Sample-Efficient Deep Learning for COVID-19 Diagnosis Based on CT Scans

Xuehai He; Xingyi Yang; Shanghang Zhang; Jinyu Zhao; Yichen Zhang; Eric Xing; Pengtao Xie

Existing methods adopt various mechanisms to preserve and sample key vectors [37] , [39] , [51] , [52] . A Siamese-like solution is to use the same network f k = f q on x k and x f simultaneously. Extreme large mini-batch (batch-size up to 8192 [39] ) is required to learn discriminative features from contrasting. This method is straightforward but incredibly expensive in terms of computational resources. Another option is to store the representations of historical keys in a negative key dictionary D k = {k i }, called memory bank [36] . At each iteration, a mini-batch of keys are sampled from the memory bank instead of using f k . The current mini-batch of queries are updated to the memory bank for replacement. This design inherently gets rid of large batch-size with an extended buffer pool. However, the key sampling step involves inconsistency for training the encoder.