🖥️ Lecture slides — Session 12 (Mon Oct 26)
Slides for a general overview that combines chapter 3.1 and 3.2 can be found here.
Learning frameworks differ in what supervises the training: labels, the data’s own structure, or feedback from an environment. A seismic analyst who has labeled ten thousand waveforms as earthquake or noise supervises a model directly. An archive of ocean profiles that nobody has labeled offers only its own structure. An autonomous glider learns from the consequences of its own sampling choices. This chapter lays out the main frameworks — unsupervised, semi-supervised, self-supervised, reinforcement, and active learning — and where each fits in geoscientific work. Supervised learning, the workhorse of this chapter, is treated in depth from 3.2 onward.
1. Unsupervised Learning¶
The Argo array has returned millions of temperature–salinity profiles from the world ocean, and no one has labeled a single one of them. How many distinct water masses are in there? Which profiles belong together, and which look like nothing else in the archive? Questions like these — structure first, labels never — are the territory of unsupervised learning.
Unsupervised learning trains models on data without labeled outputs. There is no target variable to guide the training, so the data’s own geometry does the supervising: which data samples sit close together in feature space, which directions carry the variance, which samples sit far from everything else.
Three families of techniques answer those three questions. Clustering groups data samples by feature similarity (k-means, hierarchical clustering, DBSCAN — the subject of lesson 3.3). Dimensionality reduction compresses many correlated features into a few informative directions (PCA, t-SNE; see Chapter 2.12). Anomaly detection flags the data samples that sit far from the rest (isolation forest, one-class SVM).
Geoscience examples include geochemical anomaly detection in mineral exploration, where isolation forests flag samples with unusual multi-element signatures; seismic facies analysis, where clustering of seismic attributes groups reflection data into interpretable subsurface units; and climate regime discovery, where k-means applied to atmospheric pressure fields recovers the recurrent weather regimes of the mid-latitudes.
Advantages and Challenges¶
| Advantages | Challenges |
|---|---|
| No need for labeled data, which can be costly or impractical to obtain. | Interpretation of results can be subjective. |
| Useful for exploratory data analysis. | Model evaluation is less straightforward due to the absence of ground truth. |
2. Semi-Supervised Learning¶
A landslide inventory for a mountain region confirms a few hundred failures, mapped in the field after storms, while the region holds tens of thousands of hillslopes that nobody has visited. Discarding the unvisited hillslopes wastes most of the data; calling them all “stable” plants false labels. Semi-supervised learning takes the middle path: it trains on the small labeled set and the large unlabeled one together.
The unlabeled samples cannot say which class they belong to, but they map out where the data are dense and where the natural gaps fall — and a good decision boundary should run through the gaps, not through the middle of a dense cloud. That is the information the unlabeled hillslopes contribute.
The simplest recipe is self-training: fit a model on the labeled samples, let it predict labels for the unlabeled ones, adopt its most confident predictions as if they were true, and refit. The risk is plain — a wrong but confident label gets adopted early and propagates through every later round. Variants guard against this by training several models that label data for one another (co-training), or by spreading labels along a graph that connects similar samples (graph-based methods).
Geoscience applications include land-cover mapping from satellite imagery where field-verified labels cover only a small fraction of the scene; mineral prospectivity mapping from abundant but unlabeled geophysical surveys anchored by a few drilled deposits; and landslide susceptibility mapping where inventories record only a limited number of confirmed events.
| Advantages | Challenges |
|---|---|
| Improves model accuracy with less labeled data. | Risk of propagating incorrect labels. |
| Cost-effective when labeling is expensive. | Model performance heavily depends on the quality of the initial labeled dataset. |
3. Self-Supervised Learning¶
Self-supervised learning generates its own supervision from the structure of the data: the model is trained to predict a hidden part of the input from the visible part. No human labels are involved, yet the training is fully supervised in form — the “label” is a masked value, the next sample, or the identity of an augmented copy.
This is the training mechanism behind foundation models. Large language models are trained by masked or next-token prediction on text; vision and waveform foundation models are trained by masked-patch reconstruction or by contrastive learning, where the model learns to place augmented views of the same input close together in an embedding space and different inputs far apart. Applied at scale to unlabeled archives — and geoscience has enormous ones: continuous seismic records, satellite image stacks, reanalysis fields — self-supervision produces general-purpose representations that transfer to downstream tasks with few labels.
Key Concepts¶
- Pretext Tasks: Tasks whose labels come from the data itself.
- Examples: Predicting masked segments of a seismogram, reconstructing occluded image patches, contrastive matching of augmented views, temporal order verification.
- Representation Learning: The learned features are the product; they are reused, with light fine-tuning, for classification, regression, or detection tasks downstream.
| Advantages | Challenges |
|---|---|
| Uses vast amounts of unlabeled data. | Designing effective pretext tasks can be non-trivial. |
| Produces reusable, transferable representations. | The learned representations may not transfer well to every downstream task. |
In Chapter 4 you will build this mechanism yourself: an autoencoder that reconstructs its input through a bottleneck, and a masked-pretraining exercise where the network learns to fill in hidden portions of a signal. Those two labs are the small-scale version of how foundation models are trained.
4. Reinforcement Learning¶
An autonomous glider surveying a fjord must decide, dive after dive, where to sample next. Each choice returns data of some value and costs battery, and no one can label the “correct” trajectory in advance — the quality of a choice only becomes clear later, once the survey is complete or the battery is dead.
Reinforcement learning (RL) is built for such sequential decisions. An agent (the glider) observes the state of its environment (position, battery, the field mapped so far), takes an action (dive here, transit there), and receives a reward (information gained, energy spent). Over many interactions it learns a policy: a rule mapping states to actions that maximizes the reward accumulated over the whole mission, not just the next step. A large algorithmic literature builds on these ingredients — Q-learning and its deep-network variants estimate the long-run value of each action, while policy-gradient and actor–critic methods adjust the policy directly — but the vocabulary above is enough to read an RL paper in geoscience.
RL remains rare in geoscience because few problems offer a safe environment for millions of trial-and-error interactions. Working examples exist in adaptive sampling, where autonomous underwater or aerial vehicles learn survey paths that maximize information gain, and in reservoir and groundwater management, where injection and pumping schedules are optimized against simulators. RL is also the mechanism behind the post-training of large language models from human feedback.
| Advantages | Challenges |
|---|---|
| Suitable for sequential decision-making problems. | Requires extensive interactions with the environment. |
| Capable of learning complex behaviors. | Computationally intensive and may suffer from convergence issues. |
5. Active Learning¶
A regional network records a hundred thousand small seismic events a year; an analyst can carefully review perhaps two thousand seismograms. Which two thousand deserve the hours? Picking at random spends most of the analyst’s time on easy, redundant examples the model already handles. Active learning turns the question around: the model itself selects the data samples whose labels would teach it most, and asks the expert for exactly those.
Three choices define an active learning loop. The query strategy decides which samples to send for labeling — most commonly the samples the current model is least certain about, so that each labeled seismogram lands where the model’s knowledge is thinnest. The oracle supplies the true labels: a seismic analyst reviewing waveforms, a geologist naming rock samples, a climate scientist annotating weather patterns. The stopping criterion ends the loop, usually when the labeling budget runs out or when new labels stop improving the model.
Query strategies beyond uncertainty exist — committee disagreement, expected model change, density weighting — and are cataloged in Settles’ survey of active learning; for most geoscience problems, uncertainty sampling is the place to start.
Advantages and Challenges¶
| Advantages | Challenges |
|---|---|
| Reduces the amount of labeled data needed. | Requires an effective query strategy. |
| Cost-effective when labeling is expensive. | The quality of the oracle can impact performance. |
| Can improve model performance significantly. | Determining the stopping criterion can be difficult. |
6. Comparative Overview¶
| Learning Framework | Typical usage | Data Status | Model Evaluation |
|---|---|---|---|
| Supervised | Automate workflows for data mining or develop surrogate models | Labeled | Labeled test data, held out from training: accuracy, precision/recall, F1 for classification; MSE, RMSE, R² for regression. |
| Unsupervised | Explore or understand the data’s structure. | Unlabeled | Silhouette score, inertia, or domain-specific interpretation. |
| Semi-Supervised | Automating workflows | Limited labeled data and mostly unlabeled data | Performance measured on labeled validation sets. |
| Self-Supervised | Feature representation learning; pretraining of foundation models | Unlabeled data with inherent structure | Indirectly evaluated by the performance on downstream tasks after fine-tuning. |
| Reinforcement | Sequential decisions with feedback from interactions. | An environment where the agent can interact and receive rewards | Cumulative rewards or achieving the desired level of performance. |
| Active | Continuous model training to improve performance | Mostly unlabeled, a few labeled data | Performance measured on labeled validation sets, focusing on the efficiency of the labeling process. |
The evaluation metrics named here are defined where they are first used: silhouette score and inertia in lesson 3.3, accuracy/precision/recall/F1 in 3.4, and MSE, RMSE, and R² in 3.7 — this table is a map, not a prerequisite.
Key Takeaways¶
- Framework selection: Match the learning framework to the nature of your data and the problem at hand.
- Data utilization: Get value from all available data, labeled or not, by choosing the framework accordingly.
- Evaluation discipline: Whatever the framework, evaluation happens on data the model has not seen — a rule this chapter enforces in every notebook.