Learning Techniques

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News


Learning technique is the umbrella term for the different ways a machine learning system can extract knowledge from experience. Every technique answers the same basic question, "what signal tells the model it is getting better?", but answers it differently. The signal may be a correct label written by a person, structure hidden in unlabeled data, a reward from an environment, a demonstration from an expert, a few examples in a prompt, or feedback from many devices that never share their raw data.

This page follows and extends the taxonomy popularized by Jason Brownlee's guide to 14 types of learning (linked above), which groups them into four families:

  • Learning problems: the three classic paradigms, supervised, unsupervised, and reinforcement learning.
  • Hybrid problems: approaches that blend those signals, such as semi-supervised, self-supervised, and multi-instance learning.
  • Statistical inference: how a model reasons from data to conclusions, whether inductively (general rules from examples), deductively (conclusions from general rules), or transductively (predictions about specific known cases without building a general rule).
  • Learning techniques: practical strategies layered on top, such as transfer, multi-task, active, online, and ensemble learning.

The field's foundations come from many people: Arthur Samuel coined "machine learning" in 1959; Leslie Valiant's 1984 PAC (probably approximately correct) framework formalized when learning from examples is possible; Vladimir Vapnik developed statistical learning theory and transductive inference; Richard Sutton and Andrew Barto built reinforcement learning; Rich Caruana formalized multi-task learning (1997); Chelsea Finn's MAML (2017) made meta-learning practical for deep networks; Brendan McMahan and colleagues at Google introduced federated learning (2017); and Yann LeCun has long championed self-supervised learning as the main route to general intelligence.

The paradigm shift of the past decade is that the boundaries between these techniques have blurred inside foundation models. A modern large language model is pretrained with self-supervision, adapted with supervised fine-tuning, aligned with reinforcement learning from human feedback, and then performs few-shot and zero-shot learning in context without changing its weights at all. Where older systems used one technique per model, today's systems chain several, and the main design question is which combination fits your data, labels, and deployment constraints.


There is no single way for a machine to learn. Learning techniques differ chiefly in the training signal they rely on, labels, structure, rewards, demonstrations, or context, and modern AI systems succeed by combining several of them, each covering the others' gaps.


The Mechanism

Choosing and applying a learning technique follows a fairly consistent sequence.

  1. Identify the available signal. Do you have labeled examples, only raw data, a way to score outcomes, expert demonstrations, or a pretrained model that already knows something relevant?
  2. Match the signal to a paradigm. Labels suggest supervised learning; unlabeled data suggests unsupervised or self-supervised learning; scorable outcomes suggest reinforcement learning; demonstrations suggest imitation learning; a related pretrained model suggests transfer learning.
  3. Account for constraints. Few labels point to semi-supervised, few-shot, or active learning. Data that cannot leave its owners points to federated learning. A stream of changing data points to online or lifelong learning. Many related tasks point to multi-task or meta-learning.
  4. Combine techniques. Real systems stack methods: self-supervised pretraining, then transfer, then fine-tuning with human feedback, then ensembles or distillation for deployment.
  5. Validate under realistic conditions. Hold out test data (Train, Validate, and Test) and, critically, test on data that differs from training (Out-of-Distribution (OOD) Generalization).
Paradigm Training signal Example Main limitation
Supervised Labeled input-output pairs Spam filtering, medical image diagnosis Labels are expensive and may be biased
Unsupervised Structure in unlabeled data Customer segmentation, anomaly detection Hard to evaluate; structure may not match goals
Self-supervised Labels generated from the data itself (masked words, hidden image patches, next token) Pretraining language and vision models Pretext task must be chosen well; needs heavy compute
Semi-supervised A few labels plus many unlabeled examples Speech recognition with limited transcripts Errors in pseudo-labels can reinforce themselves
Reinforcement Rewards from interacting with an environment Game playing, robotics, reasoning models Sample-hungry; rewards can be gamed
Imitation Expert demonstrations Robot manipulation, driving Mistakes compound when the agent drifts from demonstrated states
In-context Examples in the prompt, no weight updates Few-shot classification by an LLM Limited by context length and pretraining

Dependencies and conditions. Which technique works depends on the amount and quality of data, how closely training data resembles deployment data, the compute available, and whether mistakes are costly. Many techniques assume that training and test data come from the same distribution; when that assumption breaks, performance can collapse silently.

Where uncertainty remains. Researchers still debate why in-context learning works, how to make models learn continually without forgetting, and how far self-supervised learning on video can substitute for interaction with the physical world.

The Physics

The "physics" of learning techniques is the mathematics of what each one optimizes.

Supervised learning minimizes average error on labeled data plus a regularization penalty Ω that discourages overly complex models:

<math>\min_\theta \; \frac{1}{n}\sum_{i=1}^{n} L\big(f_\theta(x_i),\, y_i\big) + \lambda\, \Omega(\theta)</math>

How much data is enough? Valiant's PAC framework gives one classic answer. For a finite set of candidate hypotheses H, and data the true rule can fit perfectly, roughly

<math>n \ge \frac{1}{\varepsilon}\left( \ln |H| + \ln \frac{1}{\delta} \right)</math>

examples suffice to reach error at most ε with probability at least 1 − δ. Richer model classes need more data, which is why techniques that reuse knowledge (transfer, self-supervision, meta-learning) matter so much.

Unsupervised learning fits structure to unlabeled data, for example maximizing the likelihood of the data under a probabilistic model, or, in k-means clustering, minimizing distances to cluster centers:

<math>\min_{\{\mu_c\}} \sum_{i} \min_{c} \lVert x_i - \mu_c \rVert^2</math>

Self-supervised learning creates its own labels. Language models predict hidden or next tokens. Contrastive vision methods pull together representations of two augmented views of the same image and push apart other images, using the InfoNCE loss:

<math>\mathcal{L} = -\log \frac{\exp\big(\mathrm{sim}(z_i, z_j)/\tau\big)}{\sum_{k \ne i} \exp\big(\mathrm{sim}(z_i, z_k)/\tau\big)}</math>

where z are learned embeddings and τ is a temperature. Joint-embedding predictive architectures (JEPA) instead predict the representation of a hidden region rather than its raw pixels, so the model can ignore unpredictable detail.

Semi-supervised learning combines a supervised loss with an unsupervised one, L = Lsup + λ Lunsup, where the unsupervised term typically rewards consistent predictions under perturbation or trains on confident pseudo-labels.

Reinforcement learning maximizes expected discounted reward, E[Σ γt rt] (see Reinforcement Learning).

Transfer learning and parameter-efficient fine-tuning. Instead of starting from random weights, start from weights learned on a related task. Low-rank adaptation (LoRA) freezes the original weight matrix W0 and learns only a small low-rank update:

<math>W = W_0 + B A, \qquad B \in \mathbb{R}^{d \times r},\; A \in \mathbb{R}^{r \times k},\; r \ll \min(d, k)</math>

which can cut the number of trainable parameters by orders of magnitude.

Meta-learning learns how to learn. MAML finds initial weights from which one or a few gradient steps on a new task give good performance:

<math>\min_\theta \sum_{\text{tasks } \mathcal{T}} \mathcal{L}_{\mathcal{T}}\big(\theta - \alpha \nabla_\theta \mathcal{L}_{\mathcal{T}}(\theta)\big)</math>

Ensembles reduce variance. Averaging M models, each with variance σ2 and pairwise error correlation ρ, gives

<math>\mathrm{Var}\big(\bar{f}\big) = \rho\, \sigma^2 + \frac{1-\rho}{M}\, \sigma^2</math>

so ensembles help most when members make different mistakes (low ρ).

Continual learning and forgetting. Training on a new task can overwrite weights important for old ones ("catastrophic forgetting"). Elastic weight consolidation adds a penalty that anchors important weights near their old values, weighted by Fisher information F:

<math>\mathcal{L} = \mathcal{L}_{\text{new}}(\theta) + \sum_i \frac{\lambda}{2}\, F_i \left(\theta_i - \theta^{*}_{i}\right)^2</math>

A 2025 MIT study found that, for fine-tuning foundation models, the amount of forgetting is well predicted by a single quantity: the KL divergence between the fine-tuned and original model, measured on the new task.

Federated learning trains on data spread across many devices or institutions. In federated averaging, each client k trains locally on its nk examples and the server combines the resulting weights:

<math>w \leftarrow \sum_{k} \frac{n_k}{n}\, w_k</math>

Only model updates travel; raw data stays local, although updates can still leak information unless protected with techniques such as differential privacy.

Active learning chooses which examples a human should label next, often the ones the model is least sure about, for example those with the highest predictive entropy H(y | x) = −Σ p(y | x) log p(y | x).

Knowledge distillation trains a small "student" to match a large "teacher's" softened output probabilities (softmax at temperature T), transferring knowledge into a cheaper model.

Why This Matters

  • Data is the bottleneck. Labels are expensive, and high-quality public data is finite. Techniques that need fewer labels (self-supervised, semi-supervised, few-shot, active learning) determine what is practical.
  • Foundation models depend on combinations. Every major AI assistant is the product of self-supervised pretraining, transfer, supervised fine-tuning, and reinforcement learning; understanding the stack explains both capabilities and failure modes.
  • Privacy and regulation. Federated and distributed learning let hospitals, banks, and phones contribute to models without pooling sensitive data.
  • Robotics and the physical world. Imitation learning, simulation, and self-supervised video models are how robots acquire skills without millions of real-world trials ( Action Learning ... Embodied AI, Simulated Environment Learning).
  • Adaptability. Online, lifelong, and meta-learning aim at systems that keep improving after deployment rather than freezing at training time, a gap that remains one of the clearest differences between AI and human learners.
  • Reliability. Ensembles, multi-task learning, and careful validation improve robustness; poorly chosen techniques can learn spurious correlations (Causation vs. Correlation) that fail in new settings.
  • What remains uncertain. Whether continual learning can be solved for large models, how much in-context learning can substitute for training, and whether learning from video and interaction will close the gap between language skill and physical understanding.

Recent Developments

Self-supervised learning overtakes weak supervision in vision

In August 2025, Meta released DINOv3, a self-supervised vision model scaled to a 7-billion-parameter teacher trained on about 1.7 billion unlabeled images, curated from roughly 17 billion. Meta reported that, for the first time, a model trained purely with self-supervision outperformed weakly supervised models (trained on image-caption pairs, like CLIP) across a broad range of probing tasks, from fine-grained classification to semantic segmentation and video object tracking, using a single frozen backbone without fine-tuning. A technique called Gram anchoring kept dense, pixel-level features from degrading over long training. The 7B model was distilled into smaller variants for practical use. These are the developer's own benchmarks; independent evaluations in domains such as medical and satellite imaging are under way.

Learning world models from video

Meta's V-JEPA 2 (June 2025) pretrained a video encoder of up to 1 billion parameters on more than 1 million hours of internet video plus 1 million images, using the JEPA objective of predicting representations of masked regions. The team then trained an action-conditioned version on less than 62 hours of unlabeled robot video and deployed it zero-shot on robot arms in two labs, where it planned pick-and-place actions toward goal images without task-specific training, rewards, or data from those environments. The result supports the long-standing argument, championed by Yann LeCun, that self-supervised learning from observation can supply much of what a robot needs; success rates on more complex, longer tasks remain limited.

Test-time training and refinement loops

The ARC-AGI benchmark tests few-shot generalization on novel visual puzzles. In ARC Prize 2025 (March to November 2025), 1,455 teams submitted more than 15,000 entries; the top Kaggle score reached 24 percent on the harder ARC-AGI-2 private set at about $0.20 per task, and the grand prize remained unclaimed. The organizers identified "refinement loops," per-task optimization guided by feedback, as the defining theme. Winning approaches relied heavily on test-time training: briefly fine-tuning a model on each new puzzle's own examples, augmented with transformations, before answering. One paper-award winner, a roughly 7-million-parameter Tiny Recursive Model, reached about 45 percent on ARC-AGI-1, and another, CompressARC, trained only on each puzzle itself with no pretraining. These results blur the line between in-context learning (no weight updates) and fine-tuning (weight updates) and suggest that small amounts of learning at inference time can matter more than model size on novel tasks. ARC-AGI-3, adding interactive and agentic tasks, is in development.

Why reinforcement learning forgets less

Catastrophic forgetting remains a central obstacle to models that keep learning. In "RL's Razor" (ICLR 2026), Idan Shenfeld, Jyothish Pari, and Pulkit Agrawal at MIT compared supervised fine-tuning with on-policy reinforcement learning on large language models and robotic foundation models. At equal performance on a new task, RL preserved prior capabilities much better. The degree of forgetting tracked the KL divergence between the fine-tuned and original model on the new task, and on-policy RL is implicitly biased toward the solutions closest to the original model. A supervised "oracle" trained on KL-minimal targets forgot even less than RL, indicating that staying close to the original distribution, not the RL algorithm itself, is what prevents forgetting. This offers a practical design rule for lifelong learning: prefer updates that change the model's behavior as little as necessary.

Imitation learning at scale: vision-language-action models

Robotics has adopted the foundation-model recipe. Vision-language-action (VLA) models start from a pretrained vision-language model and learn to output robot actions from large collections of teleoperated demonstrations, a form of imitation learning. Physical Intelligence's π0.5 (April 2025; presented at the Conference on Robot Learning 2025) co-trained on heterogeneous data, including robot demonstrations from many embodiments, web data, and high-level task descriptions, and the authors reported that it could perform long-horizon tasks such as cleaning a kitchen or bedroom in homes it had never seen. Open implementations now let researchers fine-tune such models on low-cost robot arms. Independent benchmarks show pretrained VLAs generally beat imitation baselines trained from scratch but still fail frequently on precise or unfamiliar manipulation.

Distillation and the limits of self-improvement

Knowledge distillation has become central to deploying capable small models; DeepSeek, for example, distilled its R1 reasoning model into smaller open models in 2025. A NeurIPS 2025 study by Yang Yue and colleagues found that reinforcement learning with verifiable rewards mainly sharpened reasoning paths already present in a base model, whereas distillation from a stronger teacher genuinely expanded the student's capabilities, an important distinction when choosing between techniques.

Open questions

  • Can lifelong learning be made reliable for large models deployed in changing environments?
  • Is in-context learning a form of implicit gradient descent, Bayesian inference, or something else, and what are its limits?
  • How far can self-supervised learning from video replace physical interaction for robots?
  • Can federated learning scale to foundation models without unacceptable communication costs or privacy leakage?
  • When will test-time learning outperform simply using larger pretrained models, and at what computational cost?

Learning Technique Map

Core: the three classic types of learning problems in machine learning (semi-supervised and self-supervised learning, below, are sometimes counted as a fourth)

Hybrid: drawing from unsupervised and supervised learning

  • Semi-Supervised Learning ... combine a small labeled set with a large unlabeled one. Modern methods such as pseudo-labeling and consistency training (for example FixMatch, 2020) approach fully supervised accuracy on image benchmarks with only a few labels per class.
  • Self-Supervised Learning ... generate training labels from the data itself, such as predicting masked words or hidden image patches. It is how foundation models are pretrained, and in 2025 self-supervised vision models surpassed weakly supervised ones (see Recent Developments).
  • Multi-Instance Learning | Wikipedia ... labels apply to bags of examples rather than individual ones: a whole-slide pathology image is labeled "cancer" without marking which regions contain tumor. A 2019 Nature Medicine study trained on more than 44,000 whole-slide images using only diagnosis-level labels, showing the approach can work at clinical scale.

Statistical Inference: how a model moves from data to conclusions

  • Inductive Learning | Wikipedia ... infer general rules from specific examples, then apply them to new cases. Nearly all machine learning is inductive, which is why it can fail when the future differs from the past.
  • Deductive Inference | Wikipedia ... derive specific conclusions from general rules or a known model, as when a trained model makes predictions; the linked article covers deductive classifiers in knowledge representation, a related rule-based approach.
  • Transductive Learning | Wikipedia ... reason directly from the training examples to specific test examples without building a general rule, an idea developed by Vladimir Vapnik. In-context learning, test-time training, and tabular foundation models such as TabPFN behave transductively: they condition on the particular examples at hand.

Other Techniques:

  • Deep Learning ... neural networks with many layers that learn their own features; the engine behind most modern applications of the techniques on this page.
  • Transfer Learning a model trained on one task is re-purposed on a second related task. Typical practice is to start from a pretrained foundation model and fine-tune it, often with parameter-efficient methods such as LoRA that update only a small fraction of weights.
  • Few Shot Learning ... One-Shot Learning ... Zero-Shot Learning ... learn new classes or tasks from a handful of examples, one example, or only a description. GPT-3 (2020) showed that large language models can do this purely through prompts; CLIP (2021) enabled zero-shot image classification from text labels.
  • Ensemble Learning ... combine multiple models (bagging, boosting, stacking) so that their errors cancel. Random forests and gradient-boosted trees remain among the strongest methods on tabular data.
  • Multi-Task Learning (MTL) ... train one model on several related tasks so that shared representations improve each one; risks include tasks interfering with each other ("negative transfer").
  • Apprenticeship Learning - Inverse Reinforcement Learning (IRL) ... infer the reward function an expert seems to be optimizing, then learn a policy for that reward; useful when the goal is easier to demonstrate than to specify.
  • Imitation Learning (IL) ... learn a policy directly from expert demonstrations (behavior cloning). It now powers robot vision-language-action models trained on large teleoperation datasets.
  • Lifelong Learning - Catastrophic Forgetting Challenge ... keep learning new tasks over time without losing old ones; approaches include regularization (such as elastic weight consolidation), replaying stored or generated examples, and adding capacity. Recent work shows that updates which stay close to the original model forget less.
  • Neural Structured Learning (NSL) ... a TensorFlow framework (2019) that trains networks with structured signals such as graphs linking related examples, or adversarially perturbed inputs, as regularizers alongside ordinary features.
  • Meta-Learning ... "learning to learn": train across many tasks so the model adapts quickly to new ones, via learned initializations (MAML), learned similarity metrics, or black-box models. In-context learning in large language models is widely viewed as an emergent form of meta-learning.
  • Online Learning ... update the model continuously as data arrives one example or batch at a time, rather than training once on a fixed dataset; analyzed through "regret" relative to the best fixed model in hindsight. Common in recommendation, advertising, and fraud detection.
  • Human-in-the-Loop (HITL) Learning / Active Learning ... humans label, correct, or rank model outputs, and active learning selects the most informative examples to label next. RLHF is a large-scale form of human-in-the-loop learning.
  • Decentralized: Federated & Distributed Learning ... federated learning trains a shared model across devices or institutions that keep their data local (first deployed at scale for smartphone keyboard prediction); distributed learning splits training of one model across many machines.
  • Action Learning ... Embodied AI ... agents that learn by acting in physical or simulated bodies, combining perception, world models, and control.
  • Simulated Environment Learning ... train agents in simulation, where trial and error is cheap and safe, then transfer to the real world; closing the "sim-to-real gap" remains a central challenge.
  • Knowledge distillation ... compress a large teacher model into a smaller student by training the student to match the teacher's outputs; widely used to deploy capable models on limited hardware.

Choosing a Technique

Your situation Techniques to consider Watch out for
Plenty of labeled data Supervised learning, ensembles Label bias; overfitting to the test set
Few labels, lots of raw data Self-supervised pretraining, semi-supervised, active learning Pseudo-label errors; unrepresentative unlabeled data
A related pretrained model exists Transfer learning, parameter-efficient fine-tuning, few-shot prompting Domain mismatch; forgetting of general skills
Only bag-level or weak labels Multi-instance learning, weak supervision Models attending to the wrong evidence
Can score outcomes but not label actions Reinforcement learning Reward hacking; sample inefficiency
Expert demonstrations available Imitation learning, inverse RL Compounding errors outside demonstrated states
Data cannot leave its owners Federated learning with differential privacy Communication cost; data heterogeneity across clients
Data or tasks change over time Online learning, lifelong learning Catastrophic forgetting; drift detection
Many related tasks Multi-task learning, meta-learning Negative transfer between tasks
Deployment hardware is limited Knowledge distillation, quantization, small fine-tuned models Loss of rare capabilities in the student

Resources

References

Taxonomy and foundations

Recent research

Retrieved from "https://primo.ai/w/index.php?title=Learning_Techniques&oldid=46393"