Fig. 4: Image classification using 3D FeNAND array.
a Example of training and test pattern set. Two-class image set which was comprised of a total of 20 training images of black and white patterns representing line patterns with a size of 4 × 2 pixels was used as a training and test image set. b Schematic illustration of binary image classification using 3D FeNAND. The value of each pixel in the input image was converted into voltages of 0 (black) or 1 V (white) and assigned to each BL in the 3D FeNAND array. The output ISL was measured and used as the input for neurons (op-amp). c Neuron output voltages with different input patterns. The 3D FeNAND only showed a high output voltage when the black line was positioned over the white line. d Neuron output voltages according to repetitive input patterns. Pattern 1 (black line positioned over the white line) and pattern 2 (white line positioned above the black line) were used as input patterns. e Schematic illustration of MLP network for classification of MNIST hand-written digit images. 400 elements that corresponded to the number of pixels of input images (20 × 20) were used as the input and 100 hidden- and 10 output neurons were used for classification. f Comparison of simulated accuracies of MLP network based on 3D FeNAND and that based on ideal devices.
