Date / Time: Wednesday 4 November / 18:00 - 21:00 EET

Speaker: Thanos Tagaris
NTUA PostDoc researcher | Code.Hub ML/DS instructor

The past decade, advances in Deep Learning instigated a massive boost in popularity for the field of AI. The driving force of these advances was Computer Vision, where, in a couple of years, they managed not only to catch up, but even to surpass human performance in multiple tasks. AI hype surged after these remarkable breakthroughs and fueled AI’s development in the state we know today.

In this workshop we will present the key intuitions behind state-of-the-art visual models, explain the differences between popular image-related tasks and how to approach them, provide tips on how to debug and boost performance and give insights on where the industry and research might be headed. The tutorial will be complimented with a hands-on tutorial for building an interpretable visual classifier in tensorflow and keras.


Target audience:
Anyone willing to Delve into Deep LEarning

Prerequisites on Audience:
a) SW/HW:
Own laptop with wifi

b) Know-how: 
Basic concepts of supervised learning

Presentation + jupyter notebook with hands-on material.


1.5h presentation, 1.5h hands-on

  • Intro to Computer Vision
  • Traditional Neural Networks approaches and why they fall short
  • TensorFlow basics: building a Fully Connected Neural Network for the MNIST dataset (hands on)
  • Building blocks of Convolutional Neural Networks (CNNs)
  • Intuition behind CNNs
  • Building a CNN in TensorFlow/keras for the MNIST dataset  (hands on)
  • Tips for training and visualizing the performance of a model  (hands on)
  • Interpretability in visual models
  • Interpreting our CNN’s predictions  (hands on)

If time allows, workshop might also dive into Modern CNN architectures and use for image classification, state-of-the-art training techniques