Summary of "DAY 3 (Part 1) The BCI & Neurotechnology Spring School 2026"

Main ideas & concepts (what the video teaches)

• BCI / neurotechnology spring school overview

  • The DAY 3 (Part 1) session covers the pipeline from signals → meaning, including:
    • EEG-based feature extraction
    • band power
    • event-related potentials (ERPs)
    • artifact handling
  • The event is global and large-scale, framed as the “Woodstock of neurotechnology,” emphasizing shared learning across many countries and institutions.

• How EEG signal processing is done in practice (Johannes Kunvald + Sebastian Seagal Lightner)

  • A practical real-time EEG processing pipeline is demonstrated using g.tec gpipe (a Python SDK for real-time signal processing and BCI).
  • The talk moves from preprocessing basics toward feature extraction and ERP/event analysis.

• Key EEG feature extraction methods

  • Band power extraction (e.g., alpha power) in real time:
    • Band-pass filtering to isolate a frequency range
    • Squaring to convert to power-like measures
    • Moving average to average over a short window
    • Log transform to stabilize statistical behavior (helpful for classification assumptions)
    • Normalization using broadband power so power changes reflect meaningful attention/oscillations rather than artifacts or non-stationarity
  • Evoked/event-related potentials (ERPs)
    • A stimulus presentation/paradigm generates triggers
    • Use trigger nodes aligned to stimulus events (soft triggers)
    • Extract and average trial-locked ERP waveforms; discuss what happens when trials contain artifacts
  • Artifact handling
    • Artifacts can’t simply be “ignored”; they must be managed through:
      • prevention (setup, careful behavior)
      • trial rejection (for epoched data)
      • signal correction/removal (for continuous data)
    • The demo introduces OSCAR, an in-real-time artifact removal tool integrated into gpipe with ~250 ms delay

• Historical evolution of “feature extraction → decoding” in BCIs (R. “Randy” Sher)

  • Core framing:
    • BCIs decode intent from brain measurements
    • performance is constrained not just by classifiers, but heavily by the features/representations chosen
  • Historical phases:
    1. Synchrony-based EEG features (e.g., band power like alpha; ERP averaging)
    2. Better spatial information (e.g., ECoG features like high-gamma and spatial patterns)
    3. More invasive recordings (spiking, multi-unit, LFP) reveal non-stationarity and “stability crises”
    4. Shift toward latent state models: neural activity treated as observations generated from an underlying low-dimensional latent state
  • Main message:
    • “Better classifiers won’t fix a wrong question.” Representation and stability matter.

• Auditory neurotechnology signal processing & selective auditory attention (Adrian Mai)

  • Auditory attention decoding can use:
    • evoked/ERP activity
    • ongoing EEG (e.g., regression-based speech-envelope tracking)
    • EMG around the ears (post-auricular muscle activity can carry attention information)
  • Auditory evoked potentials:
    • ABR (auditory brainstem response): early waves
    • middle latency responses
    • late cortical responses
  • Improving ABR SNR:
    • stimulus design using chirps
    • PCA for broadband ABR extraction
    • wavelet transforms for time-frequency representations
    • ITPC (inter-trial phase coherence) to distinguish:
      • attention changing power vs attention changing phase consistency
  • Extracting auditory evoked potentials from ongoing speech:
    • use acoustic edge triggers

• Interpretable deep-learning for EEG feature extraction & classification (JP / Alex “Osachi”)

  • Focus on interpretable neural decoding using:
    • physiologically meaningful spatial and temporal filtering
    • learned weights to interpret which brain regions/dynamics contribute
  • Key point:
    • avoid shortcut learning (e.g., models using artifacts/muscle activity instead of brain activity)
    • separate “measurement physics” from “neural/physiology dynamics” via architecture and interpretability methods

• Practical future/ethics & clinical neuromodulation context

  • Neuroethics for industry-academia partnerships (Tristan Macintosh)
    • Why neuroethics matters: trust, safety, accountability, regulation readiness
    • Common partnership risks:
      • competing incentives/timelines
      • IP ownership disputes
      • presumed trust
      • inadequate post-trial access/support
      • communication failures/expectation gaps
    • Proposed response strategy: three-part approach (intrapersonal, interpersonal, operational)
  • Clinical epilepsy neuromodulation biomarkers (Jonathan Parker / JP)
    • Challenge in DBS programming:
      • seizure freedom isn’t immediate
      • clinicians lack fast biomarkers
      • seizure counts alone are insufficient due to rhythms and long-term effects
    • Goal: early EEG biomarkers to guide settings
  • Inner speech decoding (Erin Kun)
    • Inner speech as a more comfortable alternative to attempted speech or miming
    • System uses Utah arrays to decode phone probabilities, then a language model generates text
    • Discusses representational differences across:
      • attempted speech vs imagined (inner) speech vs listening
    • Safeguards against decoding unintended inner thoughts
    • Demonstrates decoding speed improvements and autonomy-preserving strategies (e.g., training-time “silence,” keyword gating)

• Auditory attention “hearing aid” style closed-loop concept (Nema Mascarani)

  • Decodes what the user intends to attend to, then remixes audio in real time
  • Emphasizes behavioral usability beyond accuracy:
    • switching attention capability
    • subjective preference
    • reduced listening effort
    • safety controls (e.g., turning the system off can frustrate users if it suppresses the wrong talker)
  • Envisions integrating attention-aware signals into foundation models for scene understanding

• High-performance BCI encoding for practical noninvasive use (Chingqing / Nadia / Nema)

  • Themes include:
    • robustness to weak/noisy signals
    • fatigue reduction
    • multimodal paradigms (visual, auditory, tactile)
    • generalization/personalization (small calibration, transfer learning, meta-learning)
    • scaling sensor arrays (e.g., “ultra-high density EEG”)

---

Detailed methodology / instructions presented

A) Real-time EEG preprocessing pipeline (g.tec gpipe demo)

Setup

• Use a real-time pipeline including:
  • EEG amplifier node (example: PCI Core 8)
  • visualization scope node (time-series scope)

Step 1: High-pass filter

• Add a 1 Hz high-pass filter to remove DC offset from DC-coupled amplifiers.

Step 2: Notch filter for line noise

• Add a notch (band-stop) filter for power-line interference:
  • center around 50 Hz with a narrow stop band of ±2 Hz margin
  • note: in the US/Canada, this would be 60 Hz

Step 3: Low-pass filter

• Add a low-pass filter for standard EEG viewing/analysis:
  • 1–30 Hz (noting the range can be debated, but used here for dry-electrode demos)

Outcome

• Improves readability by removing offset, reducing line noise, and keeping artifact/blink visibility manageable for demonstration.

---

B) Real-time band power extraction (example: alpha power)

Preprocessing chain

• Band-pass to target range:
  • example: 1–30 Hz broadband reference
  • then isolate the alpha band (typically 8–12 Hz)

Compute power from band-limited EEG

• Narrow to the desired band (alpha, beta, etc.)
• Square the band signal (power-like transformation) using an equation node:
  • ( x \rightarrow x^2 )

Temporal averaging

• Apply a moving average over a short window:
  • example: 0.5 s
  • with sampling rate 250 Hz → 0.5 s = 125 samples

Stabilize distribution

• Apply log transform to stabilize variance/uncertainty:
  • squared + averaged power values are non-stationary
  • log makes them more comparable across levels

Normalization (important caveat)

• Normalize alpha power using broadband power to remove confounds/artifacts:
  • broadband pipeline: square → moving average → log
  • normalized alpha:
    • in the log domain, subtract broadband power from alpha power
    • (equivalent to division in linear domain)
• Goal: preserve meaningful modulation while reducing artifact-driven changes

Visualization / validation

• Use multiple scopes to inspect:
  • raw EEG
  • band-passed EEG (alpha)
  • squared alpha
  • averaged alpha power
  • log-transformed alpha power
  • normalized alpha power
• Demonstrate modulation by asking the subject to open/close eyes.

---

C) Evoked potentials / ERPs extraction with triggered epochs (visual attention task)

Stimulus presentation

• Use a paradigm presenter to display stimuli, e.g.:
  • checkerboards with different spatial frequencies
  • faces

Triggers

• Use soft triggers emitted by the paradigm presenter that include:
  • timing + target value
• Use trigger nodes:
  • one per stimulus class
• Trigger nodes feed triggered signals into a trigger scope for ERP visualization

ERP computation / trial handling principles

• Early trials may contain movement artifacts.
• Without artifact rejection:
  • contaminated trials show unusually high amplitudes
  • later trials may converge once artifacts diminish

ERP interpretation guidance

• Compare ERP shapes across electrodes (e.g., central vs occipital channels).
• Interpret using differences between attended stimulus types (e.g., face-evoked ERP vs checkerboard).

---

D) Artifact handling (OSCAR integrated into gpipe)

Conceptual rules

• Prefer avoiding artifacts at the source.
• If needed:
  • for epoched data: trial rejection
  • for continuous data: correction/removal

OSCAR usage

• Enable OSCAR via a flag in the gpipe pipeline (example):
  • enableOscar=true/false
• OSCAR typically requires a short stable period (“a couple of seconds of good EEG”) before effective correction begins
• After enabling, demonstrate reduced artifact effects for:
  • blinks
  • head movements
  • electrode touching
  • etc.

Deployment

• OSCAR is described as rolled out across multiple systems (including PCI cores, Nelus, hybrid black) and integrated into gpipe/GIS architecture.

---

E) Auditory attention decoding approaches (Adrian Mai / Nema Mascarani)

Evoked potential approach

• Use unpredictable auditory oddball sequences (standards vs deviants, random interstimulus intervals)
• Attend to one ear/source; ignore the other
• Extract attention markers such as N1 enhancement from auditory evoked potentials

Ongoing EEG / speech regression approach

• Model EEG as convolution of stimulus features (e.g., speech envelope) with a TRF
• Use stimulus reconstruction as an inverse approach
• Compare attended vs ignored reconstruction/TRFs

Single-trial mechanisms via time-frequency analysis

• Use wavelets for time-frequency power/phase analysis
• Use ITPC to determine whether attention changes:
  • power, or
  • phase consistency (jitter reduction)

Speech edge-event segmentation

• Compute the derivative of the speech envelope
• Detect edges via threshold crossing
• Segment EEG around these acoustic edge triggers to extract evoked responses from ongoing speech

---

F) Neuroethics strategy for industry-academia partnerships (Tristan Macintosh)

Problem categories identified

• competing values/incentives
• timeline mismatches
• IP ownership disputes
• presumed trust
• post-trial access/support expectations
• expectation-setting and communication failures

Response framework (3 domains)

1. Intrapersonal (within individuals/teams)

   • test assumptions
   • run internal audits
   • pause and dialogue about tradeoffs

2. Interpersonal (stakeholder-to-stakeholder)

   • perspective taking
   • explicit communication and relationship-building
   • transparency
   • meaningful engagement of patients/clinicians early

3. Operational (process and governance)

   • explicit policies and procedures
   • clarity in contracts (timeframes, IP, support obligations)
   • planned responsibility assignment for long-term device support

---

G) Inner speech BCI safeguards / autonomy (Erin Kun)

Key concept

• Differentiate:
  • attempted speech
  • inner speech
  • listening

Safeguard strategies described

• For attempted speech BCIs (avoid decoding inner speech):
  • include imagined trials labeled as “silence”
• For inner speech BCIs:
  • use keyword detection / gating so decoding starts only when the user intends it

Training targets

• Use language models driven by RNN-decoded phone probabilities

Autonomy consideration

• Monitor and reduce decoding of unintended private inner utterances.

---

Speakers / sources featured (as named in subtitles)

1. Christoff (host/introducer; appears multiple times)
2. Johannes Kunvald
3. Sebastian Seagal Lightner
4. Reinhold “Randy” Sher
5. Adrian Mai
6. Alex Osachi
7. Nadia Lamon
8. Nema Mascarani
9. Erin Kun
10. Tristan Macintosh
11. JP / Jonathan Parker
12. Chingqing

Category

Educational

---

Share this summary

Share on X/Twitter Share on Facebook Share on LinkedIn Share on Reddit Share on WhatsApp Share on Telegram

---

Is the summary off?

If you think the summary is inaccurate, you can reprocess it with the latest model.

Reprocess summary

Video