Hiding latency with smoothing or prediction can introduce misclassification

What it is

Latency looks bad, so engineering lengthens smoothing to steady the output, or predicts to throw a cursor or label forward. Both invent intent that was not in the sensor: smoothing drags a finished event into the present; prediction treats something that has not happened as if it has. Hiding delay costs misclassification. It is not free time travel.

Why it happens

A moving average or low-pass cuts jitter and adds phase lag of about half the window; after the event ends the label is still on—false positives are stretched. Kalman prediction can cover a short lag when the model matches; when physiological intent changes (abandon mid-way, co-contraction), the model still extrapolates the old velocity and the cursor or gesture overshoots the true end. Temporal models on the classifier side (labels several steps ahead) write a prior on future class into the present; they shine on repeated laboratory trials and go brittle when the task switches. Prediction error is time-asymmetric: an early false positive is harder to undo than a late false negative, because the interface already moved.

Studying it

On tasks with sudden turns or aborts, compare raw delay, smoothing, and prediction. Factors: prediction horizon, process-noise setting. Outcomes: overshoot, executions that continue after abort, whether users notice the system jumping the gun. Measuring prediction only on smooth copy-tracking will not show the jump. Ablation should include “prediction off, in-progress state only.”

Where it stops holding

Short-horizon prediction is relatively safe on effectors with large kinematic inertia (a forearm already in flight). Any extrapolation is dangerous on fully voluntary micro-commands that can stop at any instant. Predictive rendering for touch ink is usually reversible (the stroke can change); a physiological command that already switched mode is much less so. If users are trained to “finish the motion to accommodate prediction,” the error has been institutionalized.

Applying it

• Smooth for display, not for trigger; trigger on a shorter window, display may be slower.
• Confine prediction to reversible rendering (pre-light, preload); forbid prediction from committing.
• An abort gesture must interrupt a state still being extrapolated; it cannot wait for the horizon to run out.
• Verify by inserting a clear abandon halfway through an action and counting how often the UI still completes the original command; if that is above zero, shorten the horizon or preview only.

Related

• Same group: C9.11.1 There is inherent delay from motor or affective intent to stable physiological detection · C9.11.2 Latency magnitude differs by signal type; EMG is faster than heart-rate change · C9.11.3 Physiological input latency usually exceeds pointing devices and does not fit operations that need immediate feedback
• Adjacent: C5.11 Inking Latency and Predictive Rendering · C9.06 Cost of Sensor False Positives
• Search: predictive rendering · filter lag · intent prediction error