Tasks
Task는 model이 무엇을 출력해야 하고 성공을 어떻게 측정할지 정의합니다. Architecture는 정보가 흐르는 방식을, modality는 input signal을, task는 target behavior를 설명합니다.
같은 input도 여러 task를 가질 수 있습니다.
중요한 질문은 가 무엇을 뜻하는가입니다. 는 class, scalar, ranked list, generated sequence, box, mask, answer, retrieved item, structured object일 수 있습니다.
이동 지도
| 필요 | 시작점 | 일반적인 metric 경계 |
|---|---|---|
| define the full task contract | Task specification | loss, metric, split, validity rule |
| choose the output type | Task output space | class, scalar, rank, sequence, graph, coordinate |
| connect input modality to task | Modality-task map | raw input, representation, output, metric |
| predict labels or values | Classification, Regression | calibration, threshold, error scale |
| rank or retrieve items | Ranking, Retrieval, Reranking | top-k, NDCG, enrichment, recall |
| generate objects | Sequence generation, Structured prediction | validity, diversity, likelihood, downstream score |
| localize in space | Localization, Object detection, Segmentation | IoU, mask quality, localization error |
| model molecules or proteins | Property prediction, Interaction prediction | assay context, split unit, leakage risk |
| model structure | Coordinate prediction, Graph prediction | geometry validity, invariance, equivariance |
Task Contract
Task는 입력과 출력 이름만이 아니라 decision boundary입니다.
where is input space, is output space, is allowed context, is training loss, is evaluation metric, and is the split rule.
| Contract part | Ask | Route |
|---|---|---|
| input space | what object does the model receive after preprocessing? | Modalities, Representation contract |
| output space | what counts as a valid prediction? | Task output space |
| allowed context | what information is available at inference time? | Evaluation protocol |
| loss | what signal is optimized during training? | Loss function, Learning methods |
| metric | what behavior decides success? | Metric selection |
| split rule | what unit must be held out? | Dataset split contract |
If two papers use the same model family but different , context, metric, or split, they are not solving the same task.
Output Space Map
| Output type | Typical task | Metric route | Failure to count |
|---|---|---|---|
| class | classification, detection decision | Classification metrics | threshold and prevalence mismatch |
| scalar | regression, property prediction | Regression metrics | unit, censoring, target scale |
| probability | risk, confidence, uncertainty | Probability metrics, Calibration | miscalibration |
| ranked list | retrieval, screening, reranking | Ranking metrics | candidate pool mismatch |
| sequence | text, protein, SMILES, trajectory | Generation evaluation | invalid syntax and filtered denominator |
| set or graph | structured prediction, graph prediction | Metric selection | permutation, matching, graph validity |
| coordinates | structure, pose, localization | Pose quality, Regression metrics | alignment, symmetry, atom mapping |
| action or trace | agents, tool workflows | Agent evaluation | side effect, recovery, completion evidence |
Task vs Objective vs Architecture
These are different axes.
| Axis | Example | Common Confusion |
|---|---|---|
| task | classify activity, rank candidates, generate molecule, predict coordinates | described as model family |
| objective | cross-entropy, contrastive loss, denoising score, flow matching | treated as the final task |
| architecture | CNN, Transformer, GNN, SSM, equivariant model | credited for gains caused by objective or data |
| modality | sequence, image, graph, 3D structure | used as if it already defines the output |
| evaluation | AUROC, RMSE, RMSD, NDCG, validity | metric chosen after seeing results |
For example, a Transformer can be used for classification, retrieval, sequence generation, or reranking. A diffusion model can be trained for image generation, molecule generation, structure generation, or denoising representation learning. The task note should pin down the output behavior before architecture or objective claims.
Generalization Unit
Task specification should name the unit of generalization.
| Task family | Example unit | Split unit risk |
|---|---|---|
| property prediction | molecule, material, protein | scaffold, homolog, assay source |
| interaction prediction | pair or complex | target family, ligand scaffold, pair leakage |
| retrieval | query-candidate pair | shared document, shared query template, corpus time |
| generation | sampled object | duplicated training object, invalid filtered output |
| structure prediction | sequence, complex, residue set | template, homolog, structure source |
| agent task | task instance and tool trace | repeated workflow template or hidden tool state |
The split unit should be stronger than the claim. If the claim is transfer to new targets, a random row split is usually too weak.
Task 묶음
| 묶음 | 노트 |
|---|---|
| Core ML | Classification, Regression, Ranking |
| Search | Retrieval, Similarity search, Reranking |
| Vision and spatial | Object detection, Localization, Segmentation |
| Language | Captioning, Question answering, Sequence generation, Sequence-to-sequence |
| Structured outputs | Structured prediction, Coordinate prediction, Graph prediction |
| Time and monitoring | Time-series forecasting, Anomaly detection |
Paper Reading Pattern
When reading a method paper, rewrite the task before reading the model diagram.
| Field | Write |
|---|---|
| input | raw object and model-ready representation |
| output | exact prediction type and validity rule |
| context | what information is allowed at inference |
| loss | training signal and supervision source |
| metric | primary metric and diagnostics |
| split | held-out unit and leakage audit |
| baseline | simplest relevant task baseline |
This keeps model novelty separate from task definition.
확인할 것
- output space는 무엇인가?
- 어떤 task output space가 valid prediction을 제한하는가?
- input, output, validity, loss, metric, split을 포함한 task specification이 있는가?
- raw input, representation, output space, metric을 연결하는 modality-task map이 있는가?
- example unit은 무엇인가?
- 어떤 split unit이 의도한 generalization을 검증하는가?
- model은 predicting, ranking, retrieving, generating, localizing, segmenting 중 무엇을 하는가?
- retrieval은 최종 output인가, similarity-search stage인가, reranking pipeline인가?
- task는 entity-level, pairwise, context-conditioned, structured 중 무엇인가?
- output은 independent, sequential, structured, ranked, spatial, temporal 중 무엇인가?
- metric이 user-facing behavior와 맞는가?
- data split이 이 task의 leakage를 막는가?
- output이 valid object, syntax, molecule, structure, action으로 제한되는가?