Back to How Do We Learn Dec 2025 Aliveline

Day 18 (12/22)

Energy was not there today, so I skipped. Did a bunch of end of year admin stuff instead. Meh I think I am going to take Christmas/New Years “off”. That’s where the energy is naturally going anyways.

12/21 – 1/1 will be completely off. I will do an annual review over 2-3 days in that time period, but that’s it in terms of anything work-shaped.

---

Day 18 (1/9)

OK it’s 2026 and we are back! It was nice to take a break, read tons of random things, do a lil annual review. I feel more centered. It extended a bit longer than originally planned because we had a trial-and-error week or two with sleep training the baby, but things are clearer now. Let’s hop to it.

Today I made a lot of edits to the Matlab code (obviously AI doing ~all of it). Decided to get some minimal repros done before trying to extend Asymmetric Diffusion Map (ADM) to all the worms in the QSimeon dataset. Specifically:

1. Repro the plots in the original Brennan paper using original Kato 2015 data, which is available online
2. Validate that the Kato 2015 worms in the QSimeon dataset have manifolds which match the manifolds that are repro’d in (1)

(1)

Able to successfully do (1) for the 5 worms used in Brennan, tho the pipeline is suuuper slow for these 100 neuron worms. Putting the results here for fun

Worm 1

Worm 2

Worm 3

Worm 4

Worm 5

(2)

Realized that (2) is not doable unfortunately because the 12 Kato worms in the QSimeon dataset are the ones from the Stim condition, whereas the 5 Kato worms used in Brennan are from the NoStim condition. Not sure if there’s overlap in the specific worms, but the Stim condition is not some intrinsic locomotion situation. Chat’s description:

---

Stim vs NoStim in Kato2015 really is an external O₂ drive:

• NoStim: 18 min at constant 21% O₂.
• Stim: 12 min total: first 6 min at 21% O₂, then 6 min of 30 s shifts between 4% and 21% O₂.

So if you run ADM on WT_Stim, you’re fitting dynamics of a system that’s being time-dependently forced (piecewise-stationary, with regular step changes).

What to expect if you apply ADM to WT_Stim

You don’t necessarily need to “remove global motion” as if it’s an artifact—because the O₂ steps are a real input that changes brain/behavior state. But you should expect nonstationarity to show up in the embedding/model:

• The same locomotion loop(s) may still appear (they’re strong), but the loop can thicken, split, or acquire an extra slow axis that correlates with O₂ level / time-since-switch.
• Transition probabilities can become a mixture of two regimes (4% vs 21%), which is slightly at odds with the “single stationary Markov process” assumption that ADM is implicitly leaning on.

A good way to think about it for your goal (extending ADM to qsimeon)

Since the qsimeon-preprocessed Kato source they’re using is explicitly WT_Stim.mat, the question becomes: do you want ADM to recover an “intrinsic locomotion manifold,” or a “locomotion manifold + how O₂ modulates it”?

---

Sooo yeah seems like each of these datasets may have some weird experimental conditions going on. Perhaps I should be much more selective with my dataset choices here, and think more clearly about what it means to “extend ADM to QSimeon”.