Decision Geometry

Real-data computational neuroscience with IBL Neuropixels recordings

How do sensory evidence, an animal's learned prior, and its eventual choice emerge in the activity of many neurons at once? Decision Geometry turns a published International Brain Laboratory (IBL) recording into a compact, reproducible population-dynamics analysis.

The pipeline streams real spike trains and behavior from DANDI, aligns neural activity to each decision, and asks not only whether task information is decodable, but also when it appears, how stable its neural representation is, and which recorded regions contain it.

Launch the interactive Decision Geometry explorer to scrub through time and switch between information, code-stability, and regional views.

Main result

This figure is generated by the repository from the pinned public NWB asset. It is not a mockup or synthetic example.

Reading the figure

A. Choice-conditioned neural trajectories. Each line is the average population state for one wheel choice, projected onto the first two principal components. The separation between trajectories shows that the activity of the recorded population evolves differently for clockwise and counter-clockwise decisions. Circles mark the start of the analysis window at -0.5 s; squares mark its end at +1.0 s.

B. Information over time. Cross-validated logistic decoders estimate how well a linear readout can recover stimulus side, block prior, or choice from one 50 ms population snapshot. Shaded bands are 95% class-stratified bootstrap intervals computed from held-out predictions. Chance is 0.5. Choice peaks at 82.8% balanced accuracy around 275 ms after stimulus onset with a 79.5%-86.1% interval at that time; stimulus side reaches 79.3%, and block prior reaches 69.6%.

C. Cross-temporal choice decoding. A decoder is trained at every time point and tested at every other time point. The bright post-stimulus band shows that choice information generalizes across nearby times, while its changing shape shows that the neural code is not perfectly static.

D. Regional decoding. The same choice decoder is fit separately to regions with at least five quality-filtered units. Posterior and lateral-posterior thalamic populations show strong decision-related information, with additional signals in cuneiform, trigeminal motor, and parabrachial nuclei. Region-level scores should be compared cautiously because neuron counts differ.

The scientific question

Decision-making is distributed across populations and brain regions. A single neuron can correlate with a stimulus or movement, but cognition is more naturally described as a trajectory through a high-dimensional population state. This project asks four connected questions:

1. Do leftward and rightward choices occupy distinct neural trajectories?
2. When do stimulus, prior, and choice become linearly readable?
3. Is the choice representation stable across time or dynamically reformatted?
4. Which recorded anatomical regions carry the strongest choice information?

The gap this project addresses

Large open neurophysiology archives are scientifically valuable but difficult to turn into a clear, reproducible computational story. The source asset here is about 916 MB, uses ragged spike-time arrays, and combines neural, behavioral, and anatomical tables. Many introductory analyses stop at rasters or mean firing-rate plots; same-time decoding alone also cannot reveal whether the underlying code is stable or changing.

Decision Geometry bridges that gap by providing:

• Practical access: HTTP byte-range streaming reads only the required NWB slices instead of downloading the whole recording.
• Reproducible preprocessing: explicit IBL unit-quality thresholds and a cached trial-by-unit-by-time tensor whose source and analysis parameters are validated before reuse.
• Population-level interpretation: PCA trajectories connect individual spikes to low-dimensional neural dynamics.
• Temporal tests: time-resolved and cross-temporal decoding distinguish information strength from representational stability.
• Anatomical comparison: the same analysis is repeated across recorded regions using one consistent evaluation procedure.
• Uncertainty reporting: held-out predictions are resampled within each class to provide deterministic 95% intervals around the main decoding curves.

This is the useful middle ground between a basic dataset-loading tutorial and a full consortium-scale reanalysis.

Data source

The analysis uses DANDI:000409, IBL Brain Wide Map, a public collection of Neuropixels electrophysiology and behavioral data from mice performing a visual decision task. The broader release contains recordings from many laboratories and brain areas; this repository intentionally pins one session so the demonstration remains reproducible and computationally practical.

Field	Pinned value
Dandiset	000409
Published version	0.260309.1324
Asset ID	882e2ff6-1fde-4518-8797-5d5892379739
Subject	NYU-39
Session	6ed57216-498d-48a6-b48b-a243a34710ea
NWB asset	sub-NYU-39_ses-6ed57216-498d-48a6-b48b-a243a34710ea_desc-processed_behavior+ecephys.nwb
Source size	approximately 916 MB
Source license	CC BY 4.0

The NWB file provides:

• Spike times and quality metrics for 1,366 sorted units.
• 541 behavioral trials with stimulus side and contrast.
• Wheel choice, movement onset, feedback, and reward outcome.
• Blockwise prior probability for left and right stimuli.
• Electrode coordinates and anatomical region labels.

After strict quality filtering, the default analysis retains 93 units across 12 regions. The strongest unit coverage in this session is in posterior and lateral-posterior thalamus, so this is a focused sample from a brain-wide collection, not complete whole-brain coverage in one animal.

How it works

flowchart LR
    A["Published IBL NWB on DANDI"] --> B["Remote byte-range reader"]
    B --> C["Trial and unit metadata"]
    B --> D["Ragged spike times"]
    C --> E["IBL quality filtering"]
    D --> F["50 ms stimulus-aligned bins"]
    E --> G["Trials x units x time tensor"]
    F --> G
    G --> H["PCA trajectories"]
    G --> I["Time-resolved decoding"]
    G --> J["Cross-temporal decoding"]
    G --> K["Region-wise decoding"]
    H --> L["Result dashboard and JSON summary"]
    I --> L
    J --> L
    K --> L

Loading

1. Stream and validate the NWB file

decision_geometry/data.py resolves the pinned DANDI asset, opens it through remfile and h5py, and reads the standard NWB trials, units, and electrodes tables. Only requested byte ranges are transferred.

2. Select reliable units

Units must satisfy all of the following:

• ibl_quality_score == 1.0
• presence_ratio >= 0.9
• firing rate between 0.1 and 100 Hz
• a valid maximum-amplitude electrode with an anatomical label

3. Construct population activity

For every valid choice trial, spikes are aligned to visual stimulus onset and binned from -0.5 s to +1.0 s in 50 ms bins. This produces a tensor with shape trials x units x time. The derived tensor is cached locally; raw data is never committed to Git.

4. Measure geometry and information

• PCA standardizes unit activity and estimates low-dimensional trajectories.
• Time-resolved decoding uses five-fold stratified cross-validation and balanced accuracy to decode stimulus side, prior, and choice independently.
• Uncertainty intervals use 1,000 class-stratified bootstrap resamples of out-of-fold predictions at each time bin.
• Cross-temporal decoding trains at one time bin and tests at all bins, producing a train-time by test-time stability matrix.
• Regional decoding repeats choice decoding for anatomical regions with at least five retained units.

All classifier scaling is learned inside each training fold. The logistic model uses a fixed regularization setting and class balancing, keeping the comparison consistent across time and regions.

Reproduce the analysis

git clone https://github.com/josephreggy23-coder/IBL-Brain-Wide-Map.git
cd IBL-Brain-Wide-Map
python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -e ".[dev]"
decision-geometry
pytest

The first run creates data/cache/session_population.npz. The cache records its DANDI asset, published version, time window, bin size, and unit limit. Later runs reuse it only when all of those values match; a changed analysis request rebuilds the tensor instead of silently returning stale data. Results are written to the dashboard and results/summary.json.

Useful options:

# Re-read the public NWB source and rebuild the cache
decision-geometry --force-stream

# Run a smaller population or change temporal resolution
decision-geometry --max-units 64 --bin-size 0.05

# Record a different deterministic cross-validation split in the summary
decision-geometry --seed 42

The JSON summary records the selected random_seed, bin_size_s, and max_units_requested, making the data-derived metrics and analysis settings explicit when results are compared.

--max-units must be a positive integer. When a limit is requested, eligible units are ranked by presence ratio and firing rate before the top units are used, so smaller exploratory runs remain deterministic.

Results at a glance

Measurement	Result
Valid trials	541
Quality-filtered units	93
Anatomical regions represented	12
Peak choice balanced accuracy	82.8%
Peak choice time	275 ms after stimulus onset
Choice 95% interval at peak time	79.5%-86.1%
Peak stimulus-side balanced accuracy	79.3%
Peak block-prior balanced accuracy	69.6%

Exact machine-readable values are stored in results/summary.json.

Interpretation and limits

Above-chance decoding means that a linear model can read information from the recorded population. It does not establish that a region causes the choice. Likewise, PCA is a descriptive projection rather than a mechanistic model of the circuit. The interval reported at the peak time quantifies trial-sampling uncertainty at that selected bin; it is not a correction for searching across time and should not be read as a confirmatory significance test.

This repository analyzes one session to provide a transparent, runnable example. A research claim about the full IBL population would require held-out sessions, animals, and laboratories; uncertainty estimates or permutation tests; controls for movement and reaction time; and region comparisons matched for unit count. Those are natural extensions rather than conclusions claimed here.

Repository structure

decision_geometry/
  data.py        remote NWB access, quality filtering, spike binning, cache
  analysis.py    PCA, time-resolved and cross-temporal decoding
  plotting.py    four-panel scientific result figure
  pipeline.py    end-to-end orchestration and summary export
  cli.py         command-line interface
scripts/
  export_web_payload.py  exact analysis-to-explorer data export
results/
  decision_geometry.png
  summary.json
tests/           deterministic tests for filtering, cache integrity, and decoding
web/             interactive explorer and downloadable analysis receipt

This repository is intentionally scoped to Decision Geometry. Other experiments, personal profile repositories, and unrelated literature downloads are not part of main.

References and attribution

• International Brain Laboratory. A brain-wide map of neural activity during complex behaviour. Nature 645, 177-191 (2025). DOI: 10.1038/s41586-025-09235-0.
• IBL 2025 Brain Wide Map data-release documentation.
• DANDI Archive Dandiset 000409.
• Neurodata Without Borders.

Code in this repository is MIT licensed. The source data is CC BY 4.0; cite the IBL publication and DANDI record when using it for research.