HOW INTELLIGENCE WORKS · REPRESENTATIONAL UNIT SELECTION · eduKateSG
How a Mind Chooses Whether a Pattern Should Be Stored as One Whole or as Composable Parts
Representational unit selection is the intelligence process that decides what should count as one cognitive unit. Some structures are best learned and retrieved as an integrated whole. Others become more flexible when represented as smaller parts that can be recombined into new configurations.
Experience → candidate segmentation → compare whole-pattern regularity with part-level recombination → choose unit size → store relations → test retrieval and recomposition → revise segmentation when transfer fails.
This article belongs to the How Intelligence Works series. Cognitive Compression owns reducing many elements into a usable structure. Relational Reasoning owns reasoning through relationships. Resolution Selection owns how much detail to use for a current task. Representational unit selection owns the prior encoding choice: what counts as one reusable unit in the first place?
The Mind Must Decide What Moves Together
A word can be learned as one lexical item or decomposed into morphemes. A chess pattern can be stored as one familiar configuration or as relations among pieces. A mathematical procedure can be one routine or a composition of smaller operations.
Different unit choices support different kinds of speed, transfer and recombination.
A good unit is large enough to be efficient and small enough to remain reusable.
1. Unitary Representations Preserve Integrated Patterns
When elements reliably occur together and function as one meaningful structure, representing them as a whole can reduce cognitive load and accelerate recognition.
The trade-off is that internal parts may become less independently accessible.
2. Compositional Representations Preserve Recombination
When parts recur in different combinations, representing them separately can support flexible transfer.
The trade-off is greater coordination cost during use.
3. Chunking Is One Form of Unit Selection
Practice often turns several operations into one chunk. The new unit is not merely shorter; it changes what working memory treats as one element.
Chunking is valuable when the internal structure is stable enough to travel together.
4. Over-Chunking Can Hide Important Internal Structure
A learner may memorise a whole formula pattern without understanding which component changes under a new condition.
When a unit becomes too indivisible, transfer fails at the first place where one internal part must move independently.
5. Under-Chunking Can Create Unnecessary Load
Treating every letter, step or micro-action as independent keeps too many pieces active at once.
Expertise often appears fluent because repeated structures have become larger reliable units.
6. Representational Units in Mathematics
An algebraic expression can be seen as individual symbols, terms, factors, functions or an entire transformation pattern. The useful unit depends partly on which structures recur and which must be manipulated separately.
Strong learners can often move between unitisations.
7. Representational Units in Language Learning
A phrase such as “on the other hand” is often more useful as a multiword unit than as four separate words. Yet compositional knowledge still matters when grammar requires variation.
Fluency often depends on storing some constructions whole while retaining access to parts.
8. Good Unit Selection Supports Both Recognition and Recomposition
The strongest representation often keeps a fast whole-pattern route and a slower part-level route available.
This permits fluent routine use without losing the ability to inspect or modify internals when novelty appears.
9. Representational-Unit Failure Atlas
| Failure | What happens | Repair |
|---|---|---|
| Over-fragmentation | Too many tiny elements overload active processing | Chunk stable relations |
| Over-chunking | Internal parts cannot be modified independently | Reopen the unit |
| Surface chunking | Pieces are grouped by appearance rather than function | Group by invariant relation |
| Rigid unit size | One segmentation is used for every task | Maintain multiple useful parses |
| Premature whole learning | A routine is memorised before components are understood | Teach internal structure first |
| Permanent decomposition | Learner never compresses mastered steps | Proceduralise stable sequences |
| Transfer fracture | A stored whole cannot be recombined in a new context | Build compositional representations |
10. Unit Selection and Resolution Selection Are Different
Resolution selection chooses how much detail to consult for the current problem. Representational unit selection decides how experience is segmented into reusable units in memory and reasoning.
One chooses the zoom. The other chooses what counts as an object at that zoom.
11. Teams Should Agree on the Unit of Work
Projects often fail because one person treats a task as one indivisible package while another assumes its parts can be assigned separately.
Coordination improves when the unit of work, evidence and ownership is explicit.
12. Institutions Encode Unit Choices Into Forms and Systems
A database schema, curriculum standard or reporting template decides what travels as one record and what remains separable.
Poor unitisation can make later integration, audit or migration unnecessarily difficult.
13. Artificial Intelligence and Compositional Representation
AI systems can generalise more flexibly when internal or external representations support reusable parts and relations rather than memorised wholes alone.
A stronger agent also knows when a familiar multi-step routine should be invoked as one unit and when novelty requires reopening the steps.
14. The Representational Unit Audit
- Pattern: What elements regularly travel together?
- Function: Do they perform one stable job?
- Variation: Which parts must change independently across cases?
- Reuse: Which subparts recur in new combinations?
- Load: Is the current segmentation overwhelming working memory?
- Opacity: Has chunking hidden an internal relation we need?
- Transfer: Can the representation survive a changed surface?
- Dual route: Can we access both whole and parts?
- Ownership: What should count as one unit for handoff or accountability?
- Revision: What evidence would make us re-segment the representation?
15. Systems Reading: Is This One Stem or Four Channels?
In a systems frame, a group of sounds may be bounced into one stem for fast control, or kept as separate channels for later remixing.
Bounce too early and flexibility disappears. Keep everything separate forever and the console becomes unmanageable.
16. Return to the Unit That Can Travel
Representational unit selection determines what intelligence can recognise quickly and what it can recombine flexibly.
The right unit is neither always the smallest nor always the largest.
The mature mind stores stable wholes when speed matters and preserves parts when the future may require recomposition.