Neuromorphic learning with Mott insulator NiO

Abstract

Habituation and sensitization (nonassociative learning) are among the most fundamental forms of learning and memory behavior present in organisms that enable adaptation and learning in dynamic environments. Emulating such features of intelligence found in nature in the solid state can serve as inspiration for algorithmic simulations in artificial neural networks and potential use in neuromorphic computing. Here, we demonstrate nonassociative learning with a prototypical Mott insulator, nickel oxide (NiO), under a variety of external stimuli at and above room temperature. Similar to biological species such as Aplysia, habituation and sensitization of NiO possess time-dependent plasticity relying on both strength and time interval between stimuli. A combination of experimental approaches and first-principles calculations reveals that such learning behavior of NiO results from dynamic modulation of its defect and electronic structure. An artificial neural network model inspired by such nonassociative learning is simulated to show advantages for an unsupervised clustering task in accuracy and reducing catastrophic interference, which could help mitigate the stability-plasticity dilemma. Mott insulators can therefore serve as building blocks to examine learning behavior noted in biology and inspire new learning algorithms for artificial intelligence.

Keywords: Mott insulator; neuromorphic learning; transition metal oxides.

Fig. 1.

Schematic illustration of nonassociative learning behavior including habituation and sensitization in Aplysia versus NiO devices. (A) Gill withdrawal reflex behavior of Aplysia and schematic of its neuron circuit. A light touch on the siphon leads the gill to withdraw. Repeating the tactile stimulus results in habituation learning behavior with reduced gill withdrawal. By applying a noxious stimulus, such as shock to the tail, the tactile induced gill withdrawal can be enhanced as a sensitization behavior (8). (B) The actional potential measured from the sensory neuron and motor neuron during the light tactile and sensitizing stimulus, where habituation is associated with the gradual diminishing of motor neuron potentials while sensitization is coupled with substantial enhancement of motor action potentials (11). (C) Schematic of habituation and sensitizing learning behavior in NiO devices. (D) A short exposure of the NiO device to H₂ gas causes electron filling and corresponding sharp electrical resistance increase. This response will gradually decrease, i.e., habituation, upon repeating H₂ exposure and being sensitized by exposure to an O₃ stimulus, which enriches the density of holes. (E) Example of a layer of synaptic connections in an ANN. (F) Unsupervised clustering task in a layer of an ANN employing habituation and sensitization inspired plasticity for improving accuracy and adaptability (SI Appendix, Supplementary Note 2).

Fig. 2.

Habituation and sensitization behavior of NiO devices under successive cycles of environmental stimulus. (A) Relative resistance change R/R₀ of NiO device under switching ON (15 min) and OFF (15 min) 5% H₂ for four exposure cycles at 200 °C. A reversible and nearly identical response is observed. (B) Response (R/R₀) of NiO devices under repeating expose to H₂ with an additional O₃ exposure before the eighth H₂ cycle. R₀ is the resistance of the devices at the beginning of each H₂ exposures. The H₂ OFF period is reduced to 45 s here. The O₃ exposure time is 15 s. With reducing the H₂ OFF time here, the response of NiO from the first to seventh H₂ exposures shows a continuous decrease, demonstrating a habituation learning behavior. After the O₃ stimulus, the response of NiO devices becomes even larger than that of first H₂ exposure, demonstrating a sensitization behavior. (C) Relaxation time as a function of the training process shown in B, where τ₁ and τ₂ were obtained by fitting the resistance response with a two-exponential function. (D) Time-scale dependence of habituation behavior of NiO devices. With shorter H₂ OFF time, the habituation behavior becomes more prominent, indicating a time-dependent plasticity. (E) Dependence of sensitization behavior on the strength of the O₃ stimulus. The response of NiO devices at first and sixth H₂ exposure becomes stronger with increasing the exposure time of the sensitization stimulus.

Fig. 3.

Evolution of NiO upon H₂ and O₃ gas exposures. (A) In situ resistivity-time results of NiO devices during exposure of 5% H₂ and O₃ gases at 200 °C. (B) Normalized second harmonic near-field amplitude s₂(NiO)/s₂(Pt) of NiO devices, indicating an elevation of carrier concentration with H₂, pristine, and O₃ exposure, respectively. (C) Room-temperature synchrotron XRD profile of the NiO (111) peak after the exposures. Overall, the out-of-plane lattice constant of NiO changes by 0.27% between H₂ and O₃ exposures. (D) O 1s X-ray photoemission spectra of NiO after exposures, which were fitted with lattice oxygen of NiO (O_L, orange), excessive oxygen defects (O_D, purple), and hydroxide group (-OH, blue). A predominant hydroxide group peak in H₂-exposed NiO indicates the incorporation of hydrogen into the lattice, while the enlarged oxygen defects peak in O₃ exposed NiO demonstrate accumulation of excessive oxygen defects. (E) Normalized Ni L edge XANES spectrum of NiO after exposures. Under H₂ exposure, the spectrum weight shifts to lower energy level, indicating a decrease of the Ni valence state. On the contrary, after exposure to O₃, the spectral weight transfers to a higher-energy component, indicating an increase of the Ni valence state. (F) O K edge XANES spectra of NiO after exposures. In O₃, the prepeak at the O–K edge shifts its weight from ∼532 eV to ∼529 eV, indicating the formation of the hole state, corresponding to resistivity decrease. In H₂, the diminishing of the prepeak at ∼532 eV indicates electron filling that results in substantial resistivity increase. (G) Schematic of the in situ XANES setup to monitor the response of NiO in H₂. (H) In situ Ni K edge XANES spectra of pristine and O₃-exposed NiO, where the marked area near the absorption edge is magnified in I. The lower energy shift of the absorption edge of O₃-exposed NiO (∼0.2 eV) is larger than that of pristine NiO (∼0.1 eV) after a shorter exposure period, confirming the sensitization behavior.

Fig. 4.

DFT+U computed electronic structural change in NiO with the environmental stimulus of excess oxygen. (A and B) NiO with excess O (shown in yellow) atoms can potentially locate in two energetically favorable positions, one being stable and another in the metastable configuration. We denote them as location 1 and location 2 and present them in schematic diagrams; we label various Ni positions as Ni-Near and Ni-End. On the right we present the zoom-in view of added O in two different locations of rock-salt NiO. (C and D) The total DOS for the majority spin component with and without excess O in NiO. (E and F) Projected DOS for Ni-3d and O-2p for the excess O are plotted for these two configurations. For the stable configuration, the excess O appears as the top valence band while for more interesting metastable configuration it appears as the first unoccupied band, which is located inside the gap between lower and upper Hubbard band for undoped NiO.