問24
頻出度★★☆
教師あり学習の事例に関する記述として, 最も適切なものはどれか。
令和4年度 ITパスポート試験 問24
テクノロジ系基礎理論 › 情報に関する理論
This question's explanation and your answers so far.
解説
問 24
Incorrectア
It is not given data labeled with the correct answer of “matches preference / does not match preference”; instead it automatically grasps preferences from the features of the submitted image. This is closer to unsupervised learning (feature extraction and clustering), which finds features and patterns from unlabeled data, so it cannot be called an example of supervised learning.
Incorrectイ
It makes judgments using a pre-prepared rule-based program, and there is no process of learning from data. Since it is not machine learning in the first place, it does not qualify as an example of supervised learning.
Incorrectウ
It improves by repeatedly training the dishing-up motion, which corresponds to reinforcement learning, where good actions (a policy) are learned through trial and error. This differs from supervised learning, which provides training data with correct-answer labels.
- エ
正解 Supervised learning is a method that learns from a large amount of training data in which input data is paired with its corresponding correct answer (label), so that it can predict the correct answer for unknown inputs. In choice エ, data pairing the input of a “recorded crying sound” with the correct-answer label of the “cause of crying” is collected and input for learning, and the cause is then inferred from a new crying sound. Using training data in which inputs are paired with correct answers is exactly a typical example of supervised learning, making it the most appropriate.
ここがポイント
- Supervised learning trains on pairs of “input data + correct answer (label)” and predicts the correct answer for unknown data. As in choice エ, whether the training data carries the correct answer (the cause) is the deciding clue.
- Clearly distinguish it from unsupervised learning (pattern extraction and clustering without labels), reinforcement learning (learning a policy through trial, error, and rewards), and rule-based approaches (which involve no learning).
用語