A feature (特徴量) is one of an object's properties, chosen as a clue for learning or prediction and expressed as a number.
Goal
Example features
Predict a house's price
Floor area, minutes on foot from the station, building age
Flag spam email
How often certain words appear, number of links in the body
Read handwritten digits
Number of closed loops in the strokes, number of stroke ends
Data as collected also contains information that does not help with the decision, so the properties that serve the goal are picked out, turned into numbers, and used as the input for learning. Which features you choose has a large effect on how accurate the predictions are. People used to design features by hand, but in deep learning (ディープラーニング), a method that learns through many stacked layers of a structure modeled on the brain's neural circuits, the computer finds for itself what to focus on as it learns.
AnalogyThink of someone pricing a used bicycle who ignores whether they like the color and looks at the distance it has been ridden and how worn the tires are. Only the points that decide the price get picked out as numbers.
How to rememberThe Japanese name ends in 量 ("quantity"), meaning something expressed as a number: a characteristic turned into a value.
Exam points
ExamIn deep learning (ディープラーニング), when given a large amount of data, the computer extracts by itself, during learning, the features that serve as clues for recognition.
Machine learning (機械学習) is technology in which a computer learns patterns from large amounts of data and uses them to make predictions and decisions.
Collect data
Extract features
Learn patterns from the data
Predict and decide on new data
Apply what was learned to data not seen before and predict the result.
4 / 4
The method of calculation produced by learning, which works out a result from an input, is called a model (モデル). A model can also be applied to new data that was not used for learning, so it can estimate future values from past records or decide what an image shows. Sorted by how they learn, the three main methods are supervised learning, unsupervised learning, and reinforcement learning.
AnalogyIt is like a trainee forecaster who, after comparing many past weather maps, starts to get a sense of tomorrow's weather. The knack comes from a pile of real examples.
How to rememberMachine Learning is often shortened to its initials, ML.
Exam points
ExamMachine learning (機械学習) learns patterns from stored data, so it can be used for forecasting, estimating future values from the conditions at hand.
ExamDeep learning (ディープラーニング) is one machine learning (機械学習) method, and it learns with a neural network (ニューラルネットワーク), a model based on the brain's neural circuits, built from many stacked layers.
Mix-upA rule-based (ルールベース) system runs on conditions and decision rules that people prepared in advance. Compared with machine learning (機械学習), which works out how to decide from data, the difference is who creates the rules.
Supervised learning (教師あり学習) is a method that trains on pairs of input data and their correct answers (labels), so that it can give the answer for inputs it has not seen.
Method
Output
Example
Regression
A continuous number
Estimate a used car's trade-in price from its model year and mileage
Classification
Which category it belongs to
Decide which of 0 to 9 a handwritten digit is
Adding correct answers to training data is called annotation (アノテーション), and how the answers are attached shapes the quality of the model. The main methods are regression (回帰) and classification (分類), and the finished model is tested on data whose answers are known to measure how well it performs. One such measure is recall (再現率): of the items that really are positive (the side you want to find), the share the model also judged positive.
AnalogyIt is like practicing with a workbook that comes with answers and explanations. After checking your answers again and again, you can handle problems you have never seen.
How to rememberThe Japanese name literally means "learning with a teacher": a teacher is someone who knows the answers, and you learn with those answers beside you.
Exam points
ExamAnnotation (アノテーション) is the work of adding an answer label to each item of data used for supervised learning (教師あり学習).
ExamRegression (回帰) predicts a continuous number such as sales or temperature from the input, and because it learns from data with correct answers it belongs to supervised learning (教師あり学習).
ExamRecall (再現率) is calculated with the number of items that really are positive as the denominator, as the share of those that the model also judged positive.
Unsupervised learning (教師なし学習) is a method that trains on data alone, without correct answers, and finds the groupings and structure hidden in that data.
The typical method is clustering (クラスタリング), which automatically puts similar data into the same group. It suits situations where no answer is set in advance, such as dividing customers into several segments by their buying habits. What each resulting group means is something people interpret by looking at what is in it.
AnalogyPicture a market you are visiting for the first time, where you sort vegetables you cannot name into baskets by similar shape and color. Nobody tells you the names, yet similar ones still end up together.
How to rememberThe Japanese name literally means "learning without a teacher": no answer labels, like being handed only the question sheet.
Exam points
ExamThe typical method of unsupervised learning (教師なし学習) is clustering (クラスタリング), which groups similar data together even without correct answers.
Mix-upUnsupervised learning (教師なし学習) differs from supervised learning (教師あり学習) in that its training data carries no correct-answer labels.
Reinforcement learning (強化学習) is a method that uses the reward received as a result of an action (a score that grows as the result improves) as its clue, and learns by trial and error how to act so the reward gets larger.
Choose an action
The situation changes
Receive a reward
Revise how to choose
It adjusts how it chooses actions so the reward grows, then goes back to the start.
4 / 4
It repeats the cycle of acting and getting a result many times, looking for actions that make the total reward large over the long run, not just the immediate reward. It is used for choosing moves in board games and for controlling robots so they walk without falling.
AnalogyIt is like learning to ride a bicycle with no one teaching you the steps, adjusting how you use your body by the feel of how far you got without falling.
How to remember"Reinforcement" is a psychology term for how behavior that gets praised is repeated more often. Remember it as strengthening actions with rewards.
Exam points
Mix-upReinforcement learning (強化学習) is not given a correct answer for each action and improves its actions using only the reward that matches the result. Compared with supervised learning (教師あり学習), which learns from the correct answer for each input, what it learns from is different.
Supervised Learning vs. Reinforcement Learning
Aspect
Supervised Learning
Reinforcement Learning
Key differenceWhat it learns from
The correct answer attached to each input
A reward that matches the result of an action
How it learns
Compares with the answer and moves its output closer
Repeats actions and results, searching for actions that raise the reward
Key points
A feature is a clue for learning that turns an object's properties into numbers, and the choice of features affects how accurate predictions are.
Machine learning learns patterns from data to build a model, then applies it to predict and decide on data that was not used for learning.
The data for supervised learning consists of pairs of inputs and correct answers, and attaching those answers is called annotation.
Clustering groups similar data without using correct answers and is the typical method of unsupervised learning.
Reinforcement learning uses rewards as its clue and learns how to act through repeated trial and error.