Fig. 3: The hierarchical BayesNN system based on 3D 16-layer Fe-diode array.

a The hierarchical BayesNN oriented towards MNIST recognition, with the same layer weights having the same deviation to capture weight correlations, from the input layer to the output layer, with standard deviations of the weights being, σ1, σ2, σ3 and σ4. b The workflow for in situ training based on backpropagation and expectation maximization. c The details of updating weights and hierarchical deviations. d Setup for hardware implementation of the hierarchical Bayes-NN based on 3D Fe-Diode. e 3D Fe-Diode chip. f SEM image of the test chip. g SEM image of the 16-layer 8 × 64 Fe-diode array test chip.