CS231n⚓︎

CS231n-Lecture1 Introduction⚓︎

The tools: machine learning or statistical learning theory

This class wants to explore the intersection of Computer Vision and Deep Learning

Today's agenda:

• A brief history of CV/DL
• CS231n overview

Brief History⚓︎

How does visual intelligence work? That's the core of the course. If you know the principles of human vision, you can pass. Introduce the principles of vision (Nobel Prize), and the first Ph.D. paper in CV, btw, and more history.

How to recover information from 2D to 3D is the core question. You should use some validation tools (like triangle methods). And our language is 1D (a generated "thing"), but the real world is not 1D, so language and vision are very different (CV and NLP).

AI winter (the methods and works didn't deliver) came.

The breakthrough happened in the 1980s with FNN's Backprop! (Hinton and LeCun) And we get through DL in the 2020s. Then ImageNet appears.

• 2012 AlexNet Hinton(Original 30%) exciting year!
  • Actually, there isn't too much change compared to Neocognition, 32 years apart, but there are two big things.
  • Backpropagation happened
  • Data breakthrough (ImageNet)

And the following pictures show that hardware plays an essential role in the AI era. (Fuck you, NVIDIA)

Human bias (data is produced by humans, so it naturally has bias) also exists, and the issues are not all technical issues (very vital).

CS231n overview⚓︎

• Deep Learning Basics
  • Image Classification
• Perceiving and Understanding the Visual World
  • Tasks and Models
  • Video Classification/Multimodal Video Understanding
  • Large Scale Distributed Training
• Generative and Interactive Visual Intelligence
  • Self-supervised Learning
  • Generative Modeling
  • Vision Language Models
  • 3D vision
  • Embodied Intelligence
• Human-Centered Applications and Implications
  • You can ask Hinton(AGI)