Computational neuroscientist reverse-engineering systems that were never designed to be understood — neural circuits, behavior, and lately, transformers.
Thesis: "Interference-Based Forgetting in Spatial Navigation: Latent Behavioral Signatures and the Impact of Synaptic Plasticity Disruption."
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).
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.
Simulation-based inference to identify the mechanism behind a pharmacological effect — manuscript in preparation
Hierarchical GLM-HMM recovering interpretable strategies from navigation data
Second author, IEEE MetroXRAINE 2026 (accepted)
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.
Are attention head contributions additive?
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.
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).
Is 85% of the effect really interaction — or is that a property of measuring in log-loss?
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.
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.
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.
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.
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?