1. Problem Modeling: Formulate a problem to model
2. Data curation: To inform model
3. Architecture Design: To represent the model
- Deep Feed Forward Networks (ANN) FFN
- Convolutional Networks (CNN)
- Sequence Modeling (RNNs)
- Transformers (Attention mechanism)
- Auto Encoder (AE)
4. Loss Function: Access performance of the model
5. Optimization: To train the model
- Regularization
- Optimization
6. Evaluation Metrics: For evaluation
Type of Learning | Description |
Supervised Learning Learning with a teacher.
The model is provided with labelled examples and learns to map input data to corresponding output labels.
It's like learning from a textbook with answers provided. |
Unsupervised Learning Learning without a teacher.
The model is given input data without explicit labels and must find patterns or structures in the data on its own.
It's like exploring a dataset without any prior knowledge or guidance. |
Semi-supervised Learning Supervised Learning + Unsupervised Learning.
Learning with a limited teacher.
The model is trained on a combination of labelled and unlabeled data, leveraging both the provided labels and the inherent structure of the data to improve learning.
It's like having some answers but required to figure out the rest on your own. |
Self-supervised Learning Form of Unsupervised Learning.
Learning from oneself.
The model generates its own labels from the input data and learns to predict some aspects of the data from other parts of the same data.
It's like creating puzzles for yourself and trying to solve them to understand the underlying patterns. |
Contrastive Learning Type of Self-Supervised Learning
Learning through contrast.
The model learns representations by contrasting positive examples with negative examples, aiming to bring similar instances closer together and push dissimilar instances apart in the learned representation space.
It's like distinguishing between different examples to learn meaningful representations without explicit labels.
Reinforcement Learning Learning through interaction.
The agent learns to make decisions by interacting with an environment and receiving feedback in the form of rewards or penalties.
It's like learning to play a game by trial and error. |
Active Learning Learning with guidance.
The algorithm selects which data points to label, aiming to maximize learning efficiency by choosing the most informative instances for labelling.
It's like having a tutor who selects the most relevant exercises for you to practice. |
Transfer Learning Learning with prior knowledge.
The model leverages knowledge gained from solving one task and applies it to a related task, typically by fine-tuning or using it as a feature extractor.
It's like applying what you've learned in one subject to solve problems in another subject.
Meta-Learning Learning to learn.
The algorithm learns a learning strategy or optimization procedure that enables it to quickly adapt to new tasks or environments.
It's like developing study habits that help you learn new subjects more efficiently.