Machine Learning 2: Neural Networks

Summary

This lesson on neural networks covered:

  1. How a simple set of weights and connections can “learn”
  2. The basic components of a neural network: weights, biases, layers, activation / loss functions, and gradient descent
  3. How CNN networks can process large amounts of spatial / temporal data without a ballooning number of parameters
  4. How LSTMs allow neural networks to “remember” and pick up on distant trends in temporal data
  5. How these neural networks can actually be constructed in Keras / TensorFlow
  6. How activation functions like ReLU and Softmax can be applied to convert the raw output of neurons into positive numbers or probability distributions
  7. How cross-entropy loss can capture the degree of “difference” between two distributions (categorical probabilities, in our case)
  8. How different optimizers (like Adam) can improve on naïve gradient descent performed using a fixed learning rate

Getting Help

Even if you’ve missed the session, you’ll be able to ask questions about whatever, whenever on our Discord. Feel free to drop your question in the #bootcamp channel or just message an instructor directly.

Useful Links

Recording

A bit more abstract than most of our sessions have been, but hopefully you find truly understanding neural networks as rewarding as I do!

Feedback

Don’t forget to fill in this week’s survey!