Chapter 1: Giving Computers the Ability to Learn from Data
These are my notes on reading through the book.
3 Types of machine learning
Trains a model based on training data that has labels, which are answers. Then tests that model with a different set of data and labels. Once the model is smart enough to accurately classify it can be repurposed for classifying new data and predicting where future data would appear.
- Labeled data
- Direct feedback
- Predicts outcome/future
- Useful for:
- Classifying data into one class or another
- Predicting values
Lots of data but no labels or classification for this data, a model's goal is to find patterns and clusters of information. This allows us to group results into groups that may not be apparent from human diagnosis. This same pattern recognition and clustering can be used to reduce a data set and remove "noise". Since it can detect clusters it can remove extreme outliers from the data set giving us a smaller data set to work with.
This is helpful in preprocessing data for other learning systems to reduce the "signal-to-noise" ratio and get better training data.
- No labeled data
- No feedback
- Find hidden structures in data sets
- Useful for:
- Finding clusters and patterns in data sets
- Data compression to reduce noise in pre-processing data
Given a series of challenges the model trains to overcome the challenges through trial and error. Each state the model achieves receives a value from a "reward function". Using this it can deduce how successful it is being. Over time the goal is that the machine learning algorithm can determine the best path through a series of challenges and potentially handle untrained challenges as well. The most complex form of machine learning.
- Decision process
- Reward system
- Learn series of actions