Extended Data Fig. 9: Distinct roles of input and heat in a 100-bit winner-take-all DNA neural network.
From: Heat-rechargeable computation in DNA logic circuits and neural networks
a, Weighted sum analysis of four example test patterns and a partial version of each test. When a fraction of inputs are missing, the pattern can be classified as the opposite class compared to when the full set of inputs are present. In the four examples shown here, all input bits that exist in the corresponding memory were taken out from the original tests to create the partial tests. The absence of those inputs decreased the test’s similarity to both memories in a biased fashion, reducing one weighted sum all the way to zero while maintaining a positive value for the other weighted sum. As a result, the positions of the tests moved across the diagonal line in the weighted sum space, switching the expected output from class 1 to 2 and vice versa. b, Fluorescence kinetics experiments that demonstrated incorrect output classification when inputs were presented in two subsequent batches and restored correct classification when heat was applied. Similarly, when input inhibitors were introduced to inactivate specific inputs, converting the full tests to partial tests, the output of the DNA neural network was unable to switch until heat was applied.
