Effects of presynaptic synapses on the recognition rate of a neural network

We studied the properties of a neural network where the rank of the coupling tensor is greater than two. In this case, in addition to the matrix of synaptic connections, there are presynaptic synapses. Such coupling tensors occur in hardware implementations of neural networks based on crossbar arrays. An inherent property of the crossbar scheme is the presence of parasitic currents: a signal from one wire going to a specific neuron flows through memory cells (synapses) to the wires of all other neurons. As a result, noise is superimposed on the input signal of the neuron, weakening signals going to all other neurons. In this scenario, the conductivity of the analog crossbar cell changes proportionally to the noise signal, causing the output signal of the cell to acquire a nonlinear relationship with the input. We showed that with a specific type of coupling tensor, the quality of the neural network algorithm improved significantly. We used a network similar to the Hopfield network as an example.