Professor Paul Nuyujukian MD, PhD (Brain-machine interface researcher) explains the Neuralink Monkey MindPong video and its significance for the fields of neuroengineering and systems neuroscience.

Paul Nuyujukian MD, PhD
Assistant Professor
Departments of Bioengineering, Neurosurgery, & Electrical Engineering
Director of Brain Interfacing Laboratory: http://bil.stanford.edu
Stanford University

Chapters:
00:00 – Introduction
00:37 – Disclosures
02:22 – Pager Introduction
05:06 – Bluetooth Interface
06:24 – Electrode Count
07:13 – Grid Task & Bitrate
08:30 – Electrode Tuning & Neuroanatomy
10:49 – Building a Decoder
11:18 – Neuralink Studio Console
11:56 – Console Status Pane
14:33 – Neuroelectrophysiology Stream
15:56 – Interactive Console Components
19:00 – 2D Neural Decoding
20:09 – Decoding Algorithm Observations
22:03 – Human Application Decoder Calibration
23:00 – MindPong
24:28 – Banana
24:51 – Closing Remarks

Corrections:
1) At 17:04 about Bluetooth protocol used: N1 Link uses “latest BLE” (Bluetooth Low Energy) as mentioned by one of the Neuralink engineers: https://twitter.com/djseo8/status/1300911591743827968
2) At 21:46 I misspoke saying “This is definitely running an HMM”, when I meant to say “This is definitely running a click decoder”. There’s no way for me to know based on the video what type of click decoder is running and I just slipped calling it an HMM because that is the most common decoder I have used in my work when implementing click functionality.

Additional thoughts:
I should have said a few words about the choice of 25ms bins. It’s a very smart window, as it’s plenty fast enough for nearly all neural prosthesis applications and is a pretty ideal point that balances the update rate with bandwidth minimization goals. The maximum practical update rate is 1ms (minimum duration of an action potential), which is completely overkill. Many papers in the field run at 25ms or 50ms bin widths, and Neuralink’s choice of 25ms bin widths while drawing from 250ms of history on the computer for each velocity decode is entirely sensible.
I’m sure they’re doing tight bit-packing to maximize efficiency here before transmitting over the radio. Neurons don’t fire much above a couple hundred Hertz. They are likely packing 2-3 bits per channel per 25ms bin which can encode 160Hz or 320Hz max firing rates, respectively. Even at the high end of 3 bits per channel, 25ms bins (40 bins/sec) for 1024 channels without compression comes out to a data rate of 3 bits/channel/bin * 1024 channels * 40 bins/sec / 1024 bits/Kb = 120Kb/sec. 2 bits per channel puts them at 80Kb/sec. Either choice is a totally reasonable data rate that’s comfortably below the BLE max rates. Neuralink engineers clearly did the math and made smart choices.

Comments

Comments are disabled for this post.