Scientists found an astonishing new material that behaves like nothing we’ve ever seen
...
Scientists have discovered a shocking fact about a material called Vanadium dioxide (VO2). According to research published in
Nature Electronics, VO2 can remember previous external stimuli. It’s an interesting discovery, and one that researchers say could completely change the future of computer and storage devices.
This isn’t the first time we’ve seen a material bring new properties to light, or the first time we’ve heard about researchers finding new uses for older materials. Back in 2014, researchers and engineers began looking at
graphene as a way to make smartphones even thinner. Perhaps Vanadium dioxide would provide a similar evolution for computational storage devices.
...
Memristors are not new. They've been around since the 60's. Titanium dioxide is the most popular material, but it's been done with others, even silicon dioxide.
Sheets of boron nitride look promising as memristive devices. The thing is, you need to control these devices in specific ways, to make them useful computationally. One of the basic requirements is phase coding, you need to be able to send about 90 degrees worth of a traveling wave across the spatial extent, "while" the devices are interacting with each other. So, you need to embed the memristors in a "command and control lattice", without interfering with their other ongoing internal interactions.
Memristors are "neuron-like", and networks of memristors are "brain-like". Generally they implement a "Hebbian learning rule", which means connections get reinforced on the basis of covariance (or coincidence in discrete time models). Self organizing memristor networks display the full spectrum of Kuramoto dynamics, and they tend to organize themselves into MISO-type clusters of Laguerre-filtered Volterra engines.
However this is not the whole story.
You can calculate the time it takes for an E/M signal to get from one side of the brain to the other. (Not synaptically, but rather, by volume conduction - when a nerve fires it generates a significant E/M field which is so strong it's picked up on the surface of the scalp, we're talking reversal of a trans-membrane electric field of about 10 million V/m, yes that's not a typo).
The brain is about 10 cm long, from frontal to occipital, so you end up with a fraction of a nanosecond travel time, for an EM signal at the speed of light. The duration of that event matches the duration of an action potential, which is about 1 msec.
Within that interval, you have about 80 billion neurons engaged in time-coding (just about the entirety of the cerebellum, which contains 3/4 of all the neurons in the brain), and "many" experiments have shown that the conscious brain can entrain and control the firing of a single neuron. After a little quickie math at 10k synapses per, you can calculate that the maximal temporal resolution of the human brain is on the order of femtoseconds. FEMTO-seconds being 10^-15, which is comparable to chemical reaction times or the fastest lasers.
The interesting thing about the Volterra dynamics is cellular assemblies can make the information appear somewhere (like, at the opposite end of the brain) and SEEM LIKE it gets there faster than the speed of light. What happens then is, this sets up an "expectation field" that interacts with paramagnetic ions and membrane proteins through J-coupling.
This process is not like a Kalman filter, although there are similarities in concept. The trick is we're querying the distribution in real time, "while" it's being updated, and the same neurons doing the querying are the same ones doing the updating. (These activities are "distributed" throughout the network, and their EM signatures interact holistically).
If you try this with memristors, you run immediately into the daunting technical problem that memristors don't have loads of paramagnetic ions hanging around. They can't be interrogated and controlled this way, it has to be done differently. So memristors have seen success in LSTM architectures where the control elements can be built right into the chip. And less success in the Volterra engine architectures that try to extract causality from millions of simultaneous time series