Lecture 6: Artificial Intelligence and Machine Learning

COGSCI 1 Lecture 6

Feature Engineering

Data is labeled with features for the algorithm to work on (in Machine Learning)

Representation Learning

programs are written to do their own feature engineering with raw data; no human intervention

Selectivity/Invariance Problem

System needs to develop a highly selective representation; Needs to be invariant and recognize differences from different angles or forms

Deep Learning solves this problem by coding the input in terms of increasingly complex features with each layer

Supervised Learning

Receives explicit feedback on how successful it is

Unsupervised Learning

Does not receive explicit feedback, learns to detect patterns in data on its own

Reinforcement Learning

depends on a feedback signal, but is given a reward system instead of being fed data; has to figure out how to maximize the reward on its own

Given a reward system, has to figure out how to get reward

AlphaGo

Project where AI learned to play the board game Go

Used supervised learning at first: database of 30 million moves, received feedback on success

Shifted to Reinforcement Learning after achieving a high playing strength

AlphaGo Zero

Learned quickly to beat previous versions of AlphaGo (after 3 days, able to beat all versions within 40 days); More computationally efficient

Biological Plausibility of Reinforcement Learning

Reinforcement learning is the most common type of learning in the Animal Kingdom;

Parallels between algorithms in deep neural networks and reward processing in the human brain