Research log 000

Signal acquired

What we are building, the brain signal it depends on, what published research already shows, and what we have not done yet.

MIND LabsNovember 18, 20269 min read

The idea in one paragraph

MIND Labs is building a trading terminal you operate by looking at it. Each action on screen (buy, sell, next market, confirm) pulses at its own rate. When you rest your eyes on one of them, the visual part of your brain starts to echo that rate. A headset on the back of the head picks up the echo, software works out which rate it is, and that tells the terminal which action you chose.

None of the neuroscience is ours. The effect has been studied for decades and has a name: the steady-state visual evoked potential, or SSVEP. What is new is pointing it at onchain markets, on consumer hardware, in a form anyone can try.

The signal

If a light or a patch of screen flickers steadily, neurons in the visual cortex fire in step with it. Measured with electrodes on the scalp (EEG), this shows up as extra energy at exactly the flicker frequency and its multiples. Look at a 12 Hz target and a 12 Hz component appears over the back of your head; shift your gaze to a 10 Hz target and it moves to 10 Hz.

Three properties make SSVEP attractive for an interface:

  • It needs little or no training. The response is automatic. You do not learn to produce it.
  • It is non-invasive. It is read from outside the skull. Nothing is implanted.
  • It carries a lot of information for an EEG signal. Many targets can be told apart at once, and a selection can be made in a second or two.

What published research shows

The strongest results come from laboratory systems with many electrodes, conductive gel and careful calibration. They set the ceiling, not our expectation.

  • In 2015, Chen and colleagues reported a 40-target SSVEP speller that reached about 60 characters per minute, an information rate of 5.32 bits per second, the fastest brain speller reported at that time by any method [1].
  • Open datasets exist for testing decoders without recruiting anyone. The Tsinghua benchmark contains 64-channel recordings from 35 people selecting among 40 targets [2]. BETA extends it to 70 people, recorded outside a shielded room, with target frequencies from 8 to 15.8 Hz [3]. A further dataset was recorded with a wearable headband, which is closer to our hardware [4].
  • The standard training-free way to identify the target is canonical correlation analysis (CCA), introduced for SSVEP by Lin and colleagues in 2006 [5]. It compares the recorded signal with sine waves at each candidate frequency and picks the best match.

Our approach

ChoiceWhat we pickedWhy
TargetsFour: buy, sell, next, confirmFew targets are far easier to separate than forty, which matters on low-cost hardware.
Frequencies8.57, 10, 12 and 15 HzEach divides evenly into a 60 Hz screen refresh (7, 6, 5 and 4 frames), so the flicker is exact.
DecoderCCA firstIt works without per-person calibration. Trained methods come later if they earn their complexity.
HardwareConsumer dry-electrode EEG, starting with the BrainBit MINDOThe signal is strongest there, and the product only matters if people can own the hardware.
Safety stepTwo looks per orderAn action is only sent after a separate confirm target is also selected.

What is hard

  • Dry electrodes are noisy. The published records use gel and dozens of channels. A headband has a handful of dry contacts sitting on hair. Accuracy will be lower and selections slower.
  • People differ. SSVEP strength varies a lot between individuals, and for some people it is weak at some frequencies.
  • Looking is not choosing. Eyes land on things without intent. A decoder that fires on every glance is unusable for anything involving money. This is why the confirm step exists, and why we will report false-activation rates alongside accuracy.
  • Markets move in the time it takes to decide. A two-second selection plus a two-second confirm is slow next to a mouse click. We are not claiming speed. We are finding out what this interface is good for.

Safety

Flicker can trigger seizures in people with photosensitive epilepsy. A review for the Epilepsy Foundation estimates that seizures from light stimuli affect roughly 1 in 10,000 people, and about 1 in 4,000 between the ages of 5 and 24. It finds 15 to 25 Hz the most provocative band, with risk across 1 to 65 Hz [6].

Our target frequencies sit inside that wider range, and the fastest one touches the most provocative band. We treat this as a design constraint, not a footnote. The working terminal will show a warning before any flicker starts, will not autoplay flicker, and will keep targets small and low in contrast. We are also evaluating stimulus designs that reduce visible flicker. If you have epilepsy or a history of seizures, do not use flicker-based interfaces.

MIND is not a medical device. It does not diagnose, treat or monitor any condition.

Where we are today

We want to be exact about this.

  • We have a design for the terminal and the decoding pipeline described above.
  • We have a working terminal interface with live market data. It is in development and is operated by pointer or keyboard, because the headset link is not built yet.
  • We have not yet recorded a live signal from our own headset, and we have not executed a trade from a brain signal. Anything that suggests otherwise is wrong.

What comes next

  • Log 001, decoder benchmark. Run our pipeline on the open datasets above, restricted to a few electrodes over the back of the head, and publish accuracy, selection time and false activations.
  • Log 002, first live signal. Record a frequency lock from a consumer headset in our own setup and publish the raw data.
  • Log 003, first brain-executed trade. One look, one confirm, one transaction, recorded end to end.

Each log will include what failed as well as what worked.

References

  1. Chen X, Wang Y, Nakanishi M, Gao X, Jung T-P, Gao S. High-speed spelling with a noninvasive brain–computer interface. PNAS 112(44), 2015. PMC4640776
  2. Wang Y, Chen X, Gao X, Gao S. A benchmark dataset for SSVEP-based brain–computer interfaces. IEEE Trans. Neural Syst. Rehabil. Eng. 25, 2017. doi:10.1109/TNSRE.2016.2627556
  3. Liu B, Huang X, Wang Y, Chen X, Gao X. BETA: a large benchmark database toward SSVEP-BCI application. Front. Neurosci. 14:627, 2020. doi:10.3389/fnins.2020.00627
  4. Zhu F, Jiang L, Dong G, Gao X, Wang Y. An open dataset for wearable SSVEP-based brain–computer interfaces. Sensors 21, 1256, 2021.
  5. Lin Z, Zhang C, Wu W, Gao X. Frequency recognition based on canonical correlation analysis for SSVEP-based BCIs. IEEE Trans. Biomed. Eng. 53, 2006. doi:10.1109/TBME.2006.886577
  6. Fisher RS, Harding G, Erba G, Barkley GL, Wilkins A. Photic- and pattern-induced seizures: a review for the Epilepsy Foundation of America Working Group. Epilepsia 46(9), 2005. doi:10.1111/j.1528-1167.2005.31405.x

Open the terminal →