How much can we write?
The write channel is far lower-bandwidth than the read channel. Weber's law caps each electrode at a handful of distinguishable intensity levels, and neighbouring electrodes evoke overlapping percepts, so the number of jointly distinguishable states grows much more slowly than the raw channel count suggests.
C_{\mathrm{write}} \;\lesssim\; \sum_{j=1}^{M} \log_{2}\!\big(1 + L_{j}\big) \quad \text{bits per update}A loose ceiling: with M effectively independent electrodes each supporting L_j discriminable levels, the writable information per update is bounded by the sum of \log_2(1+L_j) — and channel overlap makes the real figure much smaller.
How much information you can push into the brain per update is capped by adding up, across all your electrodes, the bits each one can carry. An electrode with L distinguishable levels carries \log_2(1+L) bits — and because real channels overlap, the honest figure is much smaller than this sum.
- C_{\mathrm{write}}
- Writable information per update, in bits.
- M
- Number of effectively independent electrodes.
- L_j
- Number of distinguishable levels electrode j supports.
- \log_2(1 + L_j)
- The bits one electrode contributes with L_j levels.
An electrode with four distinguishable levels contributes \log_2(5) \approx 2.3 bits per update.
Biomimicry vs. learned codes
The field's central open debate: should encoders mimic the natural afferent code as closely as possible, or should they use whatever pattern is stable and controllable and let cortical plasticity learn to read it? Biomimicry promises immediate naturalness and intuitive use; learned codes exploit the brain's own adaptability and may pack in more information than nature's format allows. The honest current answer is that both matter and the balance is unsettled — biomimetic transients give a strong head start, but users also demonstrably learn arbitrary mappings over time.
Scaling: more channels and current steering
Richer feedback wants more, finer contacts. Rather than driving electrodes independently, current steering splits current across neighbours to place a virtual source between physical contacts, sculpting the activated region with sub-electrode resolution. But scaling collides with physics: every added channel adds charge, and both the charge-density limit and aggregate tissue heating cap how much you can inject across a dense array at once.
Beyond somatosensation, and the stability problem
The same write machinery is being pushed toward other senses. A cortical visual prosthesis stimulates visual cortex to evoke phosphenes — spots of light — and stitches them into crude images. It is harder than touch: percepts interact non-linearly, dynamic current steering helps, and assembling coherent form from a grid of phosphenes is an unsolved perceptual problem.
Every frontier runs into stability. On the write side, percepts fade and thresholds drift as glial encapsulation raises impedance around chronic electrodes. On the read side, the decoded representation moves too, so manifold stability and recalibration matter as much for bidirectional systems as for read-only ones. The overarching open problems are clear: write more information, keep percepts natural and stable for years, and prove safety as channel counts climb into the thousands.