Concept Formation

When the colour of a previously safe plant is changed to red, the tardigrade avoids it. It never saw that specific plant hurt anyone — it generalised from the feature.

---

The Question

How does an animal learn that a category — red things — is dangerous, without anything in its nervous system being a "category detector"?

The Answer, Short Version

Feature channels accumulate valence. When a red thing causes pain, the redness channel slides negative. When any future object activates the redness channel — whether it caused pain before or not — its approach/avoid score inherits that negative weight. The category is the feature; the learning is valence accumulation through the same geometric settling that handles everything else.

The Setup

The test environment contains plants of various colours. The tardigrade eats plants. Some hurt. The question: after experiencing pain from a red plant, does the tardigrade treat novel red objects as dangerous — or has it only learned that the specific plant is bad?

The protocol:

1. Training. Run the tardigrade in an environment where red plants cause pain. Let it encounter one.
2. Transfer test. Take a previously SAFE plant (green) and change only its colour to red. Everything else about it is identical. Show it to the tardigrade.
3. Read the behaviour. Does the system approach (specific memory only, no abstraction) or avoid (concept transferred)?

A concept is formed only if the behaviour on the newly-red plant matches the learned valence of redness, not the prior valence of the specific plant it used to be.

What Emerged

After training:

Plants eaten: 1
Times hurt: 1
Feature associations:
  red   : -0.3
  green :  0.0
  blue  :  0.0

One encounter, one pain event, and the red feature channel has slid to −0.3 while green and blue sit at zero.

Transfer test:

Original colour: green
Changed to:      red
Valence:        -0.30
Result:         AVOID

The system avoids. The valence it assigns to the modified plant (−0.30) matches the learned feature valence of red, not the prior positive valence of the specific green plant. The feature won.

What This Proves

Category-level learning falls out of feature-level valence accumulation. No category detector. No category label. No "if this thing is red" check. The abstraction is structural — the redness channel exists because the sensory apparatus picks up colour, and whatever slides its valence slides the valence of every future thing that activates it.

This is the same primitive doing the work that, in a conventional reinforcement system, would be split across (a) a feature extractor, (b) a class assignment, (c) a policy table. The VINE version is one channel accumulating one scalar.

It also shows the tardigrade's learning is not case-based. If the association had been at the level of the specific plant, a recoloured version would have kept its prior positive valence. It did not. The feature dominated.

What This Does Not Prove

The training set here is extremely thin — one plant encountered, one pain event, one feature channel moved. That is enough to seed the mechanism and exhibit transfer, but it is not a thorough exploration of how the system behaves under competing features, ambiguous stimuli, or mixed training.

A fuller version of this test would include:

• Multiple training encounters at varying intensities, to map the dose-response of feature learning.
• Conflicting features — a red plant that is safe, to see how the system handles counter-evidence.
• Composite stimuli where more than one feature might carry the association, to see which channel wins and why.

Those are all later tests. This one demonstrates the minimum case: feature valence accumulates, and it transfers.

---

Raychell Langan · NEXICOG Ltd · Hampshire, UK