Acquisition and Extinction of Associative Memory Realized in Memristive Devices with Intrinsic Forgetting Effect

Abstract

Realization of the artificial intelligence (AI) has always been the dream of human beings. Recently, the Google AlphaGo beats human Go champion Lee Sedol boil the AI field. This primary AI based on software programming and traditional digital computing faces to several insurmountable obstacles of high design complexity and severe power dissipation. A potential approach to reach higher level AI is to build an artificial neural network system based on neuromorphic computing in which biological-like electronic synapse is regarded as the basic physical building block1. Memristive devices with continuously tunable resistance state can be used to emulate the synaptic plasticity effects2. Several significant synapse properties such as exciting postsynaptic plasticity (EPSC), paired pulse facilitation (PPF), spike timing dependent plasticity (STDP) have been realized in various memristive devices. In this work, continuous conductance states with intrinsic current decay, namely, forgetting effect, were obtained in the Ni/Nb-doped SrTiO3/Ti memristive device. Then, the associative learning (i.e. classical conditioning reflex of Pavlov’s Dog), a type of more advanced learning function, was realized and the forgetting effect on the acquisition and extinction of associative memory was first investigated based on the intrinsic current decay of the present devices. As shown in Fig 1(a), the current gradually increases with positive voltage sweeps and decreases with negative voltage sweeps and overlapping exist between two neighbor loops. In another way, the current increases with different positive voltage pulse sequences as shown in Fig. 1(b). Higher voltage amplitude results in more obvious current increase and obvious current decay follows after each pulse stimuli. The intrinsic current decay effect is also referred to the forgetting effect. The continuously tunable states with forgetting effect guarantee the imitation of the association acquisition and extinction functions of Pavlov’s Dog. A basic training process of Pavlov’s Dog can be simply described as follows: Dogs initially do not respond to a bell ringing, but naturally salivate when the food is presented. In the training procedure, the dogs are repeatedly fed after a bell ringing. Subsequently, the bell alone can evoke a salivary response as the food does which indicates that the association between food and bell builds. However, the salivation behavior will disappear if the bell always rings without food which indicates the association between food and bell fades. A specific device implementation of Pavlov’s Dog training process is given in Fig. 2 by embedding the Ni/Nb-doped SrTiO3/Ti synapse device into a simple electric circuit. Before training [Fig. 1(d)], there’s nearly no voltage response (4.5 mV) to the bell signal which indicates no association between food and bell. During training [Fig. 1(e)], the association between bell and food is gradually built. Here, we define tenfold voltage response of “Before Training” period (45 mV) as the signal of salivation (i.e. training success). The dog exhibits salivation behavior after 5 times food and bell combining signal training. After training [Fig. 1(f)], the salivation behavior disappears after 5 times bell signal which can be considered that the dog forgets the association between food and bell. Figure 1. Electrical conductance continuously tuning in the Ni/Nb-doped SrTiO3/Ti memristive device. (a) The current increases with positive voltage sweeps and decreases with negative sweeps. (b) The current increases after each pulse application and then spontaneously decays with time. The implementation of Pavlov’s Dog training process using the Ni/Nb-doped SrTiO3/Ti memristive device. (d) The voltage response for each bell signal, (e) the voltage response for each food and bell coupling signal, (f) the voltage response for each bell signal. 1 L. Q. Zhu, C. J. Wan, L. Q. Guo, Y. Shi, Q. Wan, Nat. Commun.5, 3158 (2014). 2 Z. H. Tan, R. Yang, K. Terabe, X. B. Yin, X. D. Zhang, X. Guo, Adv. Mater. 28, 377-384 (2016). Figure 1