† Corresponding author. E-mail:
Project supported by the National Natural Science Foundation of China (Grant No. 61332003) and High Performance Computing Laboratory, China (Grant No. 201501-02).
Memristors, as memristive devices, have received a great deal of interest since being fabricated by HP labs. The forgetting effect that has significant influences on memristors’ performance has to be taken into account when they are employed. It is significant to build a good model that can express the forgetting effect well for application researches due to its promising prospects in brain-inspired computing. Some models are proposed to represent the forgetting effect but do not work well. In this paper, we present a novel window function, which has good performance in a drift model. We analyze the deficiencies of the previous drift diffusion models for the forgetting effect and propose an improved model. Moreover, the improved model is exploited as a synapse model in spiking neural networks to recognize digit images. Simulation results show that the improved model overcomes the defects of the previous models and can be used as a synapse model in brain-inspired computing due to its synaptic characteristics. The results also indicate that the improved model can express the forgetting effect better when it is employed in spiking neural networks, which means that more appropriate evaluations can be obtained in applications.
The memristor, called the fourth fundamental element besides resistor, capacitor, and inductor, was predicted by Prof. Chua theoretically in 1971.[1] It has received broad attention since being discovered by HP labs in 2008.[2] The memristor has broad prospects in many fields: it is a promising candidate for nonvolatile memory and may change the existing computer structure,[3–5] and it can be employed in image processing and other fields.[6,7] What is more, it is a promising candidate for synaptic devices in brain-inspired computing.[8–10] In this context, it is significant to build a good model for memristor application research.
HP labs presented a drift model when they discovered the first memristor.[2] However, HP’s model has some significant deficiencies; several window functions have been proposed to improve it. Joglekar et al. proposed a window function to overcome the terminal overflow; Biolek et al. proposed a window function to solve the terminal lock and so on.[11–14] However, these window functions have their blemishes more or less. In this paper, a novel window function is proposed. It can overcome the terminal overflow and terminal lock and has good flexibility.
In later studies, researchers found the forgetting effect in several experiments, which is also called short term memory or volatile memory.[15–17] In order to describe this phenomenon, many models have been proposed.[15,18,19] Berdan et al. and Ling Chen et al. proposed different drift diffusion models based on research results of Strukov.[20–22] Though both models can describe the forgetting effect to some extent, they have some deficiencies that the forgetting effect in terminal (x = 0, 1) cannot be embodied in some situation that is common in fact. The forgetting effect has essential influences on accuracies and functions of applications, such as non-volatile memory and brain-inspired computing based on memristors. When a memristor model that cannot express the forgetting effect well is employed in simulations, it may cause serious errors. In this paper, we present an improved drift diffusion model to overcome shortcomings of the previous models. The threshold characteristic is added to our model since the existence of threshold has been discovered in several memristors. Spike-timing-dependent plasticity (STDP) of our model is verified, which means that the improved model can be employed as a synapse model. Moreover, the improved model is used in spiking neural networks (SNNs) to prove its merits. Simulation results indicate that the improved model can overcome deficiencies of the previous works and it can express the forgetting effect better when it is used as the synapse model in SNNs.
The drift model proposed by HP labs is shown in Fig.
The model is described as
In order to improve HP’s drift model, many window functions have been proposed. Jogleka’s window function is expressed as
It is easy to know that function y = 1/x is a decreasing function in (0, + ∞). So it can be exploited as a basic model of the window function, and the model can be expressed as f(x) = 1/(ax + b) + c. When the stimulus is positive and x increases continuously, it is necessary to make f(0) = 1 and f(1) = 0. A set of solutions are a = b = 1/2 and c = −1, the function is f1(x) = 2/(x + 1) − 1. When the stimulus is negtive and x decreases continuously, it is easy to obtain f2(x) = 2/(2 − x) −1. Then considering f1(x) and f2(x) comprehensively, we introduce
As shown in Fig.
In short, our novel window function has obvious advantages compared to the previous ones. It uses two parameters and two variables to overcome all shortcomings, including the terminal overflow and terminal lock. The scale of p is [0, + ∞), which means great flexibility. And j can modify the scale of f(x).
Researchers have found the forgetting effect in several memristor experiments. Some models have been proposed to describe this phenomenon. Berdan and Chen proposed different drift diffusion models based on HP’s drift model and the research results of Strukov.
Berdan’s model is expressed as
Ling Chen’s is expressed as
Obviously, both Berdan’s model and Chen’s models have deficiencies when x = 0, 1. In this section, an improved drift diffusion model is proposed to overcome problems on terminals (x = 0, 1). According to the discussion above, it is easy to find that the limitation of the window function should be decided by l−(x − ε)/τ (or l−(x − y)/Rx) when x = 0, 1. The improved model can be expressed as
The threshold characteristic has been discovered in several memristors. In order to build a good model, taking the threshold into account is necessary. Expression (
Simulation results of the improved model are shown in Figs.
In brief, the improved model overcomes the defects of Berdan’s model and Ling Chen’s model and can describe the threshold characteristic. So this model can describe the characteristics of memristors better and be used.
It has been found that memristors have synaptic characteristics and can be used as synapses, this is why memristors are used broadly in brain-inspired computing. Spike-timing-dependent plasticity (STDP) is a significant characteristic of synapses, and is usually used as an evaluation standard of artificial synapses. When a memristor model is used as a synapse model, it is important to have STDP.
As shown in Fig.
Since memristors can be employed as synapses, composing SNNs with neurons for pattern recognition is an important application of memristors in brain-inspired computing. To demonstrate that the improved model can represent the forgetting effect better in application studies, SNNs based on the improved model, Berdan’s model, Chen’s model, and the model without forgetting effect respectively are employed to recognize digit images of “0”–“9”. As shown in Fig.
Before training, the states of all memristors are mid-resistance, which means that the models are set as x0 = 0.5. During the training process, the memristors corresponding to the black blocks are trained to LRS and the memristors corresponding to the white blocks are trained to HRS, which means that all memristors are in the terminal state. The states of the memristors after training are shown in Fig.
Recognizing image “1” from other images is shown as an example. As shown in Fig.
When the forgetting effect can influence memristors’ states, taking the forgetting effect into account is essential when the memristor model is employed as the synapse in the studies of brain-inspired computing. From the simulation results, we can find that the improved model can express the forgetting effect of memristors better than the other models. The results also indicate that there will be an inaccurate evaluation of recognition if the other models are used to simulate the forgetting effect, but a more appropriate evaluation can be obtained when the improved model is used. Therefore, the improved model can work better in the situations that the forgetting effect plays a significant role.
A novel window function is proposed and has good performance. Deficiencies on x = 0, 1 of the previous drift diffusion models are analyzed, and an improved drift diffusion model is presented. The improved model deals well with influences of diffusion on drift when x = 0, 1. The threshold characteristic is added to the improved model to make it more common and the STDP curve of the improved model coincides with the biological synapse well. Different models are used as synapses respectively in SNNs to recognize digit images to examine their performances in representing the forgetting effect. Simulation results show that the improved model can express the forgetting effect better than the previous models.
[1] | |
[2] | |
[3] | |
[4] | |
[5] | |
[6] | |
[7] | |
[8] | |
[9] | |
[10] | |
[11] | |
[12] | |
[13] | |
[14] | |
[15] | |
[16] | |
[17] | |
[18] | |
[19] | |
[20] | |
[21] | |
[22] |