Study of performance improvement of single-electron associative memory circuit

The analog neural network (NN) has attracted much attention in the nano-electronics area because of having many advantages such as parallel-processing and noise-harnessing functions. However, it is exceptional challenging to... [ view full abstract ]

The analog neural network (NN) has attracted much attention in the nano-electronics area because of having many advantages such as parallel-processing and noise-harnessing functions. However, it is exceptional challenging to fabricate analog NN circuits. The crucial problem is how to implement the neurons’ behavior. It is known that a neuron circuit consisting of CMOS is quite complicated and require a large circuit area. Here we focus on single-electron circuits (SECs). It is the nano-electronic circuit which can control transportation of an individual electron tunneling by harnessing a quantum effect named the Coulomb blockade.
In this study, we design the SE Hopfield’s competitive NN circuit and confirm that it can perform the associative memory function. The Hopfield network is a NN model known for its ability to perform the associative memory function. It consists of neurons which have binary output value connected each other via the connection weight Wij. The associative memory is the function which stores some patterns and recalls them from given noisy or a related input pattern.
We prepare one type of SEC called a “single-electron box (SEB)” as a neuron and connect them each other via coupling capacitors (CW) which correspond to the connection weight. The SEB consists of a bias voltage source (Vd), a capacitor (C), a node, and a tunneling junction (Cj) in series. The tunneling junction Cj has a threshold voltage value for the electron tunneling. If the node voltage VN is increased by increasing Vd or the value of an external input and exceeds the threshold value, an electron tunnels through the junction Cj. Then, VN changes sharply.
We stored ten kinds of 5×7-pixel black/white patterns on this network by calculating weight parameters Wij using the learning rule called orthogonal leaning.
As a result of computer simulation, we could confirm that our circuit could perform the associative memory function. As a next step, to achieve higher performance, we focus on the multi-layer NN model and the back propagation which is the learning algorithm for it. We consider the method to construct the multi-layer NN by using SE circuit and apply backpropagation to it.