  ~  11 months ago
        reply Reply

        Hi Avik,

        We are convinced that you've done good work and hence we are considering this talk. However, having done good work and presenting the work at a conference are two different skills. We want to understand that you can actually present the work you've done in a professional manner. 

        We would request you to record a quick 30-60 sec video trailer of your talk and share the video with us ASAP. Here is an example.

    • Naresh Jain
      By Naresh Jain  ~  11 months ago
      reply Reply

      Hi Avik, can you please respond to Vishal's comments?