Paula Peixoto Moledo

Paula Peixoto Moledo, PhD

Computational neuroscientist reverse-engineering systems that were never designed to be understood — neural circuits, behavior, and lately, transformers.

paulapeixotomoledo@gmail.com LinkedIn GitHub Scholar Madrid, Spain — open to relocation
PythonPyTorchTensorFlow Simulation-based inferenceNormalizing flows Latent-variable / state-space modelsNeural decoding Causal intervention & attributionGPU/HPC pipelines

Experience

Cajal Neuroscience Center, CSIC, Madrid
Nov 2025 – Present
Post-Doctoral Researcher — Systems Neuroscience
Computational Neuroscience of Memory Lab, Cajal Neuroscience Center – CSIC, Madrid
  • Develop and maintain scalable Python pipelines for multimodal neural/behavioral analysis (preprocessing, model training, evaluation, visualization).
  • Implement and benchmark latent-state inference frameworks to quantify strategy dynamics and learning.
  • Collaborate across experimental and computational teams; contribute to manuscript preparation, figures, and method write-ups.
Sep 2020 – Nov 2025
PhD Researcher — Systems Neuroscience
Instituto Cajal – CSIC, Madrid / IDIBAPS – Hospital Clínic, Barcelona
  • Led end-to-end research on memory interference and adaptive behavior, integrating experimental neuroscience with probabilistic and ML-based modeling.
  • Developed probabilistic models to infer latent behavioral states, and simulation-based inference for parameter estimation, model comparison, and predictive evaluation, in PyTorch and TensorFlow.
  • Acquired and analyzed in vivo calcium imaging data; separately analyzed preliminary fiber photometry recordings acquired by colleagues. Applied pose estimation (DeepLabCut) for behavioral feature extraction.
  • Built integrated hardware–software systems (Arduino) to synchronize behavioral events with imaging acquisition.
  • Mentored Bachelor's and Master's students on thesis projects.
IDIBAPS, Hospital Clínic courtyard, Barcelona
Jun 2019 – Sep 2020
Research Technician — Systems Neuroscience
IDIBAPS – Hospital Clínic, Barcelona
  • Experimental and analytical research on immune-mediated synaptopathies affecting memory and behavior.

Education

Universitat de Barcelona logoUniversitat de Barcelona logo
PhD in Biomedicine (Neuroscience)
Universitat de Barcelona · Oct 2020 – Nov 2025

Thesis: "Interference-Based Forgetting in Spatial Navigation: Latent Behavioral Signatures and the Impact of Synaptic Plasticity Disruption."

UAB logo
MSc Translational Biomedical Research
Vall d'Hebron Campus / UAB · 2017–2018
UAB logo
BSc Nanoscience and Nanotechnology
UAB · 2013–2017

Publications

Posters & Awards

Additional

MODEP — Took part as one of three team members in MODEP, a proposed AI platform to decode and predict parkinsonian motor states from continuous movement data for therapy optimization, presented at ACTIVA-T, the CSIC programme for knowledge-based companies (Vigo, May 2026).

Paula Peixoto Moledo in a Mars-analogue spacesuit during a WoMars training mission

Co-founder, WoMars (womars.org) — Geneva-based non-profit organizing all-female space analogue missions and training. Mission logistics, team coordination, partnerships, and fundraising.

Advanced training: EMBO Practical Course — Advanced Modeling of Behavior (2022, Barcelona).

Languages: Spanish, Galician (native); English, Catalan (fluent); basic Italian, French, Portuguese.

Three research projects from my PhD and postdoc work, each built around the same habit: construct a model that recovers structure from noisy, high-dimensional biological data, then trust the result only after it survives an attempt to break it.

Thesis cover art designed by Paula Peixoto Moledo: a mouse at the centre of a radial maze dissolving into Barcelona's street grid

Causal attribution in a biophysical circuit model

Simulation-based inference to identify the mechanism behind a pharmacological effect — manuscript in preparation

~25,000 neurons 50,000 simulations 8-parameter inference GPU + HPC cluster

Latent-state decomposition of behavior

Hierarchical GLM-HMM recovering interpretable strategies from navigation data

K = 1–10 states tested leave-one-animal-out CV custom circular-data observations

Year-scale population decoding from durable single-neuron recordings

Second author, IEEE MetroXRAINE 2026 (accepted)

7 animals 30–42 sessions each 4 adversarial controls

Simultaneous behavior tracking and hippocampal calcium imaging — raw miniscope feed, processed ΔF/F traces, and extracted single-cell activity (8 example cells shown).

Real-time position decoding from that same hippocampal activity (179 neurons, ridge decoder, cross-validated on held-out data) — decoded position shown 200ms ahead to account for the calcium-signal delay.

CSIC logo
Data collected at CSIC (Consejo Superior de Investigaciones Científicas), Jercog Lab.

Are attention head contributions additive?

Personal project, August 2026 · github.com/paulapeixotomoledo/interp-induction-shapley

A from-scratch replication of induction-head behaviour in GPT-2 small, extended with an exact Shapley decomposition of head ablations. The induction-head finding itself is established literature — the point of this project is six things I found while checking it that I was not looking for, each one caught by a control before I believed the wrong number. Every chart below is built directly from this project's own result files, not illustrative data.

Finding 1 — the counterfactual changes the answer

Does how you remove a head change which head looks important?

Five induction heads, ablated individually and across all 32 subsets; Shapley values computed exactly (not sampled) so they sum to the joint effect exactly, under three different ablation counterfactuals — zero (force output to 0), mean (force output to its batch mean), and resample (swap in the head's output from a different input sequence).

Finding 2 — most of the "interaction" was a choice of metric

Is 85% of the effect really interaction — or is that a property of measuring in log-loss?

I first reported "85% of the effect is interaction." That number is real for cross-entropy loss specifically — it drops to 21% under a near-linear metric (centred logit), measured on the same 32 subsets, same Shapley computation, same forward pass. Error bars are ± 1 SD across 5 random seeds. The README was corrected in place rather than rewritten to look right the first time.

Finding 3 — the same head, twenty times the cost, depending on order

Is a head's importance a fixed property of the head?

L4H11 (the dominant previous-token head feeding the induction circuit) costs +0.250 nats when ablated first in a cumulative removal sweep, and +5.23 nats — the same head, same model, same metric — when ablated last, after the heads it was backing up are already gone. "How important is this head" turned out not to be a well-posed question without specifying what else has already been removed.

The open question

Loss per token position across [BOS] + S + S (50 random tokens, repeated). The second copy is perfectly predictable given the ability to look back — but the loss does not collapse at the first repeated token. It takes 3–4 positions to reach floor, and only one of those is explained by there being no prior occurrence yet to attend to at the exact boundary. I don't know what accounts for the other two or three positions, and the repo notes say so rather than papering over it.

Finding 4 — attention and causal importance disagree

Does the head the model "looks at" most match the head that actually matters?

L5H5 places more attention on the induction diagonal than any other head (0.907, the correlational measure from Experiment 1) — but ablating it alone costs almost nothing (+0.046 nats). L5H1, ranked third by attention, carries almost the entire marginal causal signal (+0.640 nats). Attending to the right place and being load-bearing turned out to be different properties of the same head — which is the reason this project moved from scoring attention to ablating it.

Finding 5 — where induction actually lives, across all 144 heads

GPT-2 small has 12 layers × 12 heads. How concentrated is this behaviour?

Every one of the 144 attention heads in GPT-2 small, scored for attention placed on the induction offset diagonal. The five outlined cells are the heads used throughout this project. Scores don't fall off smoothly — there's a clear gap between the 5th-ranked head (0.814) and the 6th (0.509), then a second cluster at 0.45–0.51. Two tiers, not a gradient; whether that second tier is doing weaker induction or something else this metric happens to pick up is still an open question in the repo notes.

Finding 6 — checking whether a surprising result was noise

A head's ablation effect was slightly negative — removing it looked like it helped. Real, or noise?

L7H10's marginal ablation effect was negative on the first run and small enough to write off as noise. Shown here: all 5 independent random seeds, plotted individually rather than only as a mean. Every single one is negative — about six standard deviations from zero. It wasn't noise, and I had no explanation for it then and don't invent one now. L5H1 is plotted alongside for scale: a clearly real, clearly positive effect, for comparison.

Interested in working together, or just want to talk about any of this?

paulapeixotomoledo@gmail.com Download CV (PDF) LinkedIn