learning-paradigms

Learning Paradigms - ML Strategy Selection Guide

Strategic guide for selecting machine learning paradigms when dealing with data constraints: self-supervised learning, few-shot learning, weak supervision, transfer learning, and meta-learning.

Overview

Modern machine learning extends beyond traditional supervised learning. This skill helps you choose the right learning paradigm based on your data characteristics, labeling resources, and deployment constraints.

Key covered paradigms:

• Self-Supervised Learning (SSL): Learn from abundant unlabeled data
• Few-Shot Learning (FSL): Classify with 1-10 labeled examples per class
• Weak Supervision: Leverage imperfect, incomplete, or noisy labels
• Transfer Learning: Adapt pretrained models to new domains
• Meta-Learning: Learn to learn across tasks

This skill synthesizes insights from recent research in audio classification, computer vision, and NLP to provide domain-agnostic guidance applicable to R, Python, or Julia implementations.

When This Skill Activates

Use this skill when:

• Starting a project with limited labeled data
• Facing annotation bottlenecks or high labeling costs
• Working with rare classes or long-tailed distributions
• Building systems for understudied domains (e.g., bioacoustics, medical imaging)
• Deciding between SSL pretraining vs direct supervised learning
• Combining multiple learning paradigms (e.g., SSL + FSL)
• Evaluating if weak supervision can replace full annotation

The Five Paradigms

1. Self-Supervised Learning (SSL)

What it is: Pretraining on unlabeled data by solving pretext tasks (contrastive learning, masked prediction, reconstruction) to learn useful representations.

When to use:

• ✅ Abundant unlabeled data available (1000s-millions of samples)
• ✅ Expensive or slow to label data
• ✅ Need domain-specific representations (pretrained models don't transfer well)
• ✅ Can define meaningful augmentations or pretext tasks

When NOT to use:

• ❌ Very small datasets (< 1000 samples) - not enough for SSL pretraining
• ❌ Task is fundamentally different from pretext task
• ❌ Strong pretrained models already exist and transfer well

Common techniques:

• Contrastive learning: SimCLR, MoCo, Barlow Twins
• Masked prediction: MAE (images), BERT (text)
• Consistency regularization: FixMatch, UDA

R ecosystem support:

• {torch} + {luz}: Custom SSL implementation
• {keras3}: Contrastive learning with Keras
• Limited native packages - often requires Python interop via {reticulate}

Expected gain: 5-20% accuracy improvement on downstream tasks with limited labels.

---

2. Few-Shot Learning (FSL)

What it is: Learning to classify new classes from 1-10 labeled examples per class (N-way K-shot), typically via metric learning or meta-learning.

When to use:

• ✅ New classes emerge frequently (e.g., new species, products)
• ✅ Cannot collect many labeled examples per class
• ✅ Need rapid adaptation to new categories
• ✅ Have related tasks for meta-training

When NOT to use:

• ❌ Can collect 50+ labeled examples per class - standard supervised learning works better
• ❌ Classes are very similar (inter-class similarity high) - hard to discriminate with few shots
• ❌ No related tasks for meta-training

Common techniques:

• Metric learning: Prototypical Networks, Matching Networks, Siamese Networks
• Meta-learning: MAML, Reptile
• Data augmentation: Mixup, CutMix in feature space

R ecosystem support:

• {torch}: Manual implementation of prototypical networks
• No native meta-learning frameworks - Python (learn2learn, torchmeta) dominates

Expected performance: 40-70% accuracy on 5-way 5-shot tasks (domain-dependent).

---

3. Weak Supervision

What it is: Training with imperfect labels: noisy labels, incomplete labels (only clip-level, not frame-level), or labels from multiple annotators with disagreements.

When to use:

• ✅ Labels exist but are imperfect (crowdsourced, automatically generated)
• ✅ Fine-grained labels too expensive (e.g., temporal boundaries in audio)
• ✅ Multiple overlapping events (e.g., bird chorus - know species present, not when)
• ✅ Can use rule-based heuristics or domain knowledge as weak labels

When NOT to use:

• ❌ Can afford clean labels - clean data always better
• ❌ Noise rate > 40% - model may memorize noise
• ❌ No validation set with clean labels - can't evaluate properly

Common techniques:

• Multiple Instance Learning (MIL): Treat recordings as bags
• Attention pooling: Learn which frames are relevant
• Noise-robust losses: Symmetric cross-entropy, bootstrapping
• Label smoothing: Reduce overconfidence

R ecosystem support:

• {milr}: Multiple instance learning (basic support)
• {torch}: Custom attention mechanisms
• Better supported in Python (snorkel, weakly)

Expected robustness: Tolerates 20-30% label noise with proper techniques.

---

4. Transfer Learning

What it is: Reusing knowledge from a pretrained model (trained on large source task) and adapting to target task via fine-tuning or feature extraction.

When to use:

• ✅ Limited target data (< 1000 labeled examples)
• ✅ Good pretrained model exists for related domain
• ✅ Target task shares structure with source task
• ✅ Computational budget allows fine-tuning

When NOT to use:

• ❌ Source and target domains are very different (e.g., ImageNet → medical histology)
• ❌ Target data is abundant (10k+ labeled examples) - train from scratch often better
• ❌ Pretrained model is too large for deployment

Common techniques:

• Feature extraction: Freeze pretrained layers, train only classifier
• Fine-tuning: Update all or top layers with small learning rate
• Domain adaptation: Align distributions between source and target

R ecosystem support:

• {keras3}: Excellent - direct access to pretrained models (ResNet, EfficientNet, BERT)
• {torch}: Good - torchvision models, easy fine-tuning
• {tidymodels}: Integrates with pretrained embeddings

Expected gain: 10-30% accuracy improvement over training from scratch.

---

5. Meta-Learning

What it is: Learning to learn: training on multiple related tasks to acquire a learning algorithm that adapts quickly to new tasks.

When to use:

• ✅ Have many related tasks (e.g., multiple datasets, multiple domains)
• ✅ Need fast adaptation to new tasks at test time
• ✅ Tasks share structure but differ in specifics
• ✅ Sufficient compute for meta-training

When NOT to use:

• ❌ Only one task available - no distribution of tasks to meta-learn from
• ❌ Tasks are unrelated (no shared structure)
• ❌ Limited compute - meta-learning is expensive

Common techniques:

• Optimization-based: MAML, Reptile (learn initialization)
• Metric-based: Prototypical Networks (learn embedding space)
• Model-based: Neural Turing Machines, Memory Networks

R ecosystem support:

• Very limited - requires manual implementation in {torch}
• Python (PyTorch) strongly recommended for meta-learning

Expected benefit: 2-3× sample efficiency on new tasks after meta-training.

---

Decision Framework

Primary Decision Tree

START: What data constraints do you have?

├─ Abundant unlabeled data (1000s+) but few labels?
│  ├─ Yes → Consider SSL pretraining
│  │         ├─ Then fine-tune with few labels (SSL → Supervised)
│  │         └─ Or combine with FSL (SSL → FSL)
│  └─ No → Continue
│
├─ Very few labeled examples per class (< 10)?
│  ├─ Yes → Consider Few-Shot Learning
│  │         ├─ Have related tasks? → Meta-learning FSL
│  │         └─ No related tasks? → Transfer learning + FSL
│  └─ No → Continue
│
├─ Labels exist but are noisy/incomplete?
│  ├─ Yes → Consider Weak Supervision
│  │         ├─ Clip-level only? → Multiple Instance Learning
│  │         ├─ Noisy labels? → Noise-robust training
│  │         └─ Multiple annotators? → Aggregation + uncertainty
│  └─ No → Continue
│
├─ Pretrained model available for similar domain?
│  ├─ Yes → Transfer Learning (fine-tune or extract features)
│  └─ No → Train from scratch or SSL pretraining
│
└─ Multiple related tasks to leverage?
   ├─ Yes → Meta-learning
   └─ No → Standard supervised learning

Combination Strategies

Paradigms often work better together:

Combination	Use Case	Example
SSL → Supervised	Unlabeled abundant, moderate labels	Pretrain on 100k unlabeled, fine-tune on 1k labeled
SSL → FSL	Unlabeled abundant, very few labels	Pretrain on 50k unlabeled, 5-shot classify new classes
Transfer → FSL	Pretrained model exists, few target labels	Fine-tune ImageNet model with 10 shots per class
Weak → SSL	Weak labels + unlabeled data	Use weak labels as pretext task, refine with SSL
Meta-learning → FSL	Many related FSL tasks	Meta-train on 20 datasets, fast adapt to new dataset

Implementation tip: Start simple (transfer learning), then add complexity (SSL, FSL) only if needed.

---

Paradigm Selection Cheat Sheet

Scenario	Recommended Paradigm	R Support
10k+ clean labels, standard task	Supervised learning	⭐⭐⭐⭐⭐ Excellent (tidymodels, mlr3)
1k labels, pretrained model exists	Transfer learning	⭐⭐⭐⭐ Good (keras3, torch)
100k unlabeled, 500 labeled	SSL + supervised	⭐⭐⭐ Moderate (torch custom)
5-10 examples per class, new classes	Few-shot learning	⭐⭐ Limited (manual torch)
Clip-level labels, need frame-level	Weak supervision (MIL)	⭐⭐ Limited (milr, torch)
Noisy crowdsourced labels	Weak supervision (robust)	⭐⭐ Limited (torch custom)
Many related tasks, need adaptation	Meta-learning	⭐ Very limited (Python better)

Legend:

• ⭐⭐⭐⭐⭐ Native R packages, production-ready
• ⭐⭐⭐⭐ Good support, may need some custom code
• ⭐⭐⭐ Moderate, requires torch/keras3 + custom layers
• ⭐⭐ Limited, manual implementation required
• ⭐ Use Python via {reticulate} or switch languages

---

Practical Examples

Example 1: Bioacoustics with Limited Labels

Scenario: Classify 50 frog species from 3-hour recordings. Have 10 labeled clips per species, 500 hours unlabeled.

Solution:

1. SSL pretraining on 500 hours unlabeled (contrastive learning on mel-spectrograms)
2. Weak supervision on 10 clips per species (clip-level labels, learn frame-level via attention)
3. Few-shot evaluation for rare species (5-shot prototypical networks)

R implementation path:

• {tuneR} + {torch} for audio preprocessing
• Custom SSL implementation with {luz}
• Attention pooling for weak supervision
• Prototypical networks for FSL

Reference: See SSL+FSL combination pattern in examples/ssl-fsl-combination-pattern.md

---

Example 2: Medical Image Classification

Scenario: Detect rare disease from X-rays. 50 positive cases, 1000 negative cases, 100k unlabeled X-rays.

Solution:

1. Transfer learning from ImageNet pretrained model (anatomy structure preserved)
2. SSL pretraining on 100k unlabeled X-rays (adapt to medical domain)
3. Class imbalance handling (focal loss, oversampling rare class)

R implementation path:

• {keras3}: Load pretrained DenseNet/EfficientNet
• Fine-tune with frozen early layers, train classifier
• Use {themis} for imbalance handling in {tidymodels}

---

Example 3: NLP with Crowdsourced Labels

Scenario: Sentiment classification with 5 annotators per text. 30% annotator disagreement.

Solution:

1. Weak supervision with label aggregation (majority vote or probabilistic)
2. Transfer learning from BERT pretrained model
3. Uncertainty estimation to detect unreliable annotations

R implementation path:

• {text} package for BERT embeddings
• Custom aggregation (weighted by annotator reliability)
• {tidymodels} for classifier training

---

Guidelines for R Practitioners

When R is Sufficient

• Transfer learning with pretrained models ({keras3}, {torch})
• Standard supervised learning with feature engineering
• Moderate-scale SSL (< 100k samples)
• Simple metric learning (prototypical networks)

When to Use Python Interop

• Advanced SSL techniques (MoCo, BYOL, VICReg)
• Meta-learning frameworks (MAML, Reptile)
• Complex weak supervision (Snorkel, data programming)
• Production-scale FSL (learn2learn, torchmeta)

Interop pattern:

library(reticulate)
use_condaenv("ml-paradigms")

# Python SSL/FSL training
py_run_file("train_ssl.py")

# Load embeddings back to R
embeddings <- py$load_embeddings()

# Continue in R with tidymodels
model <- logistic_reg() %>%
  fit(label ~ ., data = embeddings)

Recommended Workflow

1. Prototype in R (if possible) to validate approach
2. Switch to Python for paradigms with weak R support (meta-learning, advanced SSL)
3. Return to R for downstream tasks (analysis, reporting, integration)

---

Common Pitfalls

SSL

• ❌ Insufficient data: SSL needs 10k+ samples to learn useful representations
• ❌ Poor augmentations: Augmentations must preserve semantics (e.g., don't flip bird calls vertically)
• ❌ No downstream evaluation: Always validate SSL embeddings on downstream task

FSL

• ❌ Not enough meta-training tasks: Need 20+ related tasks for effective meta-learning
• ❌ Overfitting to support set: Use episodic training to prevent memorization
• ❌ Ignoring class imbalance: FSL assumes balanced classes - preprocess accordingly

Weak Supervision

• ❌ No clean validation set: Need clean labels to evaluate and tune noise handling
• ❌ Trusting weak labels too much: Always treat as noisy, never as ground truth
• ❌ Ignoring label dependencies: Multi-label weak supervision harder than single-label

Transfer Learning

• ❌ Unfreezing too early: Let classifier train first before fine-tuning backbone
• ❌ Learning rate too high: Use 10-100× smaller LR for fine-tuning than training from scratch
• ❌ Domain mismatch ignored: If source ≠ target, consider domain adaptation techniques

Meta-Learning

• ❌ Single task meta-learning: Meaningless - need multiple related tasks
• ❌ Insufficient compute: Meta-learning is 10-100× more expensive than standard training
• ❌ Overengineering: Often transfer learning + fine-tuning works just as well

---

Evaluation Metrics by Paradigm

SSL

• Downstream accuracy: Classification accuracy after fine-tuning
• Linear probe accuracy: Train only linear classifier on frozen embeddings
• kNN accuracy: k-nearest neighbors in learned embedding space
• Embedding quality: t-SNE/UMAP visualization of class separation

FSL

• N-way K-shot accuracy: Standard FSL benchmark (e.g., 5-way 5-shot)
• Cross-domain generalization: Test on unseen domains
• Sample efficiency curve: Accuracy vs number of shots (1, 5, 10, 50)

Weak Supervision

• Clean test accuracy: Performance on fully labeled test set
• Label noise robustness: Accuracy vs noise rate (10%, 20%, 30%)
• Calibration: Expected Calibration Error (ECE) to check confidence

Transfer Learning

• Fine-tuning gain: Improvement over random initialization
• Convergence speed: Epochs to reach target accuracy
• Forgetting: Performance on source task after fine-tuning

---

Supporting Resources

Detailed References

• Complete paradigm taxonomy: references/learning-paradigms-taxonomy.md
• SSL+FSL combination pattern: examples/ssl-fsl-combination-pattern.md
• Paradigm selection flowchart: examples/paradigm-selection-flowchart.md

Key Papers by Paradigm

• SSL: "SimCLR: A Simple Framework for Contrastive Learning" (Chen et al., 2020)
• FSL: "Prototypical Networks for Few-shot Learning" (Snell et al., 2017)
• Weak Supervision: "Weakly Supervised Learning" (Zhou, 2018)
• Transfer Learning: "A Survey on Transfer Learning" (Pan & Yang, 2010)
• Meta-Learning: "Model-Agnostic Meta-Learning" (Finn et al., 2017)

R Packages by Paradigm

• SSL/FSL: {torch}, {luz}, {keras3}
• Transfer: {keras3}, {torch}, {tidymodels}
• Weak supervision: {milr} (basic), custom {torch} implementations

---

Quick Reference Card

I have abundant unlabeled data + few labels → Self-Supervised Learning (SSL pretraining + fine-tuning)

I have 1-10 labeled examples per class → Few-Shot Learning (prototypical networks, meta-learning)

I have noisy or incomplete labels → Weak Supervision (MIL, attention pooling, noise-robust losses)

I have a pretrained model for a related task → Transfer Learning (fine-tune or extract features)

I have many related tasks to leverage → Meta-Learning (MAML, task-aware models)

I can combine multiple paradigms → Hybrid: SSL → FSL, Transfer → FSL, Weak → SSL

R support is limited for my paradigm → Use {reticulate} to call Python libraries, return to R for downstream work

---

When to Use This Skill

Invoke this skill when:

• Designing a new ML project with data constraints
• Stuck deciding between SSL, FSL, transfer learning
• Evaluating if weak supervision can reduce annotation costs
• Combining paradigms (e.g., "Should I do SSL before FSL?")
• Unsure if R ecosystem supports your chosen paradigm
• Need decision trees or cheat sheets for paradigm selection

This skill does NOT:

• Implement specific algorithms (that's for framework-specific skills)
• Provide detailed hyperparameter tuning (see r-tidymodels or r-performance)
• Cover standard supervised learning (see r-datascience, r-tidymodels)