Seismologist here. The problems you describe sound reminiscent of those in seismic imaging. Ideally you'd bury geophones below the weathered layer (typically 3 m) at the surface, to get ideal coupling. In practice that's not economic at scale, and so you plant them on the surface, and account for the non-linear wave transmission through the weathered layer by clever math and by collecting more samples.
There's been a forty-year evolution in these techniques. The cheap, noisy technique might prevail if scientists keep refining the craft by tiny improvements.
I looked into the problems in this field in the past, my impression is that the difficulty lies in how the signal diffuses through the skull and top-most layers of the brain. It’s all about resolution, and this sounds like a very similar problem. The question always is how much non-linearity is there, and what’s the frequency and density of the sensors relative to how often you have to sample to characterize it. Very excited to see how the field develops!
Neurologist here. There is an inherent problem of physics. The skull acts as a low-pass filter of the brain's activity. You lose a lot of information if you don't go under the skull.
I don't think EEGs can really give the spatial and temporal resolution we really need to extract the necessary information for thought decoding (or encoding).
you don't need to decode thought, you just need to decode information at some really low bitrate to export information (say text).
Working with a NN, the brain can probably negotiate the relevant parameters (let's say frequency modulation across one channels with four bins, across 4 spots on the brain), that's 4^4 = 64 values across whatever timebox of resolution you get. That's enough to encode english letters, which itself probably is an underprovisioned mechanism of data transfer).
Give the user real time feedback from the reader NN.
I wouldn’t be surprised if the biofeedback allows the user to retrain their brain to encode information in low frequency signals—the raw channel capacity necessary to transmit text is quite low, plus the reader NN can also use context if hooked up to a recurrent submodel.
There's been a forty-year evolution in these techniques. The cheap, noisy technique might prevail if scientists keep refining the craft by tiny improvements.