001BASELINE
Decoder evaluation across recording days
Two fixed classifiers, their confusion patterns, and the uncertainty behind a change in ordering.
RESEARCH AT DAYBREAK
Neural decoding,
beyond a
single session.
Exploring more reliable EEG decoding with less repeated calibration. A research project built around a simple question: what should a decoder carry into the next day?
Seven completed studies. New questions, documented results.
CONTINUITY / 007RESEARCH STAGE
Changing signals.
A search for lasting structure.Conceptual illustration
Not measured EEG
A search for lasting structure.Conceptual illustration
Not measured EEG
THE QUESTION THAT DRIVES US
What would it take for a neural decoder to learn today
and need less relearning tomorrow?
THE OPEN-SOURCE LEARNING LAB
Neural Dynamics Explorer
Change circuit parameters, follow population activity, and connect differential equations to phase portraits and stability.
These are simplified mathematical simulations, not measured EEG. The complete explorer and its MIT-licensed source remain available.
Open the Dynamics Explorer01 / OUR DIRECTION
An adaptable decoder
starts with better questions.
From supervised adaptation to learned augmentation. Each comparison tracks calibration cost, participant-level trade-offs, and uncertainty.
Comparisons completed · Study 002
Start with alternatives that deserve a fair test.
Our calibration study broadened the comparison beyond the original bandpower baseline to spatial and covariance-based pipelines. The same recording protocol lets us examine where the representation matters before adding a more complicated adaptation method.
QUESTION TO TEST
Does a proposed method improve on a strong, simpler baseline under the same data access?
02 / THE CURRENT EXPERIMENT
More training examples.
More useful decoding?
Study 007 compared generated examples with real-only training and simple augmentation. This first pilot found lower average accuracy with generated EEG at the fixed primary budget.
Teach with variations.
Test on reality.
Three training arms share the same compact EEGNet-style decoder, real-label budget, source checkpoint, and optimization budget. They differ in the examples used during adaptation.
The learned generator produces covariance patterns and uses them to synthesize band-limited EEG. This is a constrained conditional VAE experiment; it does not generate the full range of brain dynamics.
WHAT THE PILOT SAYS
Read Study 007 ↗Generated augmentation performed worse on average than real-only training at the fixed 24-label budget in this eight-person pilot.
Real day-two calibration trials
0%25%50%75%100%
Real EEG only63.51%
Real + simple augmentation63.41%
Real + generated EEG61.14%
Classical covariance reference74.70%
50% is the two-class chance reference. These eight participants were previously examined in Studies005–006. Three seeds quantify training variability; they do not make 24 independent people. The classical reference is recomputed on this pilot’s inputs.
FIXED PRIMARY COMPARISON
Generated augmentation versus real-only EEGNet · 24 real calibration trials
−2.37percentage points
Descriptive 95% participant interval:
−4.31 to −0.42 pp
1 improved · 7 worsened · 0 tied
After averaging three seeds per participant
Small, nonrandom development cohort. The interval assumes independent, approximately normally distributed participant differences. It does not establish reliability for new people or future days.
See every participant at 24 calibration trials
| Participant | Real only | Simple | Generated | Difference |
|---|---|---|---|---|
| 2 | 59.62 | 59.62 | 53.85 | −5.77 |
| 3 | 86.57 | 86.57 | 83.84 | −2.73 |
| 4 | 52.88 | 53.64 | 49.90 | −2.98 |
| 5 | 61.87 | 61.67 | 62.37 | +0.51 |
| 6 | 82.47 | 82.47 | 77.22 | −5.25 |
| 7 | 52.63 | 52.07 | 52.27 | −0.35 |
| 8 | 51.87 | 51.11 | 49.60 | −2.27 |
| 9 | 60.15 | 60.15 | 60.05 | −0.10 |
Fixed protocol · Real test data · Participant-level reporting
Design snapshot · ZIP ↗Code + results · ZIP ↗PREVIOUS FINDING / STUDY 006
The update gate missed its targets.
Across 53 OpenBMI participants, the gate reached 67.74% mean balanced accuracy, versus 71.87% for always updating with 24 labels. Mean loss burden fell by 20.6%, below the fixed 25% target. Both engineering targets failed.
The gate trained on 16 labels and spent 8 on selection. These are development results on the same participants as Study005. The simpler supervised update remains the benchmark.
Read the complete result ↗THE RESEARCH RECORD
A traceable progression.
Completed experiments remain available, including unfavorable comparisons and the assumptions behind each result.
Explore the archived Study 003 interactive results
A smaller adjustment.
A more careful test.
Can blending an alignment step reduce a known failure without losing what a simple label-informed update already achieves?
At a budget of 8 original trials, participant 3 improved from 52.00% with full alignment + update to 78.64% with conservative alignment + update. Updating alone reached 79.53%. Participants 1 and 2 performed worse with the conservative version than with full alignment at that budget.
WHAT THIS SUPPORTS
Read Study 003 Conservative alignment mitigated a known failure in this development analysis. It did not demonstrate an advantage over updating alone at the primary budget.
Calibration budget (original trials)
20–24 accepted calibration trials per participant after excluding artifact-flagged trials.
Budget 24 original trials. Equal-person mean for participants one through four.
0%25%50%75%100%
Frozen decoder66.00%
Update only67.47%
Full alignment + update66.73%
Conservative alignment + update67.47%
Dashed reference: 25% uniform guessing. All four participants were previously examined development data. These chart views are descriptive; changing the view does not change the fixed primary comparison below.
FIXED PRIMARY COMPARISON
Conservative alignment + update
versus updating alone
+0.007 percentage points
Conditional 95% interval: −0.67 to +0.61 pp
Equal-person mean, participants 1–4
Budget: 24 original calibration trials
Means: 67.47% updating alone · 67.47% conservative alignment + update.
No demonstrated advantage.The interval describes test-run resampling for these four participants, fixed models and calibration prefixes. It does not measure uncertainty for new participants or future days.
Historical development analysis Study 003 used previously examined participants 1–4. Study 004 subsequently evaluated the frozen method on reserved participants 5–9. Rerunning identical data checks reproducibility, not generalization.
THE RESEARCH RECORD
The completed study archive.
Read the findings, inspect the code,
and keep the limitations in view.
002CALIBRATION
Comparing calibration budgets
Broader pipeline comparisons and adaptation strategies under defined access to calibration data.
003ALIGNMENT
Testing conservative alignment
A targeted failure mitigation, compared with full alignment and updating alone. The primary advantage remains unproven.
004COMPLETED
Evaluation on reserved participants
Five BNCI participants. Conservative alignment minus updating: −0.765 pp; 95% participant interval −2.250 to +0.720. No demonstrated added benefit.
005COMPLETED
An independent dataset: OpenBMI
53 reserved participants. At 24 calibration labels, updating gained +6.38 pp on average; 38 improved, 14 worsened, and one tied.
006COMPLETED
Can a small validation set select updates?
The gate missed both targets: −4.14 pp versus always-update24, with a 20.6% reduction in mean loss burden. Exploratory reuse of Study 005 participants.
Running Study 003 on Windows ARM64?Download environment requirements
03 / THE PATH FORWARD
One question.
One honest experiment at a time.
The evidence has expanded to a second dataset. The next method still has to earn its place against a simpler baseline, then face new sessions and new people.
- 01COMPLETED BASELINE
Make the comparison reproducible.
Document the data, exclusions, model choices, errors, and uncertainty in a runnable notebook and working paper.
- 02COMPLETED DEVELOPMENT ANALYSIS
Measure the calibration trade-off.
Compare established pipelines and adaptation strategies at defined calibration budgets on participants 1–4.
- 03COMPLETED DEVELOPMENT ANALYSIS
Test a more conservative alignment.
Mitigate a known failure and compare with update-only. The primary 24-trial result shows no demonstrated advantage.
- 04COMPLETED · STUDIES 004–006
Expand the evidence. Keep the failures.
Evaluate reserved BNCI participants, test supervised updating on OpenBMI, and retain the update gate’s failed development screen.
- 07COMPLETED · DEVELOPMENT PILOT
Test learned augmentation.
Compare real-only, simple, and generated training examples at the same real-label budget. Use this pilot to decide whether a larger, independently evaluated study is warranted.
A FEW USEFUL DISTINCTIONS
A clear view
of the work.
What are you building?
We are investigating EEG decoding methods that need less repeated calibration. This section preserves our reproducible offline studies. Decoder Studio turns evaluation concepts into a guided browser workspace, while the research explores methods we have not established as superior commercial decoders or clinical systems.
What have the earlier studies established?
Study 004 found no demonstrated benefit from conservative alignment over updating alone on five reserved BNCI participants. Study 005 tested supervised updating on an independent dataset, OpenBMI: the mean gain at 24 labels was 6.38 percentage points across 53 participants, but 14 people worsened. Study 006 then reused those people to test an update gate; it missed both development targets.
Is Study 007 independent validation?
No. It is a fixed eight-person development pilot using OpenBMI participants already examined in Studies 005 and 006. The protocol was saved before the new computation, but this is not external preregistration or prospective validation. Training seeds and generated examples do not increase the number of independent people.
What does generated EEG mean here?
A small conditional variational autoencoder learns covariance patterns from permitted real training data. A stochastic synthesis step converts those patterns into band-limited signals. It is a restricted generator, not a model of complete brain dynamics. Generated examples are used only for training; performance is measured on real later trials.
What would make this useful to another research team?
A worthwhile pilot would test a fixed decoder against that team’s baseline on new sessions, measuring accuracy, calibration time, individual regressions, and integration cost. Our present work provides an inspectable research starting point, not a promise of those benefits.
LET’S START A CONVERSATION
The next question
could begin with you.
Have a research question, a potential collaboration, or feedback on the work? Get in touch with our team.