Deep neural networks can be trained under several different learning paradigms, distinguished by what kind of feedback the network learns from. All five paradigms use the same underlying network machinery — what differs is the source of the training signal.
The Five Paradigms
| Type | Feedback Signal | DL Example |
|---|---|---|
| Supervised Learning | Labeled input-output pairs | A CNN trained on images with class labels (cat/dog/etc.) |
| Unsupervised Learning | No labels — find structure | An autoencoder compressing images into a low-dimensional representation |
| Semi-Supervised Learning | Small labeled set + large unlabeled set | Fine-tuning a network on 500 labeled medical scans after pretraining on 50,000 unlabeled ones |
| Self-Supervised Learning | Labels generated automatically from the data itself | Predicting a masked word from surrounding words (this is how every modern LLM is pretrained) |
| Reinforcement Learning | Reward signal from an environment, not fixed labels | An agent learning to play a game from win/loss reward, or RLHF fine-tuning a chatbot from human preference scores |
Self-Supervised Learning Deserves Special Attention
Self-supervised learning is arguably the paradigm behind the current AI boom, and it's the one most learners confuse with unsupervised learning. The difference: unsupervised learning looks for structure with no target at all (e.g. clustering); self-supervised learning manufactures a target from the raw data itself — masking a word and asking the model to predict it, or hiding part of an image and asking the model to reconstruct it. This "free" supervision is why models can pretrain on essentially the entire internet without any human labeling.
How Each Paradigm Shows Up Later in This Hub
- Supervised: nearly every CNN and classification example in the CNN Fundamentals, CNN Architectures and Computer Vision categories.
- Unsupervised: autoencoders and GANs in the Generative Deep Learning category.
- Self-supervised: its own dedicated category later, plus the pretraining stage of every LLM in LLM Fundamentals.
- Reinforcement learning: touched on in LLM Fundamentals (RLHF) and Advanced Deep Learning.
Common Mistakes
- Calling self-supervised learning "unsupervised" — it's technically supervised (there is a real prediction target), just with labels generated automatically instead of by a human.
- Assuming reinforcement learning needs a "correct answer" per step — it only needs a reward signal, which can be sparse and delayed (e.g. win/loss at the end of a game).
Interview Relevance
Q: "How does self-supervised learning differ from unsupervised learning?" Unsupervised learning has no prediction target at all — it looks for structure (clusters, compressed representations). Self-supervised learning creates its own prediction target from the raw data (e.g. predict a masked token), making it technically a form of supervised learning where the labels are free.
Practice Question
A company has 2 million unlabeled product photos and 3,000 manually labeled ones. Propose a training strategy using at least two of the five paradigms above, and explain why.
