Fig. 6: Transfer learning with the hybrid FeCAP/memristor memory circuit.
From: A ferroelectric–memristor memory for both training and inference
a, Schematic showing the transfer learning strategy with initial MobileNet-V2 architecture pre-trained on a base dataset using a quantization-aware training strategy to obtain a fixed quantized feature extractor (QFE). b, The MobileNet-V2 quantized feature extractor was used in conjunction with a newly initialized classifier for transfer learning of the pre-trained model toa new dataset. The proposed HM-based training strategy was used to train the classifier. c, Test accuracies on the five evaluated datasets for different classifier (C) configurations, varying in precision (32-bit floating point or 4-bit) and training strategy (offline or HM-online). The bar chart represents the mean value, and the dot plots show the corresponding individual data points (n = 10).
