ESTIMATE
Predict the response
Multimodal models take audio, video, and text and estimate the stimulus-evoked cortical response, second by second.
Canvascope Research
Predicting how the cortex responds — and proving it.
We build computational tools that estimate the cortical response a stimulus evokes, map that estimate onto interpretable brain regions, and hold it to the standard this field has rarely met: it has to work on material it has never seen.
Berkeley, California
MAP
Put it somewhere legible
High-dimensional predictions are summarized across interpretable cortical regions and functional networks, so a result can be inspected rather than trusted.
VALIDATE
Test it on what it hasn't seen
Predictions are measured out-of-sample against a strong baseline — never against nothing, and never in-sample fit dressed up as prediction.
The work
What we're actually working on.
Multimodal brain encoding
Models that map naturalistic audio, video, and language onto stimulus-evoked cortical response patterns.
Cortical region and network mapping
Reducing vertex-level predictions to region and network summaries that a researcher can read, argue with, and check.
Cross-subject decoding
Whether a model pretrained across many people can read a new person's signal from minutes of calibration instead of a full session.
Leakage-audited evaluation
Pre-committed splits, counterfactual unit tests, and audits that fail loudly — built to catch the failure mode this field is known for.
Pre-registration and provenance
Endpoints registered before the analysis runs, with execution lineage and receipts kept so a result can be reproduced by someone who doubts it.
The critiques, collected
Failed replications, reverse-inference critiques, and reliability limits are gathered as carefully as the supporting work. They decide what we don't say.
Claim discipline
The limits are part of the work.
Prediction is not identification, and neither is explanation. We keep the three separate in every result we publish.
We predict responses, not states of mind. Reading a mental state back out of a response pattern is the inference that discredited this field. We don't make it.
Out-of-distribution performance degrades. Models trained on one kind of material do worse on another. We expect that, measure it, and say so first.
A result is one signal among several. Our work belongs alongside other evidence in a research workflow. It is not a verdict, and never a readout about an individual person.
How a result becomes a claim.
STEP 01
Register the endpoint
The question, the splits, and the success criterion are written down before any modelling starts. Post-hoc promotion doesn't count.
STEP 02
Beat a strong baseline
Held-out performance is compared against the best simple alternative — content features, metadata, self-report — not against a straw man.
STEP 03
Publish it either way
The result goes out honestly, including when the baseline wins. A claim we can defend under scrutiny is worth more than a broad one that isn't.
Come argue with the evidence.
We work with people across machine learning, neuroscience, neuroimaging, EEG, and measurement methodology — on study design, model evaluation, data partnerships, and tooling. The shared standard is careful validation and claims that survive a hostile read.