Ningbo Materials is in the field of research on flexible oxide neuromorphic transistors

"Artificial Intelligence (AI)" was introduced in the 1950s and has experienced a period of slow development. However, since the advent of AlphaGo in 2016, AI has become one of the global research hotspots and has received much attention.

It is worth noting that the existing AI technology is mainly based on the traditional von Neumann architecture, which needs to be implemented with more complicated computer code. The calculation module is separated from the storage module, so its parallel computing capability is limited, and the energy consumption is relatively low. High, it has certain limitations for the processing and calculation of unstructured big data in the future.

At the same time, in recent years, the construction of artificial biological nervous systems based on the device level is also becoming an important branch of the AI ​​field. As the basic unit of human brain cognitive behavior, synapse is the key part of the connection between neurons and an important starting point for constructing artificial neural networks.

In the field of synaptic biomimetic electronics, the current research mainly includes two-end resistive devices and three-terminal transistors. These devices have mimicked various synaptic functions and neuron functions from simple to complex, and have potential application prospects. .

They fabricated a ITO synaptic transistor with a learning behavior using a chitosan film as a gate dielectric on a flexible PET substrate. After 1000 times of mechanical bending stress, the performance parameters of the device remained stable; After 8000 seconds of compressive stress, it was found that the threshold voltage of the device showed a certain drift, indicating that the developed transistor has learning ability. Subsequently, three synaptic functions were simulated on the developed flexible ITO thin film transistor: post-synaptic excitatory current (EPSC), double pulse facilitation (PPF), and spike timing dependent plasticity (STDP). In 1968, Atkinson and Shiffrin proposed the "multiple memory model of human brain" from the psychological level: the process of perceptual memory (SM) to short-term memory (STM) and short-range memory to long-term memory (LTM). The team used the design of the gate pulse stimulation frequency and the gate pulse stimulus intensity to simulate the "human brain multiple memory model" on a single synaptic transistor. The above results are published in ACS Applied Materials Interfaces, 2018, 10 (19), 16881-16886.

The physiologically famous "Pavlov dog conditioned reflex" (ie, classical conditioning experiment) is an important type of associative learning behavior, which reflects the influence of conditional and unconditioned stimuli on neuronal activity, in a single The imitation of this association learning behavior on the device is an important research content of brain-like neuromorphic devices. It is worth noting that the STDP learning rule is an important synaptic learning behavior, which plays an important role in the cognitive behavior of the nervous system. It reflects the influence of presynchronous and post-synaptic stimulation on synaptic weights and is an important step in regulating high-level neural activity. Touch the learning mechanism. It can be seen that conditioned reflection has certain similarities with the STDP learning rule. Inspired by this, the team developed a re-stickable oxide neuromorphic transistor using transparent polyimide (PI) tape as the substrate, followed by design. Synaptic stimuli of different waveforms successfully mimicked four types of STDP learning behaviors in biological synapses on a single device, including Hebbian STDP, anti-Hebbian STDP, symmetric STDP, and visual STDP. The test curve fitting parameters of Hebbian STDP are similar to those measured on the biological synapse, indicating that the neuromorphic transistor has brain-like operational characteristics. Based on the STDP learning rule, the imitation of classical conditioning behaviors, including information acquisition, regression, and recovery, can be achieved on a single neuromorphic transistor without the need for additional complex circuitry and components. In addition, the conditional inhibition behavior in classical conditioning is successfully simulated, which is the first report in the study of neuromorphic devices. The results are entitled "Restickable Oxide Neuromorphic Transistors with Spike-Timing-Dependent-Plasticity and Pavlovian Associative Learning Activities" and published in Advanced Functional Materials 2018, 28 (44) 1804025.

The above work was funded by the National Natural Science Foundation of China, the Zhejiang Outstanding Youth Fund, the Chinese Academy of Sciences Youth Innovation Promotion Association, and the Ningbo Science and Technology Innovation Team.