Why Corrupted Training Data Doesn't Show Up as High Loss
A developer's controlled fault-injection study shows that corrupted training data, such as shuffled labels or white noise, can reduce loss by 62% without any visible spikes, making it indistinguishabl…