Understand AI concepts and learning approaches — decision exercise
Exercise: review a community archive proposal
Original fictional case. No paid services or cloud account required. Difficulty: beginner · Estimated duration: 15 minutes
An archive wants to categorize digitized letters, find groups among untagged descriptions, and draft summaries for a curator. It has some reviewed category labels. A proposal also says that rewriting the prompt learns soft vectors and that any language model can interpret a scanned image.
| Work | Supplied evidence |
|---|---|
| Category prediction | Letters with reviewed category labels |
| Explore descriptions | Untagged text; no target groups supplied |
| Draft summaries | Curator approval required |
| Interpret scans | Candidate input capability not documented |
Your tasks
- Explain the roles of language processing, machine learning and generation in this proposal.
- Explain foundation-model reuse, the unresolved image-input requirement and the basic diffusion concept as a separate generation approach.
- Correct the prompt-engineering claim; contrast learned soft-prompt tuning with model fine-tuning.
- Match the labeled and untagged tasks to learning approaches. Contrast them with an agent learning from action rewards.
Reference solution
Language processing concerns the letters’ language. Learning the supplied categories fits supervised learning; grouping the untagged descriptions fits unsupervised exploration. Drafting summaries is generation, with curator approval retained. A reusable foundation model can be considered, but its exact input/output capabilities must be checked; a language-model label does not prove scan/image support. Diffusion describes learning to reverse a noising process for generation and is not a synonym for every language application. Rewriting instructions is prompt engineering. Learned soft-prompt tuning trains task-specific vectors while the other model parameters stay frozen. Model fine-tuning instead performs further task-specific training to refine pretrained model parameters; rewriting prose alone does neither. Reward-driven policy learning describes a different reinforcement setup; the archive case has not supplied one.
Rubric (8 house points)
Two points each for: correct task/field distinctions; model/modality and diffusion boundaries; the three-way engineering, soft-prompt and model-tuning distinction; all three learning-approach distinctions. Any claim of verified image input without evidence loses that boundary point.