Pairs a minimally-trained novice with an AI model to catch cerebral ischemia from intraoperative EEG—matching board-certified neurophysiologists (statistically non-inferior) while halving the AI's false alarms. Expert-grade surgical monitoring, without the expert.

Benchmarks five machine-learning models for detecting cerebral ischemia from quantitative EEG during surgery. Random forests lead on sensitivity and precision—and the top predictors, alpha-band power and hemispheric asymmetry, mirror exactly what expert readers watch for.

When expert neurophysiologists are scarce and AI alone isn't safe enough for the operating room, can a non-expert help close the gap? Human–AI deferral routes only the hardest cases to a minimally-trained novice—and with just 5–10% novice involvement lifts precision and sensitivity by ~40% over the AI alone. The proposed method, Learning to Augment (L2A), softly blends novice and AI predictions to stay robust and well-calibrated even when the novice is unreliable, concentrating its rare hand-offs so long stretches of monitoring run untouched.

Not all spurious features are harmful—a shortcut only hurts generalization when it's easier to learn than the real signal. By watching a network's training dynamics, the paper shows these harmful shortcuts reveal themselves in the early layers, early in training—catchable and fixable before they ever reach deployment.

An accurate model that's also overconfident is dangerous—in the clinic or the driver's seat. ACE fine-tunes a classifier on counterfactual augmentations—realistic images nudged across its decision boundary—so it flags ambiguous and out-of-distribution inputs instead of guessing confidently, all without sacrificing accuracy.

The first system to re-identify individual, unmarked fruit flies across days—no tags, no paint, no frame-by-frame tracking. Trained on ~3M images of 60 flies, it names and beats a key failure mode—"cross-day accuracy decline"—using domain-adversarial training to learn day-invariant features, keeping most flies above 90% accuracy.

Could a fruit fly recognize another fruit fly? Building a "fly-eye" network that mirrors Drosophila's real visual connectome, this work shows their low-resolution vision—just ~29×29 pixels—carries enough to tell individuals apart, something humans do near chance. A far richer visual world than assumed.

The undergraduate work behind the WACV 2019 paper: recognizing individual Drosophila with ResNet CNNs, tag-free. Same-day accuracy is near-perfect—but a few "problem flies" collapse on later days as their looks drift, revealing temporal shift, not model capacity, as the real obstacle. Domain-adversarial training then learns cues that identify the fly while ignoring the day, lifting ≥90%-accuracy re-identification from 47 to 57 of 60 flies.