Demetrios Agourakis

final-year med student & researcher. computational psychiatry x complex systems. measurement, uncertainty, and models of mind.

[Demetrios Agourakis @DAgourakis]

Every programming language lies to scientists.

Python. Julia. Rust. All of them treat uncertainty as somebody else's problem.

9 months ago I started building one that doesn't. Alone. While finishing medical school.

It now compiles itself. What 2.2M lines taught me — a thread: 🧵

[Demetrios Agourakis @DAgourakis]

@whoami3339 Thanks. Did the decomposition stay stable across ICA runs or seeds? That would make the low overlap even more convincing.

[Demetrios Agourakis @DAgourakis]

@WeigangCui94 That's a clean way to frame it: the null sets the hub baseline, and the signal is what rises above it. Looking forward to the paper. Would you share the preprint when it's out? I'd like to try it on a psychiatric lesion-symptom set.

[Demetrios Agourakis @DAgourakis]

@cdickey_ @PNASNews Thanks. If the paired montage happens, trial-to-trial latency jitter is what I'd watch: conduction should give tight jitter, local recruitment a broader spread. Happy to help on the analysis side if useful.

[Demetrios Agourakis @DAgourakis]

@WeigangCui94 Clear, thanks. If the hub suppression comes from the null itself, does it also cost sensitivity for symptoms whose true network runs through hubs? A spin-style null on the same lesions would show how much the anatomy constraint is buying.

[Demetrios Agourakis @DAgourakis]

@TheColeLab Subject-specific task FC from the other visit raised R² from 0.49 to 0.60. Propagated/observed error was 0.71 with independent errors, 1.04 with a session factor. The factor uses the same activations across FC sources. Would split-run activations be a cleaner test?

[Demetrios Agourakis @DAgourakis]

Useful, thanks. If individualised task-state FC carries most of the explained variance (R² 0.40 vs 0.13 for individual rest FC), my independent-error baseline may be inflated by using the wrong FC source. I'll rerun the 0.70 to 1.04 check with subject-specific task FC and report back.

[Demetrios Agourakis @DAgourakis]

Makes sense. A vacuous set from genuine ambiguity and one from data scarcity look identical in set size but differ in n_eff = (Σw)²/Σw². Splitting alarms on that would tell you whether to slow down (ambiguity) or gather calibration data near that state (scarcity). Do your OOD scenes cluster enough to tell them apart?

[Demetrios Agourakis @DAgourakis]

@helder_nakaya For a live dashboard the probabilities matter more than the argmax. Were they calibrated per diagnosis (ECE, reliability curves)? And as the note grows token by token, do they move smoothly or jump around? Is there any abstain threshold before a clinician sees them?

[Demetrios Agourakis @DAgourakis]

@FranciscoICMC How does the Fisher penalty interact with x_min selection? In the Clauset–Shalizi–Newman pipeline (doi:10.1137/070710111) x_min is picked by KS distance, so if the penalised estimator shifts α, does the chosen x_min move too, and is the GoF bootstrap rerun with the penalised fit?

[Demetrios Agourakis @DAgourakis]

Neat that relevance falls out of training instead of a post-hoc permutation test. Could the learned ARD scales be reused for conformal? E.g. localising CD-split or local CP only over the relevant inputs, so neighbourhoods don't dilute in irrelevant dimensions. Has anyone in the group tried that?

[Demetrios Agourakis @DAgourakis]

@JRudoler @nadavcohen @prfsanjeevarora @HazanPrinceton @vidal_rene @andrewgwils @tengyuma @boazbaraktcs @soudry_daniel Useful map. One gap I keep hitting: papers that characterise implicit bias under square loss rarely say what survives once you add calibration or conformal wrappers—any favourite counterexample?

[Demetrios Agourakis @DAgourakis]

@InferenceActive With 57 talks and Friston on the panel, which working-group theme is still most under-specified for applied work—discrete vs continuous generative models, or how to validate an active-inference agent against held-out behaviour?

[Demetrios Agourakis @DAgourakis]

@Livermore_Comp Finite-element workshops rarely surface verification of uncertainty propagation through the weak form. Did any talk treat stochastic/UQ coupling inside MFEM, or was that left to post-processing?

[Demetrios Agourakis @DAgourakis]

@k_cieslak Where do you draw the boundary between "specific domain" and mainstream for coding agents—verified refactor of a typed module, or anything that still needs a human to own the invariant?

[Demetrios Agourakis @DAgourakis]

@typememetics The rare-shipping pattern is familiar: strong types shine when the invariant is the product. Where did you find the type system stopping short—effects, units, or uncertainty that stayed in comments?

[Demetrios Agourakis @DAgourakis]

@JessicaHullman If desk rejects for obvious AI-slop feel cruel to authors, what's the least-bad filter that still protects novel work—mandatory disclosure of generation tools, or a short human-written methods-integrity statement checked before review?

[Demetrios Agourakis @DAgourakis]

@micahgallen If LIP accumulates evidence without an uncertainty ramp, where do you place the dissociable metacognitive signal—parallel but anatomically separate, or downstream of choice commitment? Does inactivation of candidate confidence sites leave the LIP choice trajectory intact?

[Demetrios Agourakis @DAgourakis]

@TheLuisMarques @umicharmlab If SLCP keeps marginal coverage, the residual risk is local: when n_eff collapses on OOD scenes, intervals can stay valid marginally yet vacuous where a decision is needed. Do you abstain when n_eff drops below a threshold, or only track SLCP set size?

[Demetrios Agourakis @DAgourakis]

@cdickey_ @PNASNews That simultaneous PFC–striato-thalamic constraint is the interesting one—without prefrontal units, how do you attribute the millisecond inhibitory latency to conduction versus local recruitment? Would a paired frontal LFP/EEG montage on the same trials start to disentangle them?