Fig. 1

Architecture of the neural networks. a Upper panel: schematic view of the three-layer back propagation neural network (BPNN). Each input value is passed through the neural network. The output value is compared with the desired target output, an error is computed and this error is propagated backward through the network to each node. Lower panel: graphical representation of the model neuron j or threshold unit. The threshold unit receives input, called Xi, from m other units. The associated weight is called W _i . The total input A _j is the sum over all inputs. The activation function f (A _j ) of the neuron is a sigmoid and Y _j is the output of the neuron. b Schematic diagram of the algorithm used to detect 2D diffusion modes within a trajectory. The trajectory is split in overlapping segments of length S1 using a sliding window. The MSD curve is calculated for each segment, normalized, presented to the neural network and classified according to three main diffusion modes (Brownian, directed or confined). A score (output value) is obtained for each frame of the movie and attributed to all the frames of the sliding window