Deep Learning Machinery in Computer Vision
The Deep Learning Machinery in Computer Vision has several objectives:
- demystify what is at stake in the learning process of deep learning models
- build an incremental intuition of the machinery behind the scene
- think out of the box of current frameworks such as PyTorch and TensorFlow
- grasp the main concepts to design and build a computer vision framework from scratch
How to read?
The Blog
The articles are presented from the newest to the latest. Reading the articles in the chronological order should be the way to go.
This order should propose an understanding path. It is close to the path I actually followed some years ago, as I built my own computer vision framework from scratch.
The Table of Contents
The articles are re organised into chapters. We will elaborate on 4 main chapters:
- In the
Deep Learning Meta Walkthroughwe explore the core of the learning process: forward pass, backward pass level 1, weights update… - In
From a Layer Perspective, we dig deeper into the layers: neural structure, backward pass level 2… - In
From a Network Perspective, we talk about deep learning models: supervised, unsupervised, maximal activation… - In
From a Developer Perspective, we elaborate on the steps needed to build a framework from scratch: impact graphs, gradient checking, GPUs…
Disclaimer
I will not try and give any code example in the current existing frameworks: there are already many resources on the Internet.
I will try to use the least mathematical material needed and justify their use by intuition. The mathematical content mainly deals with: functions, derivatives. Knowing about vectors and multivariate functions is a plus.
In order for mathematically shy people not to feel overwhelmed, I will try to use these signs:
means the content can be skipped after this sign
means reading should be safe after this sign