notes.txt

convoluted neural network
ensure learning rate is just right
- too low: take too much time
- too high: undefined behaviours
confusion matrix
overfitting
- overreliance on obvious features (especially similar ones)
- if the input data is too similar
- the model will only learn to match a certain orientation instead
  of the overall shape
data biasness
- unequal amount of data
- 2 classes:
  - class 1: 10000 samples
  - class 2: 5 samples
- confidence bias towards class 1

include proficiency in programming languages/frameworks

facial detection
- convert to black and white
- for every pixel
  - draw an arrow pointing in the darker direction in a gradient
    - consistent despite lighting
- center the face (if at an angle)
  - by using a neural network to find face landmarks
- make facial measurements using another neural network

activation functions
- modern non-linear activation functions typically use max(0, x)

gradient descent
- find the lowest point possible
- compute gradient

convolutional neural network
- split into rgb channels
- pixel values fed into convolutional layer
- convolute image (3x3 -> 1x1)
- pass images through filter
  - to increase contrast (max pooling)

max pooling
- out of the 2x2 pixels select most intense
- increase contrast

ml
- topology
  - "shape" of neural network
- innovation number
- crossover
- speciation
  - bots with similar neural network grouped together
  - remove topologies that don't work well
  - keep bots with some potential by preventing better bots from smothering them