Multilayer Reservoir Computing Based on Ferroelectric α‐In₂Se₃ for Hierarchical Information Processing

Abstract

monolayer limit, [2,4,7] many of which possess semiconducting properties simultaneously, [8,9] making them promising candidates in memory and neuromorphic computing. [10][11][12][13] Extensively studied, van der Waals semiconducting α-In 2 Se 3 exhibits both IP and OOP ferroelectricity with dipole locking effect, [3,14] facilitating resistive switching in vertical [8,15] and lateral junctions [8,16] as well as a modulation effect from a third terminal. [8,17,18] Ferroelectric semiconductor field-effect transistors (FeSFETs) were also fabricated using non-ferroelectric gate stack and ferroelectric α-In 2 Se 3 as a channel for artificial synapses. [10,12,19] Reservoir computing (RC) is a kind of recurrent neural network (RNN), suited for temporal and sequential information processing. The reservoir computing system is composed of an untrained reservoir to map the input into a high-dimensional space and a trained readout layer for classification. Short-term memory (STM) and nonlinearity are two key features for a reservoir, which can be physically implemented by hardware such as memristive devices. [20] Tremendous progress has been made using different types of memristive devices for pattern classification, [21][22][23][24] spoken-digit recognition, [25][26][27] and time series prediction. [21,25,27] Various forms of nonlinearity have been explored and utilized to realize RC, such as electronic, spintronic, photonic, mechanical, and biological mechanisms. In general, the stability of nonlinearity plays an important role in RC performances, which can be quantitatively evaluated by a local Lyapunov exponent. In particular, a negative local Lyapunov exponent indicates that the system is stable, whereas a positive local Lyapunov exponent indicates chaotic behavior. Previous theoretical studies have revealed that the memory capacity of the reservoir is increased when the system stability is decreased. [28] The memory capacity (MC) is maximized at the edge of chaos and then decreases quickly when the system becomes chaotic. Therefore, a form of nonlinearity that can drive the system to the edge of chaos and hence maximizing its memory capacity is desirable. Based on standard RC, different RC architectures have been proposed in pursuit of more complex dynamics and advanced information functionalities. [28,29] A wide RC is composed of several standard RCs in parallel, which has been actually implemented in previous works to improve network performance. [25,27] On the other hand, deep RC consists of multiple Dynamic physical systems such as reservoir computing (RC) architectures show a great prospect in temporal information processing, whereas hierarchical information processing capability is still lacking due to the absence of advanced multilayer reservoir elements. Here, a stackable reservoir system is constructed based on ferroelectric α-In 2 Se 3 devices with voltage input and output, which is realized by dynamic voltage division between a ferroelectric field-effect transistor and a planar device and therefore allows the...

All-ferroelectric implementation of reservoir computing

“…6b, the trained all-ferroelectric RC can correctly classify the sine and square waveforms into their corresponding categories with an NRMSE of 0.13. This NRMSE value is sufficiently low, even lower than the value of 0.2 reported recently in the α-In2Se3 FeFET-based RC system 11 which used high-precision floating point-based weights for the readout network (note: in this work the readout weights are mapped onto the measured conductances of nonvolatile FDs). Such low NRMSE of our all-ferroelectric RC system is attributed to the capability of the volatile FD-based reservoir to produce sufficiently high feedback strength and state richness (Supplementary Fig.…”

“…Recently, emerging hardware-based RC systems have attracted great attention, not only because they have achieved prediction performance comparable to that of the software-based ones in many tasks (e.g., pattern classification 4,5 , speech recognition [6][7][8] , chaotic system forecasting 6,7,9 , and others [10][11][12] ), but also because of the boosted energy efficiency 6,13 . For the hardware implementation of an RC system, the constituent reservoir and readout network need to be implemented on memory devices with distinctly different switching characteristics, i.e., volatile and nonvolatile switching characteristics, respectively.…”

“…Owing to these advantages, ferroelectric memristors have recently been extensively investigated as synaptic devices for FNNs [31][32][33][34] . However, their use in RC systems is still scarce and mainly restricted to the reservoir 11,27,[38][39][40][41] . Hence, allferroelectric implementation of a whole RC system still remains undemonstrated.…”

“…Hence, allferroelectric implementation of a whole RC system still remains undemonstrated. The reason for this is probably because the ferroelectric memristors used hitherto in RC systems, e.g., ferroelectric tunnel junction (FTJ) 27 and ferroelectric field-effect transistor (FeFET) 11,40,41 , possess inherently large depolarization fields (Edps) [42][43][44] , making them easy to exhibit volatile characteristics while difficult to be engineered into nonvolatile memristors to implement the readout network.…”

Memristive, Spintronic, and 2D‐Materials‐Based Devices to Improve and Complement Computing Hardware

“…[117] Finally, memtransistors operate similarly to its two-terminal counterparts (memristors) with the exception that the resistance of the device is gate controlled. In fact, several mechanisms governing resistive switching in memtransistors have been demonstrated, such as grain boundary migration, [112] FE switching, [118] and gate-controlled vdW heterojunctions. [119] Of course, with any of these technologies, due to the analog nature of computations, the idealized vector-matrix computation in Figure 5a is often difficult to achieve.…”