Research Survey: How Do You Validate AI Outputs Derived from Brain Signals?

Announcements
,
1

Hi Neurostars community,

I’m conducting an early-stage research survey to understand how researchers and practitioners validate AI-generated outputs derived from neural signals, including EEG, fMRI, MEG, and ECoG.

One question I’m particularly interested in is:

How do you determine whether an AI-generated output genuinely reflects the original neural signal, rather than patterns learned by the model or other unrelated factors?

For example, when a model generates text, images, or other representations from neural data, what methods do you use to evaluate how much the output actually depends on the input signals?

The survey explores:

  • Current validation methods and workflows

  • Difficulties encountered when evaluating neural signal-based AI outputs

  • Potential needs for independent validation tools

If you work with neural signal analysis, brain decoding, neuroimaging, or related machine learning applications, I would greatly appreciate your perspective.

Survey link (approximately 4–6 minutes):

Participation is voluntary. Responses will be used for early-stage research and customer needs assessment.

I’d also be interested in hearing about validation approaches you currently use, if you’d prefer to share your experience in the discussion.

Thank you for your time and insights!

2

I’d separate “does the output depend on the neural input?” from “is it decoding the intended variable?”. One possible control is to compare correctly paired trials with mismatched neural inputs while keeping prompt and stimulus conditions fixed, using a null that respects the trial/session structure. A drop would support input dependence, but shared eye movements or task timing could still explain it.

Are those two validation questions separated in the survey? They seem likely to need different tools.