Fig. 2: CNN training of ML algorithm to predict molecular structure.
From: Machine learning for laser-induced electron diffraction imaging of molecular structures
a The 2D-DCS is convoluted by different filters to generate a collection of feature maps. b The collection of feature maps is first flattened into a one-dimensional array and multiplied by the weights in each neuron of all layers to predict the atomic position for each atom in the molecule. c Contour plot of the cost function for two weights (ωi and ωi+1). The blue dot represents the cost function value, and the red arrows show the direction of the gradient of the cost function. After five iterations, the cost function is minimized.
