- Deep Learning is made up of 2 major domains
- Computer Vision
- Natural Language Processing
- Why I choose Computer Vision to learn first
- Because it's easier to see & understand images compared to numeric vector representation of a word in NLP's black box Neural Networks.
- Can use Neural Network Visualizations to understand the behaviour of the Neural Network better.
- Why I chose learn ONLY Image Classification in Computer Vision
- Image Classification is easiest problem to solve & understand in Computer Vision.
- Being new to Deep Learning, I decided to focus on easiest problem & learn essential fundamentals first. Then learning more complex parts of Deep learning will become easier.
- Steep Learning curve leads to very difficult journey or complete abondonment.
- Lots Hands on experimentation with Pytorch & keras too (for easy to understand code)
- Building my own Code Cookbook for entire
Neural Network Ecosystem, from network to it's visualizations
Found a lot of different learning resources from Internet. All Learning resources are here. Few good resources, fewer great resources & Most are bad resources, avoid them.
Compile good courses from diverse POVs, not just a Single POV of Data Science.
- Deep Learning needs a multi disciplinary POV - (Data Scientist + Neuroscientist + Programmer + Business Analyst). So we need to learn all of these not just 1 single POV
- Every course is biased because of Instructor's POV. So need diverse types of courses
Increamental complexity of Image Classification Problem. From scale of
| Competition | Progress |
|---|---|
| 10 Digits Recognition(MNIST) | |
| 10 Small Objects Recognition(CIFAR10) | |
| Imagenet |
- Alexnet
- Resnet