Sancheeta Kaushal

Musings of my world

Strong convictions precede great actions.


Notes for Deep Learning fast.ai MOOC!

Lesson 0:

  • Incredibly flexible function (NN), all-purpose parameter fitting(Backprop), fast and scalable(GPU) made DL possible
  • ?? before any function describes the function in Ipynb
  • Shift + Enter in Ipynb
  • Try things with your own data

Lesson 1:

  • Infinitely flexible function, all purpose parameter fitting and tuning which is fast and scalable.
  • Neural network is the universal approximation function.
  • Gradient Descent & Backpropagation is the tuning part.
  • NVIDIA GPU supports CUDA which is efficient in running deep learning computations.
  • P2 and t2.micro
  • AWS Setup
  • AMI: Amazon Machine Images: Snapshot of instance at a given point in time
  • Literate programming
  • Cmd + Shift + P for keyboard shortcuts on jupyter notebook
  • tmux
  • Test, Train and Validation Datasets. Also, introduced to sample data.
  • Magic function
  • Pretrained Model - Model with Learned Parameters
  • VGG 16 model
  • Keras.json & theanorc file for configuration
  • Concept of Batch and Mini Batch
  • GPU has limited memory and the data transfer is a costly operation
  • Finetuning replaces the last layer of a pretrained network with the current classes