Inductive Learning

• Bayesian is inductive learning
• Learning is identifying which hypothesis set is a concept
• Hypotheses don’t disappear, they just become less likely
• Learning develops through more experience
• One challenge of Bayesian learning is that any small subset is consistent with many hypotheses
• Different hypotheses have different likelihoods based on the examples we are exposed to
• But in the end we also prefer smaller hypotheses over larger ones: The size principle
• Simple Clustering methods can be used to get the data to automatically create the hypothesis space needed for Bayesian modelling
• Probabilities of different sets then match with human judgments surprisingly well
• Clustering based on biology worked worse!
• Clustering using linguistic co-occurrences with Latent Semantic Analysis also worked worse!
• Human subject judgements of similarity worked best
• Suggests some human reasoning relies on Probability
• Bayesian learning can also learn categories
• Models are capable of making generalizations about the specific objects as well as the appropriate generalizations about categorization (superordinate categories!) in general.
• Advanced learning means learn constraints on what is a possible hypothesis
• Hierarchical Bayesian Modelling (HBM) can explain how we acquire Overhypotheses
• using observations from the lowest level (data) and calculating statistical inferences