An objective function is a mathematical statement of a model’s goal. It can require minimizing task errors or reducing the difference between predictions and actual values. section 1.1
Training algorithms adjust neural-network weights and bias terms repeatedly to improve performance against that objective. The precise mathematics is left aside, keeping the idea at an intuitive level. section 1.1
The objective changes with the learning setting:
- Supervised learning aims to predict outputs for unseen inputs. section 1.1
- Unsupervised learning aims to discover patterns such as clusters. section 1.1
- Reinforcement learning seeks an optimal strategy through rewards and penalties rather than a static dataset. section 1.1
- In mixture-of-experts models, selecting the loss function carefully helps balance training across experts. section 2.6
Table 1.1 Types of machine learning
| Supervised learning |
Unsupervised learning |
Reinforcement learning |
|
| Description |
The model learns by mapping labeled inputs to known outputs. |
The model is trained without labels and without a specific reward. |
The model learns from its environment based on rewards and penalties. |
| Data |
Labeled data |
Unlabeled data |
No static dataset |
| Objective |
To predict the output of unseen inputs |
To discover underlying patterns in the data, such as clusters |
To determine the optimal strategy via trial and error |
Compare learning types by their data sources and objectives. section 1.1
