BrainGate Neural Cursor (Hochberg et al., 2006)
Measured by Hochberg, Serruya, Friehs, Donoghue et al. · Nature 442 (2006)
Inputs
The measured or assumed values behind the calculations, each with its source.
- N = 16
- Grid task with 16 square targets, described in the Nature supplementary information.
- grid = 4x4
- The 16 square targets are treated as a 4x4 square grid for the same Fitts grid correction used on Webgrid and BrainGate2 grid tasks.
- P = 0.578
- MN selected among 16 targets with 57.8 +/- 25.9% accuracy (500 ms dwell; control 5.56 +/- 7.26%, n=9, p<0.0001), per the Nature 2006 Supplementary Information, 'Grid task' section (not the main text). The coarser 64-square grid dropped to 16.7 +/- 13.2%.
- t = 5 s/trial
- Each grid-task trial expired after 5 seconds. Mean acquisition time was not reported, so this uses the timeout as a conservative timing denominator.
- array = 96
- 96-microelectrode intracortical array implanted in primary motor cortex.
Strictest ITR
Each scoring method is an upper bound on the channel, so the headline is the strictest (smallest) one for this entry. Use the score selector on the home page to view any single method across entries.
-
Grid geometry to movement difficulty
4x4 board: cell width W = S/4; random grid targets -> mean amplitude A ~= 0.52 S, so A/W ~= 2.1
Same grid correction used for Neuralink Webgrid and the BrainGate2 cursor grid entry. The target is cued, so the movement carries the Fitts index of difficulty, not log2(16).
-
Information per movement
ID = log2(A/W + 1) = log2(2.1 + 1) ~= 1.62 bits/movement
The public text does not publish effective target width or endpoint scatter, so this uses nominal grid-cell width.
-
Net movement rate from timeout and accuracy
net movements/s = 0.578 / 5 = 0.116/s
The paper reports a 5 s trial timeout rather than mean acquisition time. Using the timeout makes this a lower-bound-style estimate.
-
Fitts throughput
1.62 bits/movement × 0.116 movements/s = 0.188 bits/s
What counts as a bit depends on the action space. The number of distinguishable actions and how likely each one is are design choices of the task, not the sensing hardware. The same modality can present a fixed set of targets, a set pruned per step by a grammar or language model, or a continuous control space. Each of these changes how many actions are live and how the probability mass is spread, and therefore the information per selection. Read the action space below before comparing headline numbers across entries.
Action space
What the user can produce at each step, and how those options are distributed.
- Structure
- Continuous control space
- Size
- Continuous
- Prior
- Uniform: all actions assumed equally likely
- Notes
- The first widely cited BrainGate human demonstration: a 2D neural cursor used for center-out and grid tasks, simulated email, simple games, television control, prosthetic hand opening/closing and rudimentary robotic-arm control. Because this is a continuous cursor channel, the reference calculation uses a Fitts-style grid correction on the 16-target task. The paper does not publish pixel geometry, endpoint scatter or mean acquisition time for the grid task, so this is an estimated lower-bound-style historical benchmark using the 4x4 grid layout, reported accuracy and 5 s trial timeout. The Wolpaw-over-16 figure is kept as supplementary.
Comparability The strictest bound here is Fitts throughput: the index of difficulty, log₂(A/W + 1), per movement. Directly comparable to the other continuous-pointing entries (mouse, trackball, stylus, gaze and the cursor BCIs). Set against the text entries (keyboards, spellers, speech) it crosses methods: both report bits/s, but one measures movement difficulty and the other text information, so compare within the family first.
Other score types
Bounds the atlas keeps out of the default strictest headline: as-reported figures, alternate task conditions, or raw-channel ceilings that shouldn't win the headline by default. Each still carries a score type, so the home-page selector ranks this entry on it when you choose that type. Read its derivation before comparing across entries.
-
Per-selection information with errors
B = log2(16) + 0.578 log2(0.578) + 0.422 log2(0.422 / 15) = 1.37 bits/selection
This uses the standard Wolpaw discrete-selection formula on the 16-target grid result. The task was cued cursor control, so this is best read as a historical benchmark, not text communication.
-
Selection rate from timeout denominator
60 / 5 = 12 trials/min
The public supplement reports a 5 s trial timeout, not mean acquisition time. Using the timeout for every trial makes the throughput a lower-bound style estimate.
-
Information transfer rate
1.37 bits/sel × 12 sel/min ÷ 60 s/min = 0.274 bits/s
-
Achieved-bitrate credit per correct acquisition
N = 16 cued targets → log2(N − 1) = log2(15) = 3.91 bits per net-correct selection (field-standard achieved bitrate, e.g. Webgrid; Nuyujukian 2015, which introduced the metric, used log2(N)).
-
Net-correct acquisition rate
incorrect = 0, so net-correct = 0.578 of trials at one per 5 s → 0.578 / 5 = 0.116 correct/s.
A missed target is a 5 s timeout marked incorrect, not a wrong-target selection (Supplementary Information: 'had 5 seconds to acquire the target, or the trial was marked as incorrect'), so there are no false selections to net against. Same timeout denominator (5 s) and 57.8% accuracy as the entry's Fitts and Wolpaw calcs. Because misses are timeouts rather than false selections, the achieved bitrate credits log2(N − 1) per successful acquisition; the Wolpaw figure instead spreads the 42.2% error mass over the 15 non-targets, so it runs lower here.
-
Achieved bitrate
3.91 bits × 0.116 correct/s = 0.45 bits/s.
Source
- Authors
- Hochberg, Serruya, Friehs, Donoghue et al.
- Publication
- Nature 442, 2006
- Paper
- 10.1038/nature04970
- Reference
- Nature supplementary information
- Reference
- BrainGate publication page