Other

What we didn’t (much) cover

• Hypothesis testing
  • p-values: the p-value is the probability of obtaining a test statistic at least as extreme as the one that was actually observed, assuming that the null hypothesis is true
    • Bonferroni correction: divide the p-value by the number of hypotheses tested
  • Confidence intervals, standard error estimates
• Randomized search for MLE/MAP
  • Gibbs Sampling /MCMC (alternative to EM)
• Metric learning
  • {$x^\top A x$}
• Domain adaptation
  • adapt model from one distribution {$p(x,y)$} to another
• Structured data
  • {$x$} can be a graph: use graph kernels, graph Laplacians
  • {$y$} can be a structure (e.g. a parse tree)
• Meta-learning (auto-ML)
  • Search over hyper-parameters, network architectures, …

We only touched on

• Multitask learning
  • simultaneously predict multiple {$y$}s from the same features. (e.g. CCA)
• Reinforcement learning
  • Choose sequence of actions (or policy) to maximize expected reward
  • Markov Decision Processes (MDP, POMDP)
• Time series in deep learning
  • GNNs, LSTM’s generalize HMMs

Back to Lectures