Izhikevich et al. demonstrated the self-organisation of the spiking neurons by a computer simulation using 100,000 artificial neurons and synapses (see the picture here).
The behaviours of neurons and synapses are known to be described by simultaneous differential equations. Then, the neural networks are to be described by the physics of nonlinear dynamics in nonequilibrium systems. The spiking pattern formation of the neural network corresponds to a phase transition (spontaneous symmetry breaking) controlled by the attractor dynamics of the system.
We have an idea that this self-patterning of the neural network can be utilised for developing a neuromorphic architecture to be equipped in any edge-devices. The devices are autonomously controlled and are very low power-consuming. The devices are not the IoT device because they are not necessary to be connected to supercomputers through the internet to ask the decision of the extremely power-consuming deep-learning.
The goal of this research is to demonstrate the game-changing neuromorphic circuit based on the attractor dynamics.
E. M. Izhikevich, J. A. Gally, and G. M. Edelman., Cereb. Cortex 14, 933 (2004)
Our work is not to do computer simulation. We aim at making an original neuromorphic architecture (electronic circuit). For that purpose, we need artificial synapses and neurons. We have fabricated FETs on SrTiO3 with a double layer gate insulator made of a high-k oxide (16nm) and the Parylene-C (6nm). The critical point here is the Parylene-C, which prevents the surface of SrTiO3 from creating oxygen vacancies while applying the gate voltage. The FET on SrTiO3 shows field effect mobility of around 10 cm2/Vs while accumulating carriers from 0 up to 1014 /cm2 continuously. The surface becomes metallic, but the carrier density is still kept increasing, and the conductance also increasing while increasing the gate voltage. This increase is because of the inhomogeneous formation of the metallic region on SrTiO3. Then, we found this conductance change in the metallic region due to the inhomogeneity can be used as spike-timing-dependent plasticity (STDP) required for artificial synapse. We have demonstrated the property on our SrTiO3 FET, and the work was finally reported in IEDM 2017.
However, we know that fabricating artificial synapse is rather trivial. Any analogue resistance-change devices, such as PRAM, MRAM, ReRAM, memristor, can be used for the artificial synapse. Researches along the direction are now very active worldwide. Those people call their researches "neuromorphic", though we think those researches are nothing but to integrate the analogue memory devices.
The more complicated target is the artificial neuron. It requires the property of Leaky Integrate and Fire (LIF). We believe the "fire" must be done by an external circuit such as FPGA. Some people say that our brain is such an analogue circuit that the neuromorphic circuit should be entirely analogue. However, we do not think so. If a digital circuit can do better, we should utilise it even in neuromorphic circuits. The leaky integration, however, must be done by the analogue device. To realise this in a digital circuit, it becomes quite complicated. A serious problem here is that we need a large capacitor for integrating charges for up to a few seconds.
We found that our SrTiO3 FET can solve the problem! Without the help of the large capacitor, our device can integrate the charge (spikes) for the order of a second. This integration is due to the shift of the threshold voltage of the SrTiO3 FET. We are now fabricating a simple architecture to demonstrate the neuron behaviour.
Since Oct. 2019, I have been the Research Director of a CREST project of JST to study a brand-new biomimetic computation on a spiking neural network using its autonomous spatiotemporal pattern formation (dynamical attractor). My research with the CREST project is reported elsewhere.
SrTiO3 is one of the most studied transition-metal oxides in the history of condensed matter physics. It is a simple band insulator (band gap of ~3.3 eV) but exhibits various unique and interesting properties. SrTiO3 undergoes an antiferrodistortive phase transition at ~105 K due to the staggered rotations of TiO6 octahedra around the [001] axis (see the pictures below). Many studies suggest that this antiferrodistortive transition suppresses the ferroelectric phase transition at relatively high temperatures. However, it remains unclear why the ferroelectricity is suppressed down to very low temperatures (at least 350 mK) despite its phonon structure with polar soft modes remaining. Because of this missing ferroelectricity, a huge static dielectric constant ε ~ 24,000 is observed at low temperatures, resulting in a very large effective Bohr radius of ~0.5 μm. Thus, slight carrier doping of even 2×1016 cm−3 leads to the appearance of an extraordinary dilute metallic state. It is generally believed that the missing ferroelectricity, even at low temperatures, is entirely due to quantum fluctuations: i.e., zero-point motion preventing the complete softening of the transverse optic phonons. The low-temperature phase is positioned close to a quantum critical point (QCP), where different phases compete (such as paraelectric, antiferrodistortive, and ferroelectric states with similar energies). Near the QCP, any residual interactions may drive the system to a superconducting state.
In the early '90s, it was believed that La doped SrTiO3 does not show superconductivity. Only Nb-doped and oxygen-deficient SrTiO3 showed superconductivity after the discovery in 1965. The Nb doping is the Ti substitution, and oxygen deficiency breaks the Ti-O bonds, hence both should affect the transport properties including superconductivity significantly. On the contrary, the La cation substitutes for the A-site cation Sr, thus, it does not affect the Ti-O networks directly.
We thought the La doping is better for superconductivity, why not?
Our trial was indeed successful. We have discovered superconductivity of the La-doped SrTiO3 for the first time in the literature.
In the early '00s, some reports on "quantum critical point (QCP)" shook the world of condensed-matter physicists. Around QCP, where the phase transition happens at zero temperature, quantum fluctuation becomes significant and the ordered ground state (at that time only magnetic ordered state were discussed) is completely suppressed. Then, weird superconductivity appears around QCP.
Pierce Coleman said there is a material event horizon at QCP as an analogue of the event horison in the general relativity theory. Because I was in a Low Temperature Physics lab in Cambridge, which was the centre of the QCP research at that time, I was of course so much fascinated by the physics behind.
Moreover, Mitsuru Itoh's group in Tokyo Instite of Technology reported in 1999 that oxygen isotope exchange of SrTiO3 drives the material ferroelectic!!! This is a great target material to investigate whether the ferroelectric QCP would be a material event horizon. Nobody had investigated it yet.
However, once we started to substitue La for Sr in SrTiO3 to make the sample metallic, no oxygen isotope exchange was successful at all. We raised the temperature up to 2000°C and kept it for one week in the 99% 18O gas atmosphere, no weight change happened. Chemistry is awesome.
This work was thus unsuccessful. Because I had other research topics at that time, this theme was put on hold. No output paper on this topic to list below. It is a shame... however, a big progress comes in the third stage!
We have finally found a way to revenge on the failure of the previous stage. We have indeed succeeded to prepare Sr1-xLaxTiO3 single crystals for many x values and their 18O exchanged single crystals. Tc has boosted up to around 0.6K!!!
Urgent challenge for condensed matter physicists is to develop alternative electronics beyond the present Si ones, which are now confronted a lethal problem called “miniaturisation limit.” In 2020, a typical size of transistors reaches 10nm which contains only about 10 carriers. This will decrease the switching energy smaller than the thermal noise limit (~100kBT), and “on/off” states are no more distinguished. The only saviour of our fully electronics-depended society from the looming crisis not by the traditional artificial way but rather “fundamentally” is to use Mott insulators. At the Mott transition, even in such a nm-scale bulk, more than 100,000 charges exist. They all are localised due to the strong electron correlations, but suddenly delocalised by applying electric field producing an ample amount of itinerant carriers. Thus, the Mott FET is believed to have no worry of the miniaturisation limit.
However, the fabrication of Mott FET has come up against a severe obstacle: essentially, the Mott insulator behaves as an ionic crystal due to the charge localisation, and ionic defects can form quite easily. Hence, the Mott transition is not controlled straightforwardly by simply applying the gate electric field. This renders the Mott FET unsuitable for integration to the present solid-state electronics.
Our mission is to overcome this obstacle and establish a method of the electrostatic carrier-density control of correlated materials for the future Mottronics.
As a springboard of this research, we fabricated Al2O3/SrTiO3 and Al2O3/KTaO3 field-effect transistors (FET), and demonstrated typical n-channel FET characteristics successfully for the first time in the world.
However, the direct deposition of Al2O3 on oxides always produces oxygen deficiencies. Thus, we tried to use an organic gate insulator, Parylene-C, instead of Al2O3. The Parylene-C is inert to any materials and does not damage the channel interface when deposited. We have examined this idea by using SrTiO3 as a channel material. Although SrTiO3 is not the Mott insulator, it is widely used as a substrate for fabricating thin films of Mott insulators. Our FET has demonstrated the field effect mobility as high as 1 cm2/Vs even at room temperature indicating an extremely clean interface, but the carrier density is not so high because of the small dielectric constant of the Parylene-C.
Then, we have have fabricated FETs using an electric double-layer of ionic liquid. By this fancy gating method, we have successfully accumulated carriers in the channel more than 1014 /cm2, and realised Mott transitions! However, ionic liquid is not suitable for practical electronics, moreover, careful control is necessary to avoid electrochemical reaction at the interface of ionic liquid and oxide.
Now this work is suspended but we are planning to step forward soon.
For half a century, metal/oxide/metal sandwich structures have been intensively examined. Especially, it shows numerous interesting properties upon “electroforming”—applying a voltage above a certain critical value to the sandwich to produce a permanent nonvolatile change in its electric properties.These electroformed sandwiches often exhibit a negative differential conductance concomitant with electron emission, electroluminescence, and resistance switching. Typical examples were Al2O3 and SiOx. Around the year of 2000, there is a renewed interest in this area that was prompted by a new generation of experimental and theoretical works: TiOx, NiOx, HfOx, TaOx, CoOx, and so on are widely investigate these days. These phenomena were studied intensively in a bid to put them to practical use, a futuristic non-volatile memory called RRAM. It is widely accepted now that the most important facts for the phenomena are voids, dislocations, defects, and so on, which are, in a word, nonstoichiometry inevitable in every oxide. We did model calculations and suggested the importance of the metal/oxide interfaces. We also did experimental works and pointed out the electric current constricting structure at the interface plays a key role on the resistance switching phenomenon.
We studied the electronic structure of the organic conductor (DMe-DCNQI)2Cu,where DMe-DCNQI is dimethyl-N,N'-dicyanoquinonediimine, by x-ray photoemisson spectroscopy, and clarified that the ratio of Cu+ to Cu2+ is nearly 2 : 1. I also found that Cu is in a valence fluctuating state in the metallic (DMe-DCNQI)2Cu, and a Cu+ - Cu2+ charge ordering coupled with cooperative lattice distortion takes place below the metal-insulator transition temperature in other (DCNQI)2Cu salts.
We performed the ultraviolet photoemission spectroscopy near the Fermi level (EF) of (DMe-DCNQI)2Cu and (Me,Br-DCNQI)2Cu using an undulator radiation. We found a disappearance of sharp Fermi edges in both salts as was observed in other one-dimensional conductors; this is one of the first experimental evidences of the Luttinger liquid.
We showed the first experimental manifestation of how the single-particle density of states around EF changes near a Mott transition in a perovskite-type 3d1 Mott-Hubbard system Ca1-xSrxVO3. As one decreases x, the spectral weight of the coherent band near EF is gradually transferred to that of the precursor of the lower Hubbard band about 2 eV below EF. We considered that the momentum-dependence of the quasi-particle self-energy and hence the long-range exchange or correlation becomes progressively important as one approaches the Mott transition from the metallic side.
We synthesized single crystals of the perovskite-type 3d1 metallic alloy system Ca1-xSrxVO3. The substitution of a Ca ion for a Sr ion reduces the band width W due to a buckling of the V-O-V bond angle without changing the number of electrons making Ca1-xSrxVO3. We found that the Sommerfeld-Wilson's ratio (~2), the Kadowaki-Woods ratio (in the same region as heavy Fermion systems), and a large T-quadruple term in the electric resistivity, even at 300K, substantiate a large electron correlation in this system, though the effective mass, obtained by thermodynamic and magnetic measurements, shows only a systematic but moderate increase, in contrast to the critical enhancement expected from the Brinkmann-Rice picture. we proposed that the metallic properties observed in this system near the Mott transition can be explained by considering the effect of a non-local electron correlation.
We observed a quantum oscillation of a magnetisation (de Haas–van Alphen effect) in the perovskite CaVO3. This is the first observation of the de Haas–van Alphen effect in the perovskite-type transition-metal oxides. Our experimental and calculated Fermi surfaces are in good agreement, but only if we ignore large orthorhombic distortions of the cubic perovskite structure. Subtle discrepancies may shed light on an apparent conflict between the low energy properties of CaVO3, which are those of a simple metal, and high energy probes which reveal strong correlations that place CaVO3 on the verge of a metal-insulator transition.