Back to How Do We Learn Dec 2025 Aliveline

Day 23

I have noticed that through these daily working notes I am repeating ideas and not really getting into the level of depth I would like. Part of that is because there was a 2-3 week break in the middle of this whole thing, which probably caused me to lose a bunch of intuition that was brewing. I think I have a pretty interesting experiment I can run to connect LOOPER and the DD-DC view using QSimeon (and/or a simulation), so I will start on that soon. But for now I think I could get a lot of value out of converting my understanding/hypotheses into Evergreen Notes.

Or is Evergreen Notes the best modality to do this? Perhaps writing out a longer form hypothesis/narrative would also work?

OK what might the Evergreen Notes look like, just to test it out?

Each neuron is a DD-DC
- Can’t you test this by seeing what happens if you perturb a single neuron and see if it behaves like a feedback controller?
Moving faster allows enables/strengthens learning through proprioception
- Hypothesis: because each neuron is a DD-DC, speed increases quality & quantity of feedback
Locomotion is implemented as loops in a low-D phase space in C Elegans
Each brain has its own neural coordinate system to implement the same behavior
Neural population activity encode tasks like motor control and even higher order cognition as manifolds
- As things become more predictable, tangling decreases (manifold geometry becomes simpler)
- hypothesis: As things become more predictable, some analog to Kolmogorov complexity goes down
What might a useful KC analog definition look like?
- Minimum number of neurons that need to be activated to achieve a particular neural state space trajectory?
- what if we frame this in terms of feedback/control??
Theories of the brain have shifted from a single-neuron view to global population dynamics due to advancements in measurement technology
- What’s the natural extension here? What kind of measurement would enable this?
The space of neuromodulators is massive, even in a small nervous system like C Elegans
- Ex: Cytokines paper
Can the Tracy-Widom distribution (or other universality class distributions) describe the transition of neural population activity before/after a rule is learned?
Behavior is a top-down computational constraint on a neuronal network
How does DNA bridge the micro (local neuronal rules) to the macro (evolutionarily relevant behavior)
- hypothesis: perhaps in local neuronal rules are in the language of feedback control laws?
What is the appropriate measure of ‘complexity’ or ‘representational capacity’ for a brain?
How does the parallelism of the brain tie into the feedback control / nonlinear dynamical system view?
This comment: In CN, the trend is to take algorithms from computer science and statistics and map them onto biology. What’s far rarer is extracting new ML algorithms from the biology itself.
What if intelligence/the brain were simple and they seem complex because the environment is complex? simple in the sense of having a relatively small set of rules

That’s… quite a list. It’s probably worth developing them as Evergreen Notes though to sharpen my claims / see connections better / be clear about what I actually understand.

LOOPER/DD-DC experiment:

Yes — there’s a pretty clean “highest value” experiment you can run on qSimeon that ties the two stories together without needing to literally prove “neurons are optimal controllers.” It tests the part where DD-DC says “local (piecewise) linear closed-loop rules are enough,” and LOOPER says “global activity organizes into 1-D strands that branch/merge.”

Highest-value dataset experiment: scaffold-conditioned local linear dynamics + “control leverage” per neuron

Idea in one sentence

Use LOOPER/ADM to label where you are on the 1-D scaffold (which strand + phase), then ask: does a piecewise-linear model predict the next population state much better than a single global model, and which neurons have the most leverage on moving along the strand vs switching strands?

That’s the tightest bridge:

LOOPER: provides the scaffold (1-D trajectories, branches/merges).
DD-DC: predicts that behavior can be governed by local linear rules / switching between regimes, i.e., “a controller bank.”

What you do (concrete)

Build a scaffold label for every timepoint
- If you have LOOPER: use its strand ID + 1-D coordinate (position/phase).
- If not: use your ADM embedding and define:
  - a main loop phase (e.g., angle in the first 2 diffusion components), and
  - a discrete “regime” label (e.g., clusters / branches via transition matrix structure).
Fit a simple local linear predictor of population dynamics
For each regime (and optionally phase-bin within regime), fit:
[
x_{t+1} \approx A_{r} x_t + b_r
]
where (x_t) is the full neural population vector at time (t), and (r) is the regime/strand (optionally phase-bin).

Use ridge regression (you’ll need it because (n) can be large vs samples per bin).
Cross-validated test: piecewise vs global
Compare prediction error of:
- a single global (A)
- your piecewise (A_r) (and/or (A_{r,\phi}))
If LOOPER’s scaffold is “real computational structure,” the piecewise model should beat global especially near branch/merge regions.
Extract “control leverage” of each neuron
For each regime/bin, quantify neuron (i)’s leverage as something like:
- state influence: ( |A_r[:, i]| ) (how much neuron (i) predicts changes in everyone else next step)
- phase influence: regress phase velocity (\Delta \theta_t) on (x_t), and look at coefficients per neuron
- switching influence: logistic regression predicting branch choice (r_{t+1}) from (x_t); coefficients give “who drives switches”
Do the key comparison across worms
This is where it becomes a real test of “degeneracy vs constraint”:
- Are the top-leverage neurons consistent across animals (suggesting genetically constrained micro-implementation)?
- Or does each worm use a different set of neurons to implement the same scaffold motion (degeneracy with conserved macro dynamics)?

What would count as a strong tie between the papers

You’d be able to say something like:

“LOOPER’s strands correspond to regimes where a local linear predictor is stable and accurate.”
“Branch/merge regions correspond to sharp changes in the best-fit local linear map (switching dynamics).”
“A small subset of neurons has high leverage on phase progression or regime switching, consistent with a controller-like role — but that subset is either conserved or degenerate across worms.”

That’s basically “switching DD-DC controllers generate LOOPER scaffolds,” tested in the only way you reasonably can with imaging data.

Why this is high value

It’s fast to implement (it’s just regression + your existing embedding/transition structure).
It produces crisp plots: prediction error vs phase; leverage heatmaps; worm-to-worm consistency.
It directly addresses your “local rule → global scaffold” question.

Main caveat (worth keeping in mind)

With calcium imaging you don’t have true causal interventions, so “control leverage” is effective influence (predictive, not causal). Still, it’s the right first pass.

If you want an even cleaner story: do a small simulation in parallel

If you want something that really looks like DD-DC → LOOPER end-to-end:

Minimal simulation recipe

Latent motor phase (\theta_t) on a ring (oscillator).
High-dimensional neural state (x_t) that is a nonlinear embedding of (\theta_t) plus noise.
Add a switching mechanism (e.g., two gaits / turn vs forward) that changes the local linearization.
Implement neurons as adaptive AR/ARX filters (DD-DC-flavored) that stabilize the oscillator and enact switching when certain sensory inputs occur.

Then run LOOPER (or LOOPER-like extraction) on the simulated (x_t) and show it recovers:

1-D strands (the phase trajectory),
plus branches/merges (switching),
and your piecewise linear fits recover the ground-truth regime switches.

This gives you a “proof of concept” that the conceptual link is mechanically plausible, and your qSimeon analysis becomes “do worms behave like this class of systems?”

If you only run one thing

Run the scaffold-conditioned local linear predictor + phase/switch leverage on 2–3 conditions (e.g., spontaneous locomotion + stimulated) across multiple worms.

If you tell me which qSimeon split you trust most (Kato-only? WT_Stim?), I’ll outline the exact minimal implementation choices (binning scheme, regularization, metrics, and the 3 plots that make it persuasive).