Figure 6. The haptic-temperature fusion based on
VO2 volatile memristor and multilayer perceptron by
simulation. a) The circuit diagram of the haptic-temperature fusion. b)
Neuronal response of the haptic-temperature fusion under 100 g & 27 °C,
200 g & 17 °C and 200 g & 32 °C. The output frequencies are 0.5, 0.6,
and 1.1 MHz at the
haptic-temperature fusion under 100 g & 27 °C, 200 g & 17 °C and 200 g
& 32 °C, respectively c) Schematic of the multilayer perceptron for
pattern classification. d) Evolution of the training accuracy with
training epoch. The line with blue circles is the training accuracy
based onV max-V mean-finputs, which reaches 91.35% after 200 epochs. The line with orange
triangles is the training accuracy based onV max-f inputs, which reaches 86.03% after
200 epochs. The line with green squares is the training accuracy based
on V mean-f inputs, which reaches 83.25%
after 200 epochs. e) Confusion matrix of the testing results, showing
that the test inputs are well classified after training.
In order to further illustrate the potential of the present
VO2 neuron in multisensory fusion, we have tested a
total of 11 combinations of different pressures and temperatures (Figure
S14, Supporting Information), and the measurements were repeated 10
times for each individual combination to form a small dataset. A
multilayer perceptron (MLP) network was constructed as shown in Figure
6c, which is composed of 20 hidden neurons and 11 output neurons, and
the total 110 data samples were used to train the network to identify 11
different combinations using backpropagation. According to the
experimental data in Figure S14, it can be found that only the
oscillation frequency also cannot classify all 11 situations.
Fortunately, it has been revealed that the temperature elevation also
leads to variations in the amplitude of oscillation (Figure 5e). As a
result, the maximum and mean of the output voltage signal have also been
extracted (Figure S14) and included as input variables together with the
oscillation frequency (f ). We have investigated three
haptic-temperature input combinations, namely,V max –f , V mean – f andV max – V mean – f ,
and the training is performed for 200 epochs. Figure 6d shows that the
classification accuracies of the network are 86.03% and 83.25% based
on V max – f andV mean – f inputs after 200 epochs,
respectively. When V max –V mean – f information is used as input,
the classification accuracy can attain to 91.35% after 200 epochs,
which has a better training performance. Figure 6e further shows a
confusion matrix of the testing results for the 11 cases based onV max – V mean – finputs. As a measure on the classification accuracy, the confusion
matrix in Figure 6e displays the classification result in each column
while the expected (actual) result in each row, where the number of
instances is depicted by the color bar, demonstrating that the test
inputs are well classified after training.