↑↓ move↵ open[] previous / next lesson? all shortcuts
Appearance
Theme
Type size
Reading

Focus mode hides the sidebars. Press f to toggle it anywhere.

Keyboard shortcuts
/
Search
[ ]
Previous / next lesson
t
Jump to the table of contents
r
Review queue
f
Focus mode
?
This panel
ProjectA month$0Tier 0

Project E: Intracortical data

Human motor cortex recordings from published BCI studies, downloaded free. Spike sorting, local field potentials, and a decoder on the real thing. The end of the road that started with a cockroach leg.

AssumesProject E: Large-scale neural dataBayesian decoding and the Kalman filterSpineDecoding / signal processing / MLComputational neuroscience
On this pageA month
  1. The data
  2. Three analyses
  3. Speech
  4. What you will understand at the end

The and the data supplements of recent Nature papers hold recordings from in the motor cortex of people with paralysis: the actual data behind the cursor and speech-decoding results. In format, with the behaviour aligned. You have built every tool needed to work with them: detection from Phase 1’s thinking, the Kalman from Phase 3, the population analysis from Phase 4. Now run them on a human brain.

Predict before you look

A Utah Array has 96 recording electrodes. In a participant implanted three years ago, roughly how many electrodes still show clearly sortable single neurons?

Often a dozen or two clean single units, with many more channels carrying threshold crossings from several unsortable neurons and some carrying only field potentials. Recordings degrade over years as the tissue responds and fail, which is why most human now use threshold crossings and -band power rather than sorted units, and why the foreign body response is the field’s central engineering problem.

The data

Search DANDIDANDI, Distributed Archives for Neurophysiology Data IntegrationDANDI Consortium · DANDI Archive · 2019Neurophysiology in NWB format, including the human intracortical data behind the speech BCI papers. for human intracortical datasets: the BrainGate consortium has released cursor-control and handwriting data; the Willett et al. 2023A high-performance speech neuroprosthesisWillett, Kunz, Fan, et al. · Nature · 2023Speech decoded from motor cortex at conversational rates, with data public for your own reimplementation. speech dataset and the Metzger et al. 2023A high-performance neuroprosthesis for speech decoding and avatar controlMetzger, Littlejohn, Silva, et al. · Nature · 2023The second 2023 speech BCI, decoding speech and facial movement together, with public data. dataset are public through Dryad and the papers’ supplements. The Neural Latents BenchmarkThe Neural Latents BenchmarkPei, Deng, O'Shea, et al. · NeurIPS · 2021Standard monkey datasets with held-out evaluation, the right place to compare a decoder against the field. packages several (mostly monkey) datasets in a standard format with held-out evaluation, which is the right place to compare a against the field. Load with pynwb or the NLB tools.

Three analyses

Figure 1. Three analyses on human motor cortex, in order. The first is the one nobody publishes and everybody has to do, and it is what makes the third believable.

Signal quality across channels. For each : the -band (300 to 3000 Hz) RMS, the threshold-crossing rate, and, where waveforms are provided, a sorting. Plot the array as a 10-by-10 grid coloured by signal quality. Count the clean single units. Compare to the paper’s methods section. This grid is the honest picture of what a chronic array delivers.

Decode kinematics. On a cursor or handwriting dataset: bin threshold crossings at 20 ms, fit the from Phase 3 on the block, test on the held-out block, report correlation between decoded and true velocity. Then fit the same on sorted units only and compare; then on -band power; then on features. The comparison says which signal carries the information, and it usually surprises people how well the unsorted signals do.

Population dynamics. PCA on the binned population activity during reaches. Plot the trajectories per reach direction. See the rotational structure that the neural dynamics literature describes, in a human motor cortex. Then try a latent-variable model (the NLB baselines include several) and report held-out likelihood.

Speech

If you take on the speech dataset: neural features per 20 ms, attempted-speech phoneme labels, a recurrent network mapping one to the other, a language model on top. Weeks, not days, and the result is a reimplementation of a paper that made headlines. Report phoneme error rate and word error rate alongside the paper’s.

What you will understand at the end

What a human actually looks like after years, and why were redesigned to cope. Why threshold crossings won. How much of the population’s information lives in a few dimensions. And that the arc from a cockroach leg on a speaker to a paralyzed person’s motor cortex on your laptop is one arc, made of the same physics, the same amplifiers, and the same statistics.

  1. Choose a dataset and load it. Plot raw traces from a good channel and a bad one.
  2. Build the array quality grid. Count units.
  3. Kalman on threshold crossings, held-out block; then on sorted units, -band power, and . Table.
  4. PCA trajectories by reach direction.
  5. Optionally, the speech .
  6. Write it up as a reproduction, in the Phase 3 format.
Recall
Why do most human BCI decoders now use threshold crossings rather than sorted units?
Chronic arrays lose sortable single units over years; threshold crossings and spike-band power remain informative on many more channels and are more stable, and they decode as well or better.
Recall
What does the array-quality grid show and why is it important?
Per-electrode signal quality across the 10-by-10 array, revealing how many channels carry clean units, multi-unit activity, or only LFP; it is the honest picture of chronic recording performance that motivates materials research.