Skip to content

Machine Learning Hot Sheet

Use this page when the repo's exact ML route is:

  • binary labels
  • one linear separator
  • perceptron updates only on mistakes
  • linearly separable training data

Do Not Use When

  • probabilities or calibrated scores matter
  • the data is not separable and no surrogate-loss route is being taught
  • the real topic is Gradient Descent, logistic regression, or neural networks
  • the task is normal CP modeling with full offline input

Exact Route In This Repo

  • classifier: sign(w · x + b)
  • labels: -1 / +1
  • update on mistake:
  • w <- w + yx
  • b <- b + y
  • stop after one full pass with no mistakes

Recognition Cues

  • linearly separable data
  • deterministic online update rule
  • one hard classifier, not one probability model
  • benchmark wants the perceptron specifically

Core Invariants

  • include a bias term
  • update only when y (w · x + b) <= 0
  • one clean pass is the convergence signal only under separability
  • final weights are not unique; correctness is classification, not one exact parameter vector

Main Traps

  • forgetting the bias update
  • silently assuming nonseparable data will still converge
  • interpreting perceptron as a generic optimizer for all classifiers
  • overclaiming this lane into full ML coverage

Exact Starter In This Repo

Reopen Paths