It is the era of computer-aided diagnosis, and ten thousand mammograms cross your system — but only about half a percent hide a genuine malignancy, and the costs of the two errors could not be more lopsided. Flag which scans warrant a radiologist's second read. A naive classifier maximizing accuracy will simply call everything benign and be 99.5% right while missing every cancer; meanwhile each false alarm spends a radiologist's scarce attention and, repeated, burns the trust that keeps them using the tool at all. Balance sensitivity against false alarms with the real asymmetric costs baked in. Get it wrong and you either miss the tumors the system exists to catch, or cry wolf so often the radiologists switch it off — extreme imbalance is the whole difficulty.
Chose Porting attention to a new modality — right call.
Vaswani's defining contribution is the 2017 transformer — an architecture breakthrough in sequence modeling that later colonized vision. In a modern pipeline, a transformer backbone could certainly featurize mammograms, and Vaswani operates comfortably in the large-scale training regime such a system needs. But the transformer solves a representation and parallelization problem, not this problem: attention does nothing intrinsic about a 0.5% positive rate, asymmetric misclassification costs, or the calibration of an operating threshold that keeps radiologists from switching the tool off. Those difficulties live in the loss function, the sampling scheme, and the decision layer — territory where Vaswani has no signature methods. He supplies excellent modern machinery for the easy part of the task while the hard part, extreme-imbalance cost-sensitive decision-making, lies outside his published toolkit.
Clauset's toolkit, built in the 2000s-2010s, is the statistics of rare and extreme events: rigorous power-law fitting, community detection, and a career-long insistence that heavy-tailed claims be tested rather than eyeballed. That temperament transfers well to a 0.5% prevalence problem — he would immediately distrust naive accuracy and demand honest evaluation of the tail class, likely via careful resampling and calibrated base rates. What he lacks is the classifier-construction machinery itself: he did not develop cost-sensitive learning methods, image features, or ROC-threshold optimization, and mammography pixels are far from network data. He is the right auditor for someone else's CAD system — checking whether its rare-event claims survive scrutiny — rather than the person who builds the sensitivity-versus-false-alarm tradeoff into the model. Useful, but adjacent.
Battle #106 · 8/10/2026, 11:38:17 AM · this result is deterministic: the same two personas on this problem always resolve the same way.