brain reading and writing

off-the-cuff broad strokes

You can basically think of wearable neurotech about doing two things with the brain without drilling holes in the skull - reading (recording) and writing (stimulation).

Reading -

You can read two functional signatures from the brain to measure its activity -

1. Electrical activity from neuronal activations -

a) Using EEG - measuring the actual millivolts on the scalp from summed up neuronal action potentials in the brain

• dog-eared, well-understood, temporal resolution in 'ms, spatial resolution in cms, limited to cortex

b) Using optically pumped MEG - measuring the magnetic fields on the head from the flow of current in neuronal circuits in the brain

• MEG is half-a-century old but optically pumping it to make it wearable is around a decade new, temporal resolution in ‘ms and spatial resolution in mm, can go to deeper brain regions

2. You can record metabolic activity from changes in blood flow -

a) Using fNIRS/HD-DOT - the fluctuating oxygenation of blood via the diffrential diffraction of infrared light based on hemoglobin saturation

• also around a half-century old and well-understood, spatial resolution in mm, temporal resolution in '00ms, limited to cortex

b) Using fUS - using the doppler effect to measure the movement of blood

• newer, temporal resolution in low '00ms, spatial resolution in μm, can go deep

Writing -

You can write to the brain in three ways -

1. By electricity

a) direct current (tDCS - transcranial direct current stimulation)

b) alternating current (tACS - transcranial alternating current stimulation)

2. By magnets (TMS - transcranial magnetic stimulation)

3. By ultrasound (LIFU - low-intensity focused ultrasound)

Reading or Writing?

You always want to have better reading capabilities than writing at any point in time, because you don’t fully know what and how well you’re writing otherwise. If this is true, then the biggest moat to a good writing model would be a very good reading model.

Good models of reading from the brain are downstream of a good data-set, specifically four things -

a) spatiotemporal resolution - <mm and <ms

b) whole-brain view - % of brain covered

c) scale - individuals x time

d) context - across the spectrum of sensing, moving, feeling, and decision-making

While a) and b) are highly contingent on technology with upper-limits decided by our SoTA understanding of physics and materials, c and d depends on form-factor, adoption, distribution, use-case of PMF and savvy ground-truth generation. Models with excellent a and/or b can likely be outclassed by ones with better c or d, and given there is a certain inverse correlation between the two (e.g. better resolution generally comes at the cost of form-factor/distribution), it might not always be clear what to optimize for in the trade-off. This is assuming that you still cross a certain bare-minimum resolution and field view of data that it is meaningful, something which arguably commercial EEG has crossed in the last 5 years.