Memristors: The Neural Bridge between Silicon and Biology
Most electrical engineers can easily list the fundamental circuit components: the resistor, capacitor, and inductor. Each of these components relate different circuit quantities to each other non-linearly; resistors relate voltage and current, capacitors relate voltage and charge and inductors relate flux and current. Missing from this list is a non-linear relationship between flux and charge. The final piece of this puzzle is called a memristor, the circuit component linking flux and charge.
First theorized by Leon Chua in 1971 memristors are finally being implemented, and their unique properties are of great interest to neurotechnologists.
The most interesting property of memristors is aptly named memristance. Memristance involves the dual utilization of resistance, which is the ability to impede the flow of electricity. Memristors have programmable resistance, meaning the level of resistance changes based on the direction of the current. If the current flows one way through the memristor its resistance increases, and if it flows the other way the resistance decreases. The memristor is also non-volatile –it remembers its current state even when there is no power. For example, a memristor removed from the circuit will remember its’ individual resistance even in absence of the circuit.
Modern invasive BCI technologies use microelectrodes and ECoG Arrays to record brain activity. In many cases, spike data from neurons are recorded and then translated by a machine learning algorithm into actionable insight. This approach assumes it is possible to understand the brain with non-neuromorphic machine learning. Most machine learning algorithms don’t have much relation to the way humans learn, because it is hard to implement biologically inspired learning rules on traditional hardware. It is not yet known if simply looking at neural data and trying to understand it without capturing the stochasticity, non-linearity, and changing nature of the human brain will be fruitful.
In the case that it is not, neuromorphic BCIs are a possible solution. Neuromorphic BCIs use algorithms on Spiking Neural Networks(SNNs), machine learning models that simulate neurons. Chips that can efficiently run SNNs exist which you can read more about in my other article here. However, one of the shortcomings for SNNs that run on chips is that they take some time to converge onto an output. In order for quicker response times, SNNs should be run as a network of physical artificial neurons. This way, spike data can be taken directly from neurons and passed into the artificial neurons. These artificial neurons can then function just like regular neurons, passing signals back and forth through their synapses. In an ideal network, the latency would be around the same as for biological neurons.
But in order to connect these artificial neurons with each other, and with biology, we also need artificial synapses. These artificial synapses need to follow biological learning rules, like Spike Timing Dependent Plasticity– neurons that fire together, wire together. This is where memristors come in. Memristors can be easily programmed to have both long and short-term plasticity. Like synapses, they remember their state and can adjust based on input. This means memristors allow for the wetware of the human brain and hardware of things like prosthetic limbs and computers to be connected directly at the most fundamental level.
In 2020, researchers at the University of Southampton, TU Germany-Dresden and the University of Padova used memristors to connect artificial neurons to biological ones. An artificial neuron was used to provide input stimuli to a biological rat neuron, which was connected in series with another artificial neuron. The resistive state of the memristor synapses between the neurons was used as the “synaptic weight” –the strength of the connection between the neurons. The input artificial neuron was able to influence the output artificial neuron using the biological neuron as an intermediary, and both memristive synapses showed potentiation and depotentiation at adequate types of input stimuli.
Of all the potential applications for this technology, the most exciting lies in the field of neuroprosthetics. Instead of needing machine learning to translate biological motor neuron input into action, the biological motor neurons could be directly connected to artificial motor neurons. This would allow for very low-latency, low-power prostheses, which could be the basis for a seamless patient experience.
Memristors, therefore, function as artificial synapses, and they open up a new world of connection between man and machine.
