Introduction to ML
Field where systems improve performance through experience rather than explicit rules
Learning (or ML)
Learning from labelled input-output pairs
Supervised
Full form abbreviation used for PAC
Probably-Approximately-Correct
Set of all hypotheses consistent with training data seen so far
Version-space
Umbrella field of which Machine Learning is a subset
AI (Artificial Intelligence)
A candidate function proposed by a learning algorithm
Hypothesis
Learning from unlabelled data to find structure
Unsupervised
Symbol for the accuracy/error parameter in PAC learning
Epsilon
Earlier, simpler algorithm that only generalizes from positive examples
Find-S
ML subfield using multi-layer neural networks
Deep-learning
Set of all candidate hypotheses an algorithm can choose from
Hypothesis-space (or H)
Learning via rewards and penalties from an environment
Reinforcement
Symbol for the confidence parameter in PAC learning
Delta
Boundary of most specific consistent hypotheses
S (Specific-boundary)
Ability of a model to perform well on unseen data
Generalization
The true function the learner is trying to approximate
Target (concept)
Predicting a continuous numeric value
Regression
Minimum number of examples needed to guarantee a PAC bound
Sample-complexity
Boundary of most general consistent hypotheses
G (General-boundary)
ML enables chatbots and virtual assistants to understand human language by learning patterns from large amounts of text — this ML subfield is called?
NLP (Natural Language Processing)
Data used to check generalization on unseen examples
Test-data
Predicting a discrete category/label
Classification
A class of concepts that can be learned within PAC guarantees is called
PAC-learnable
Algorithm that updates S and G as new examples arrive
Candidate-Elimination
In healthcare diagnostics, ML-based systems have largely replaced traditional hand-coded rule systems, which are known as?
Expert-systems