Learning Methods
Learning method note는 특정 architecture와 독립적인 training objective와 representation-learning strategy를 설명합니다.
대부분의 learning method는 어떤 target signal 를 만들고 어떤 loss를 최적화하는지로 설명할 수 있습니다.
For supervised learning, is a human or assay label. For self-supervised learning, is derived from the input itself.
Route Map
| Question | Start | Evidence Boundary |
|---|---|---|
| what kind of objective is being optimized? | Objective taxonomy | objective family, target construction, metric alignment |
| is the target externally labeled? | Supervised learning, Semi-supervised learning | label semantics, noise, split unit |
| is the target derived from the input? | Self-supervised learning | augmentation, masking, collapse, transfer metric |
| is the model reused on a downstream task? | Pretraining, Transfer learning, Fine-tuning | downstream protocol and data split |
| is representation quality the claim? | Linear probing, Representation evaluation | frozen encoder, probe capacity, task diversity |
| is behavior learned from actions or preferences? | Reinforcement learning, Preference optimization | reward source, off-policy data, evaluation loop |
| is the data distribution changing? | Domain adaptation, Active learning | source/target split and sampling policy |
Learning Contract
Learning method note should state the training signal before discussing architecture.
where is the training unit sampled by the method. Depending on the method, may be a labeled pair, masked view, augmented pair, noisy sample, preference pair, or trajectory.
| Contract part | Ask | Route |
|---|---|---|
| training unit | what object is sampled by the objective? | Example unit, Task specification |
| target construction | where does the target or feedback come from? | Objective taxonomy |
| loss | what exact quantity is optimized? | Loss function |
| sampling policy | how are masks, views, negatives, prompts, or trajectories sampled? | Augmentation policy, Sampling strategy |
| model selection | which validation signal chooses checkpoint or threshold? | Evaluation protocol |
| downstream evidence | what metric justifies the learning claim? | Representation evaluation, Metric selection |
Signal Map
| Signal source | Learning family | Main risk |
|---|---|---|
| measured labels | supervised or semi-supervised learning | label semantics, noise, censoring |
| hidden input parts | masked modeling | target too local or leakage through context |
| augmented views | contrastive, JEPA, invariance learning | augmentation changes semantic identity |
| corrupted/noisy samples | denoising, diffusion, score, flow objectives | corruption process mismatched to downstream task |
| teacher outputs | distillation | teacher bias and hidden data access |
| demonstrations | imitation learning | behavior cloning distribution shift |
| preferences | reward modeling or preference optimization | preference proxy mismatch |
| environment reward | reinforcement learning | reward hacking and off-policy bias |
| newly queried labels | active learning | sampling policy changes data distribution |
Task vs Objective vs Evidence
These should be recorded separately.
| Axis | Question | Example |
|---|---|---|
| task | what behavior is needed? | rank ligands, classify images, predict structure |
| objective | what training signal is optimized? | cross-entropy, InfoNCE, masked likelihood, reward |
| architecture | what computation carries the signal? | Transformer, GNN, CNN, SSM |
| data | what examples and feedback are available? | labels, unlabeled corpus, demonstrations, preferences |
| evidence | how is success measured? | AUROC, retrieval recall, RMSD, task success |
An SSL pretraining loss can improve a downstream classifier, but the pretraining objective is not the downstream task. A preference objective can improve chat behavior, but it is not the same as agent task completion.
Evaluation Boundary
Learning-method claims often fail because the evaluation budget changes with the method.
| Claim | Must Hold Fixed |
|---|---|
| better pretraining | downstream split, probe budget, fine-tuning budget, data access |
| better representation | frozen encoder, evaluator capacity, task suite, retrieval corpus |
| better fine-tuning | pretrained checkpoint, target data, search budget, early stopping |
| better domain adaptation | source/target split, label availability, target-domain leakage |
| better preference learning | preference data source, reward model, evaluation prompt/task set |
| better RL method | environment, reward, simulator version, rollout budget |
If the comparison changes architecture, data, objective, and evaluation at once, the note should say the claim is system-level rather than learning-method-specific.
Supervision and Transfer
| Method | Use For |
|---|---|
| Supervised learning | direct labels from humans, assays, annotations, or measurements |
| Semi-supervised learning | small labeled set plus larger unlabeled set |
| Pretraining | learning reusable parameters before downstream adaptation |
| Transfer learning | moving representations or weights to a target task |
| Fine-tuning, Fine-tuning protocol | adapting pretrained models with explicit data and metric contracts |
| Knowledge distillation | training a smaller or deployable model from teacher outputs |
| Instruction tuning | aligning language-model behavior to task instructions |
| Curriculum learning | changing sample difficulty over training |
| Imitation learning | learning from demonstrated actions |
Self-Supervised Learning
| Method | Core Target |
|---|---|
| Self-supervised learning | target signal derived from the input |
| Masked modeling | reconstruct hidden tokens, patches, residues, atoms, or regions |
| Contrastive learning | pull positive views together and push negatives apart |
| JEPA | predict latent representation of a target view |
| Augmentation policy | define invariances through view construction |
| Representation collapse | failure mode where embeddings lose useful variation |
| Representation evaluation | measure transfer beyond the pretraining loss |
Reinforcement and Preference Learning
| Method | Core Object |
|---|---|
| Reinforcement learning | policy, state, action, reward, return |
| Policy gradient | gradient estimator for expected return |
| Reward modeling | learned proxy for human, assay, or environment preference |
| Preference optimization | pairwise or listwise preference signal |
Objective Lens
Learning method note는 training signal을 명시해야 합니다.
where defines how examples and targets are sampled. For papers, this means the method is incomplete until the note states:
- example unit and split unit;
- target construction rule;
- loss and metric;
- augmentation, masking, negative sampling, reward, or preference source;
- evaluation task used to justify representation or behavior claims.
Domain Boundaries
| Domain | Extra check |
|---|---|
| molecules | scaffold leakage, chemical state, negative construction, activity cliffs |
| proteins | homolog split, family split, MSA/template leakage, residue indexing |
| structure models | coordinate frame, template use, pocket definition, geometric validity |
| LLMs | pretraining/test contamination, instruction data, preference source, tool access |
| agents | trace success, tool side effects, recovery, completion audit |
Move object-specific details to Computational Biology when the issue is protein, molecule, ligand, pocket, assay, or structure identity rather than the learning signal itself.