{
  "id": "train-xor-mlp",
  "title": "Train a 2-4-1 network on XOR",
  "kind": "train",
  "summary": "Train a small sigmoid network with full-batch gradient descent until it labels noisy XOR points with 95 percent held-out accuracy.",
  "difficulty": "easy",
  "tags": [
    "mlp",
    "backpropagation",
    "gradient-descent",
    "xor"
  ],
  "lesson": "phases/03-deep-learning-core/03-backpropagation",
  "cheatsheet": "neural-net-basics",
  "minutes": 30,
  "entry": [
    "train",
    "predict"
  ],
  "collections": [
    "train-tiny-models",
    "deep-learning-core"
  ],
  "time_limit_s": 20,
  "budget": {
    "steps": 2000
  },
  "target": {
    "metric": "held-out accuracy",
    "value": 0.95
  },
  "mutants": {
    "lr_wrong_sign": "Adds the gradient to the output weights instead of subtracting it, so the loss goes up.",
    "zero_init": "Starts every weight at zero, so all hidden units stay the same and the network cannot learn XOR.",
    "missing_hidden_derivative": "Forgets the sigmoid derivative h * (1 - h) when it sends the gradient back to the hidden layer.",
    "over_budget": "Trains for 3000 steps, more than the budget of 2000.",
    "global_rng": "Draws the first weights from np.random and not from the rng argument, so two runs with one seed differ.",
    "centers_input_in_place": "Centers X with X -= mean in place, so the caller's array changes."
  }
}