Neurotechnology & Human-Machine Interfaces¶
-
Exit capability
Decode, model, stimulate, and close loops around nervous-system and physiological signals while treating adaptation, safety, identity, and user utility as first-class.
-
Frontier targets
BCIs; neural prostheses; sensory augmentation; speech restoration; adaptive exoskeletons; memory interfaces.
-
Choose this track if
Best if you want computation and machinery to become extensions of human capability.
Neural dynamics¶
Week 1: Biophysics of neurons¶
Reading/source: Neuronal Dynamics Ch. 1-2 + physiology membrane concepts
Know: Own membrane RC analogy, ionic currents, resting potential and spike-generation models.
Reconstruct: Derive leaky integrate-and-fire equation and time constant.
Do: Simulate current-to-spike response; compare LIF with richer nonlinear/spiking model.
Context: Neuronal Dynamics provides online theory plus Python exercises.
Defend: What biological features does LIF deliberately discard?
Gate: Pass if simplification is tied to intended decoding/control task.
Source: source
Week 2: Hodgkin-Huxley and excitability¶
Reading/source: Neuronal Dynamics Ch. 2-4
Know: Understand conductance-based dynamics, threshold, refractory behavior and adaptation.
Reconstruct: Reconstruct current-balance equation and gating-variable concept.
Do: Simulate HH-like model under current steps; map excitability regimes.
Context: Mechanistic models are useful for stimulation and interpretation even when decoders are learned.
Defend: Which parameters are physiologically identifiable from extracellular signals?
Gate: Pass if model/measurement observability is discussed.
Source: source
Coding¶
Week 3: Spike trains and neural coding¶
Reading/source: Neuronal Dynamics coding chapters
Know: Represent spike timing/rates, tuning curves and variability.
Reconstruct: Derive firing-rate estimate and Poisson likelihood/basic information intuition.
Do: Generate synthetic population code and decode stimulus from rate vs timing features.
Context: BCI design depends on what information neural activity carries and how stable it is.
Defend: What information did binning destroy?
Gate: Pass if decoder comparison uses same latency/data budget.
Source: source
Week 4: Population models and GLMs¶
Reading/source: Neuronal Dynamics GLM/population material
Know: Fit probabilistic encoding models and quantify predictive uncertainty.
Reconstruct: Derive Poisson-GLM log likelihood/gradient intuition.
Do: Fit GLM to synthetic spikes; inspect residuals and misspecification.
Context: Modern neural decoding often combines statistical models and deep learning.
Defend: Does predictive accuracy imply mechanistic truth?
Gate: Pass if residual structure is checked.
Source: source
Decoding¶
Week 5: State estimation from neural signals¶
Reading/source: Wave-1 estimation + Neuronal Dynamics
Know: Decode continuous/discrete intention while propagating uncertainty.
Reconstruct: Derive linear/Kalman decoder and Bayesian classification formulation.
Do: Build cursor/kinematic decoder from synthetic neural population data; add drift.
Context: BCIs are estimators embedded in feedback loops.
Defend: What state is user intending versus what state your labels assume?
Gate: Pass if calibration/uncertainty is reported, not accuracy only.
Source: source
Week 6: Adaptive decoders and nonstationarity¶
Reading/source: Neuronal plasticity + online learning
Know: Handle electrode/unit drift, learning, fatigue and nonstationary mappings.
Reconstruct: Derive recursive/online parameter update and forgetting-factor tradeoff.
Do: Simulate decoder drift; compare static, periodically recalibrated and adaptive models.
Context: Long-term practical BCIs must work outside one lab session.
Defend: When does adaptation chase noise or override user learning?
Gate: Pass if adaptation improves long-horizon utility without catastrophic instability.
Source: source
Interfaces¶
Week 7: Signal acquisition and artifacts¶
Reading/source: Wave-2 electronics/signals + neurophysiology
Know: Understand invasive/non-invasive signal bandwidth, noise, referencing and artifacts conceptually.
Reconstruct: Build full signal chain/noise budget from neural source to ADC/decoder.
Do: Simulate EMG/ECoG/EEG-like signals with line noise/motion artifact; design filtering without erasing target signal.
Context: Measurement modality constrains achievable information before ML begins.
Defend: Which 'neural' feature could actually be artifact?
Gate: Pass if artifact control/negative control is designed.
Source: source
Week 8: Latency, bandwidth and closed-loop utility¶
Reading/source: Control/information + BCI context
Know: Treat communication rate, latency, error correction and user correction as system variables.
Reconstruct: Derive simple bits/minute or task-time utility metric and latency stability effect.
Do: Closed-loop cursor/typing simulator: vary decoder error vs latency vs correction ability.
Context: Clinical usefulness is not captured by offline classification accuracy.
Defend: Would you choose a slower, more accurate decoder?
Gate: Pass if metric maps to user task, not abstract model score.
Source: source
Human systems¶
Week 9: Motor control and sensorimotor loops¶
Reading/source: A&P motor systems + control theory
Know: Understand spinal/cortical motor pathways, proprioception and feedback/feedforward control.
Reconstruct: Draw nested human-device control loops and delays.
Do: Model powered-assistance/exoskeleton control with human adaptation.
Context: Human operator is part of the controller, not an external disturbance.
Defend: Who adapts to whom?
Gate: Pass if coupled adaptation is represented.
Source: source
Week 10: Sensory coding and augmentation¶
Reading/source: A&P sensory systems + neural coding
Know: Map stimulus transduction, receptive fields and perceptual adaptation to artificial sensory channels.
Reconstruct: Derive dynamic-range/compression mapping for an augmented sensor.
Do: Design simulation converting IR/ultrasound sensor values into haptic/auditory code; quantify learnability/information.
Context: Sensory augmentation requires a human-learning channel, not just more sensors.
Defend: What makes a code intuitive versus merely decodable after training?
Gate: Pass if user-learning burden is an explicit metric.
Source: source
Stimulation¶
Week 11: Neural stimulation as control input¶
Reading/source: Neurophysiology + safety literature conceptually
Know: Understand stimulation as uncertain input to excitable/adaptive tissue; distinguish activation from desired function.
Reconstruct: Build simplified stimulus-response and safety-envelope model.
Do: Simulate closed-loop stimulation controller with uncertain thresholds/adaptation; no operational clinical settings.
Context: Bidirectional interfaces require both decoding and controlled perturbation.
Defend: What unintended populations/effects can the stimulation recruit?
Gate: Pass if control uses conservative constraint independent of performance objective.
Source: source
Replication¶
Week 12: Speech/motor BCI system reconstruction¶
Reading/source: NIH 2026 home speech-BCI summary + underlying public paper if available
Know: Backchain a real practical BCI from implant/signal -> decoder -> synthesis -> home use.
Reconstruct: Reconstruct architecture, training, latency and evaluation metrics from sources.
Do: Build reduced speech/sequence-decoding analogue on public/synthetic data.
Context: NIH reported July 2026 home use of a speech BCI by a person with paralysis.
Defend: What changed when the system moved from lab to home?
Gate: Replication Gate: distinguish reported evidence from your extrapolation.
Source: source
Physiology¶
Week 13: Cardiovascular and respiratory control¶
Reading/source: OpenStax A&P cardiovascular/respiratory chapters
Know: Understand circulation/gas exchange and compensatory control as engineered-system analogues.
Reconstruct: Derive cardiac output, oxygen delivery and ventilation mass-balance relations.
Do: Build lumped cardio-respiratory model under exercise/altitude/device assistance.
Context: Neurotech often interacts with whole-body physiology.
Defend: Which compensation hides device failure until late?
Gate: Pass if multiple feedback loops interact.
Source: source
Week 14: Endocrine and autonomic regulation¶
Reading/source: A&P endocrine/autonomic sections
Know: Model slow hormonal/autonomic loops and multi-timescale homeostasis.
Reconstruct: Derive negative-feedback hormone toy model.
Do: Simulate fast neural + slow endocrine control of one physiological variable.
Context: Human augmentation cannot assume one control timescale.
Defend: How does delay produce overshoot/oscillation in physiology?
Gate: Pass if timescale separation guides controller design.
Source: source
Week 15: Renal/metabolic homeostasis and artificial organs¶
Reading/source: A&P renal/fluid/electrolyte chapters
Know: Understand clearance, fluid balance and biochemical regulation relevant to artificial-organ interfaces.
Reconstruct: Derive clearance/mass-balance model.
Do: Simulate dialysis/insulin-like control abstraction with sensor delay and safety bounds.
Context: Artificial organs are cyber-physiological systems.
Defend: What does replacing average function miss about homeostatic regulation?
Gate: Pass if disturbance recovery and fault state are tested.
Source: source
Evaluation¶
Week 16: Clinical/user-centered endpoints¶
Reading/source: NIH BCI reports + causal/statistical core
Know: Define user utility, communication independence, fatigue, training burden and durability endpoints.
Reconstruct: Write target-trial/longitudinal evaluation for assistive neurotechnology.
Do: Design synthetic longitudinal dataset and analyze dropout/adaptation/confounding.
Context: Practical utility must survive months, context shifts and user priorities.
Defend: Whose definition of success is in the objective?
Gate: Pass if endpoint set includes patient/user-reported and functional measures.
Source: source
Ethics/safety¶
Week 17: Agency, privacy and mental data¶
Reading/source: Security/humanities core + neurotechnology
Know: Treat neural/physiological data as high-stakes signals with consent, access, inference and misuse risks.
Reconstruct: Build threat model and authorization/recourse map.
Do: Red-team a BCI data pipeline for inference, insider, model-update and device-control abuse.
Context: Neural interfaces collapse boundaries between measurement and intervention.
Defend: What data/inference should never be required to use the device?
Gate: Pass if local fail-safe and user override are explicit.
Source: source
Safety¶
Week 18: Closed-loop safety and graceful degradation¶
Reading/source: Control/safety + physiology
Know: Guarantee bounded behavior under decoder/stimulator/sensor failure.
Reconstruct: Derive invariant/safe-set constraint for simplified assistive controller.
Do: Inject drift, dropout, adversarial artifact and actuator saturation into closed-loop simulator.
Context: Useful neurotech must fail safely while biological state remains partially hidden.
Defend: Which failure mode is detectable before harm?
Gate: Pass if independent monitor catches unsafe command path.
Source: source
Research¶
Week 19: Failure mining in BCI/neuroprosthetic loops¶
Reading/source: All track traces/current literature
Know: Create taxonomy: signal, decoder, user adaptation, physiology, interface, task, safety.
Reconstruct: Build causal graph and rank by severity/frequency/tractability.
Do: Generate 100 simulated failure episodes and classify mechanisms.
Context: The research target should be persistent failure, not leaderboard fashion.
Defend: Which failures are fundamentally information-limited?
Gate: Pass if taxonomy predicts intervention class.
Source: source
Week 20: Reproduce an open neurotech analysis¶
Reading/source: Choose public neural dataset/paper
Know: Practice reproducible decoding/encoding research.
Reconstruct: Reconstruct preprocessing, train/test split, temporal leakage and metric.
Do: Reproduce central decoding result and rerun with session/subject-held-out split.
Context: Neural datasets are especially prone to non-independent samples.
Defend: Does performance survive new-day/new-subject evaluation?
Gate: Extension Gate: reproduction + leakage/OOD audit.
Source: source
Week 21: Independent adaptive-interface extension¶
Reading/source: Selected failure mechanism
Know: Design extension improving robust user utility rather than offline score.
Reconstruct: Write mechanism and expected benefit/cost.
Do: Implement adaptation, uncertainty, active calibration or feedback change; pre-register test.
Context: A good extension changes closed-loop behavior.
Defend: What user population/context could be harmed by your adaptation?
Gate: Pass if improvement holds under drift and user-model variation.
Source: source
Translation¶
Week 22: Hardware, packaging and long-term operations¶
Reading/source: Electronics/materials/manufacturing/safety core
Know: Backchain prototype to power, telemetry, packaging, calibration, maintenance and update lifecycle.
Reconstruct: Create reliability/maintenance/firmware-update trust model.
Do: Design field-deployment architecture for non-clinical BCI analogue or assistive interface.
Context: Home deployment exposes support/reliability problems hidden in lab studies.
Defend: Who restores safe function after software/hardware update failure?
Gate: Pass if lifecycle and rollback are explicit.
Source: source
Capstone¶
Week 23: Integrated human-machine interface¶
Reading/source: All track sources
Know: Integrate signal -> estimator -> controller/feedback -> human adaptation -> safety -> utility.
Reconstruct: Regenerate closed-loop architecture and objective hierarchy.
Do: Capstone: public/synthetic-data BCI, adaptive assistive controller or sensory augmentation simulator with failure tests.
Context: No claims of medical efficacy beyond evidence.
Defend: Where does the human compensate for your technology?
Gate: Systems Gate: signal-processing + control + human-factors + safety reviewers.
Source: source
Week 24: Research and deployment defense¶
Reading/source: NIH current reality + all work
Know: Produce roadmap that distinguishes lab capability, assistive utility and speculative augmentation.
Reconstruct: Write evidence ladder and kill/revision criteria.
Do: Technical paper + reproducible code + user-value metrics + privacy/safety case + 12-month program.
Context: Current speech BCI progress shows practical home use is becoming a real benchmark, not sci-fi alone.
Defend: What would make the user choose not to use your system?
Gate: Capstone Gate: technical reviewer + clinician/human-factors proxy + privacy/safety reviewer.
Source: source
Research gates¶
G1 Replication¶
Required performance: Reproduce public neural decoding/encoding result with strict session/subject/time split.
Minimum artifacts: Data/code; preprocessing; leakage audit; latency/calibration metrics.
Pass criterion: Offline random-split accuracy is insufficient.
G2 Extension¶
Required performance: Improve closed-loop user utility under drift/adaptation rather than benchmark score alone.
Minimum artifacts: Adaptive/uncertainty method; simulated user variation; OOD sessions; safety constraint.
Pass criterion: Must report latency, calibration, robustness and user-task metric.
G3 System Closure¶
Required performance: Close signal->estimator->feedback/stimulation abstraction->human adaptation->device lifecycle/safety.
Minimum artifacts: Noise/latency budget; privacy threat model; safe-state monitor; maintenance/update plan.
Pass criterion: Independent safety constraint cannot be overridden by learned policy.
G4 Research Defense¶
Required performance: Separate assistive evidence from speculative augmentation and defend user agency.
Minimum artifacts: Technical paper; longitudinal eval design; privacy/safety case; 12-month roadmap.
Pass criterion: Must answer why a user might rationally reject the system.
Frontier technologies primarily routed here¶
- 24. Advanced prosthetic limbs: class A
- 25. Neural prostheses: class B
- 26. Speech/motor brain-computer interfaces: class A
- 38. Virtual reality for training/therapy: class A
- 40. Holographic/spatial interfaces: class B
- 56. Sensory augmentation: class B
- 57. Memory prostheses: class B
- 58. Non-invasive neural interfaces: class B
- 59. Brain-to-brain communication: class D
- 88. Full-dive neural VR: class D
- 91. Whole-brain emulation: class D
- 92. Mind uploading: class D