Fig. 4: The conductance of devices at the BayesNN’s output layer. Each BL connects 16 layers, using a total of 160 BLs (80 positive, 80 negative) to store the output layer weights. | Nature Communications

Fig. 4: The conductance of devices at the BayesNN’s output layer. Each BL connects 16 layers, using a total of 160 BLs (80 positive, 80 negative) to store the output layer weights.

From: Bayesian neural network with unified entropy source and synapse weights using 3D 16-layer Fe-diode array

Fig. 4

a–c show the initial, intermediate, and final states of conductance values during in situ training, respectively. d Comparison of minimum entropy and MNIST recognition accuracy based on Fe-diode-based entropy source and CMOS Latch-based entropy source under different temperature conditions. e Accuracy trend over training epochs. In situ training results follow the software simulation closely, with a 3–5% gap. f A certain example generated by BayesNN. g An uncertain example generated by BayesNN.

Back to article page