Summer School in Computational Biology 2026 — University of Coimbra
Renato Duarte · CNC, CiBB, University of Coimbra renato.duarte@cnc.uc.pt
Seven students · basic Python, no neuroscience background required · laptop only — no cluster, no GPU, no data-access paperwork.
Neural population recordings are high-dimensional, redundant, and rarely interpretable at face value. The working hypothesis of modern systems neuroscience is that computation lives in a small number of latent modes shared across many neurons, rather than in individual cells. That hypothesis licenses almost all population-level analysis, and it is usually assumed rather than tested.
Three consequences, and they are the substance of this project:
- Structure is method-dependent. PCA, UMAP, Isomap and t-SNE impose different assumptions. Whether the class structure survives all of them is an empirical question.
- Visible structure is not usable structure. A separation you can see in three dimensions may be an artefact of the projection. Decoding accuracy is a common, defensible currency.
- Decodability does not localise. Knowing information is present says nothing about what carries it — and in a redundant code, the obvious answer is frequently wrong.
The payoff is the relationship between the three: does any geometric property of a representation predict what a downstream linear reader can extract, and can that capacity be traced to an identifiable subspace?
Spiking Heidelberg Digits — spoken digits pushed through a biophysical model of the inner ear and emitted as spike trains: ~700 channels, ~10 000 trials, 20 classes, 12 speakers. The answer is known for every trial, so every structural claim is checkable.
Four single trials. Each dot is one spike. Channel index is tonotopic — it maps monotonically onto frequency — so the diagonal sweeps are formant transitions. The structure is visible by eye; whether it is usable is the question the project answers.
git clone https://github.com/CNNC-Lab/compbio-2026.git
cd compbio-2026
conda env create -f environment.yml # or: mamba env create -f environment.yml
conda activate compbio-2026
jupyter lab notebooks/01_data.ipynbThe dataset downloads itself on first use and is cached in data/. Nothing else to fetch.
from compbio2026 import data, geometry, decoding, selection, plotting
plotting.apply_style()
shd = data.load("train")
X, y, t = data.build_design_matrix(shd, bin_ms=25.0) # (trials, channels, bins)
geometry.summarize(data.flatten(X), y) # participation ratio, class separation, ...
decoding.decode(data.flatten(X), y) # accuracy with its chance level
selection.compare_selectors(X, y, k=50) # sparse group lasso vs. heuristics, ablatedFull instructions, including what to do when something breaks: docs/00-setup.md.
Ten, in order. Every one is executed in this repository, so you can read the outputs before running anything. Notebooks 1 to 7 need only environment.yml; 8 to 10 need environment-routeb.yml because they simulate.
| Notebook | What you get out of it | |
|---|---|---|
| 1 | The data | What SHD is, and the two preprocessing choices that are scientific decisions rather than defaults |
| 2 | Geometry | Five embeddings, one table, and the shuffled-label control that makes the table mean something |
| 3 | Decoding | Accuracy with its controls; the published baseline reproduced; the redundancy result |
| 4 | Interpretability | Sparse group lasso, the heuristics it must beat, and ablation as the necessity test |
| 5 | Hair cells | Recovering the layer of the inner-ear model underneath the data |
| 6 | Filtering and readout | Spikes to continuous signals; what binning costs; pooling versus selecting channels |
| 7 | Decoding in the projection | Classifying inside an embedding, the leak that inflates it, and what clustering actually groups |
| 8 | Driving an HH layer | Spikes into conductance-based synapses — and the reduced pipeline that makes training affordable |
| 9 | The full spiking model | Route B's actual question: is the simulated layer worth more than the layer you were given? |
| 10 | A state matrix | One population vector per presentation, an AdEx layer, and why a single state vector is a rate code |
Every dimensionality-reduction method tells a different story about the same data. t-SNE and UMAP look the most convincing, which is exactly why they are dangerous: distances between clusters in those embeddings are not meaningful.
So the deliverable of Stage 1 is a table of numbers — participation ratio, class separation, trustworthiness — not a gallery of pictures. See docs/03-dimensionality-reduction.md for what each method assumes and where it misleads.
Train a linear readout, and give every number its controls: chance level, a shuffled-label null, a speaker-held-out split.
The right-hand panel is the important one. Accuracy is near its ceiling with a small fraction of the 700 channels — the code is massively redundant. That fact drives everything in notebook 4.
The naive question is "which neurons?", and it is the wrong unit. A population code is distributed: stimulus identity lives in a shared subspace, not in a list of important cells. Sparse group lasso (one group per cochlear channel) gives a candidate basis for that subspace; ablation tests whether it is necessary.
Remove the 50 channels a selector called essential and the decoder barely notices — the remaining 650 carry the same information. That number is the result, not a failed experiment. It is the honest, quantitative answer to "which neurons matter?" in a distributed code.
Required framing, to be stated explicitly in your report: selection and ablation here establish causal relevance for the decoder, not for the circuit. The data are fixed recordings and nothing is intervened upon in the biological system. Distinguishing decoder-causality from circuit-causality is a live confusion in the interpretability literature, and getting it right is part of the assessment.
SHD ships only the last stage of the LAUSCHER chain:
audio ──► basilar membrane ──► hair cell ──► bushy cell ──► SHD
(Meddis) (LIF, 40 afferents)
continuous spikes
probability
The hair-cell stage emits a continuous firing probability, and bushy cells map one-to-one onto hair-cell channels — the 40 afferents are stochastic draws within a channel, not a fan-in across channels. That makes recovery a well-posed per-channel inverse problem: measure the forward map by simulating the published bushy-cell model, then invert it.
It works, with limits that are part of the science rather than footnotes to it — a refractory ceiling that censors high rates, a smoothing kernel that is the time resolution, and a rate that identifies the model's output but not its internal state. See docs/04-hair-cell-recovery.md.
Pick one and justify the choice. They are different projects, not difficulty levels.
| Route | Finishes in 10 days | Simulation | Ground truth | |
|---|---|---|---|---|
| A | Analyse this population directly | reliably | no | yes |
| B | Use it as input to a biophysical circuit; analyse the second layer | if disciplined | yes | no |
| C | Decompose across layers of the auditory pathway | yes, for the recovery direction | some | yes, via LAUSCHER |
Trade-offs and risks: docs/05-routes.md.
| 00 · Setup | Install, download, troubleshoot |
| 01 · Dataset | What SHD is, where the channels come from, the splits that matter |
| 02 · Stages | What each stage is for, and what doing it well looks like |
| 03 · Dimensionality reduction | What each method assumes and where it misleads |
| 07 · Algorithms | Catalogue of methods with code and papers, from PCA to CEBRA |
| 04 · Hair-cell recovery | The inverse problem, its limits, and how to validate it |
| 05 · Routes | Three ways through the project |
| 06 · Reading | Three starred papers, and the rest on demand |
src/compbio2026/ data · geometry · decoding · selection · filters · lauscher · spiking · plotting
notebooks/ ten executed notebooks (1-7 analysis, 8-10 route B)
docs/ setup, dataset, stages, methods, recovery, routes, reading
tools/ SpikeList / StateMatrix, carried over from the 2025 course
data/ downloaded dataset (gitignored)
plots/ your figure output (gitignored)
- A short written report built around one defensible result, with methods, controls and stated limitations
- Final presentation, Saturday 12 September
- Reproducible analysis code, in this repository
A negative or null result, properly established and controlled, counts fully.
The stages are a spine, not a recipe. You design the plan and the experiments.
| Date | |
|---|---|
| Wed 2 Sep, 17:00 | Project pitch; Pizza & Games 19:30, Casa das Artes |
| Thu 3 Sep, 14:00 | Student distribution, room F42 — work begins |
| Fri 4 – Tue 8 Sep, 17:00–18:00 | Daily project blocks |
| Wed 9 Sep, 10:00–13:00 & 15:00–17:00 | Computational Neuroscience / NeuroAI teaching block |
| Wed 9 Sep, 17:00 | Beer & poster session, Physics Department atrium |
| Thu 10 Sep, 16:00–18:00 | Progress reports |
| Sat 12 Sep, 15:00 | Final presentations |
The final three days are hosted at CNC-UC, Polo I (Rua Larga, Faculty of Medicine building).
Dataset: Cramer, Stradmann, Schemmel & Zenke (2020), IEEE TNNLS. Conversion chain: LAUSCHER. tools/ from the 2025 course repository.