九九福利网

This material, which switches states depending on the wavelength of light shone on it, could have applications in computers that mimic the human brain.

Traditional computer hardware architecture is limiting AI computing

Some tech enthusiasts and entrepreneurs claim we鈥檝e entered the era of technological singularity, the point at which artificial intelligence surpasses human intelligence and control. Or at least, they say, .

For others, we still have a long way to go, given how computers 鈥渢hink鈥� very differently from the way the human brain does.

One thing that makes computers and brains very different is the architecture that supports their electrical circuits and enables learning and memory. Within the processor or 鈥渂rain鈥� of most computers, memory (storage) and computing (learning) units are and connected by a 鈥渂us.鈥� The memory unit stores data and instructions.

Passing large amounts of data back and forth between memory and compute units on a chip (or between chips) consumes an incredible amount of energy. This so-called 鈥渕emory bottleneck鈥� has become a central challenge in AI computing and is a key contributor to the growing energy demands of large-scale data centers. 

Yet highly efficient, low-energy artificial intelligence is theoretically possible, if we can learn enough from the most efficient 鈥渃omputer鈥� in nature: the human brain. 

In the brain, learning and processing are integrated 

鈥淚n the brain, processing and memory reside in the same system, in the same hardware. Data doesn鈥檛 have to flow between two different places for learning to happen,鈥� said V铆ctor Garc铆a-L贸pez, an assistant professor in the Department of Chemistry in the LSU College of Science. Garc铆a-L贸pez explains that learning and memory are built into how neurons communicate with each other. 

Neurons talk to one another through electrical 鈥渟ignals鈥� that travel along neurons and between them, across gaps called synapses. Each electrical wave triggers the opening of ion channels鈥攑roteins located within the lipid membranes that form the 鈥渨alls鈥� of each neuron. 

When an electrical wave arrives at a synapse, ion channels open and calcium ions flow in. These ions activate proteins that trigger neurotransmitters to exit the neuron鈥檚 walls and travel to neighboring neurons, passing on the electrical signal. 

Here鈥檚 the cool part: The more often a signal travels along a particular neuronal pathway, the stronger the connection between those neurons becomes. Repeated activity leads to a buildup of proteins that trigger neurotransmitter release, making signals easier to send and more effective at activating the next cell. 

This is how learning happens, built directly into the way signals move through the brain. 

鈥淭hat's why they say, if you want to learn something, you have to repeat it,鈥� Garc铆a-L贸pez said. 鈥淵ou are training your synapses, but at the very molecular level, the repeated activity of ions crossing across the neuronal membrane is triggering stronger signals.

鈥淭he question is, can we develop computers that work more like the human brain and model how neurons communicate?鈥� 

Many experts are pursuing this very idea in the field of neuromorphic computing. But how do you create artificial neuronal pathways that can 鈥渓earn鈥� based on repetition and reinforcement? 

This requires devices known as memristors (which allow current to flow) and memcapacitors (which store charge) that can 鈥榬emember鈥� past signals while actively processing new ones, bringing computers a step closer to how the brain works. 

The problem is that most experts are still trying to create these devices with solid-state materials. 

鈥淚f we want to mimic the brain, why don鈥檛 we use materials that are in the brain?鈥� Garc铆a-L贸pez said. 

The key to neuromorphic computing might be a lipid membrane鈥� and a light-controlled molecular switch 

Garc铆a-L贸pez has been working with Dr. Charles Collier and Dr. John Katsaras at Oak Ridge National Laboratory to design soft materials built from biological components, such as lipids, that can 鈥渞emember鈥� past electrical activity. These materials exhibit memory-like electrical behaviors. 

In their newly published study, the researchers describe the creation of bioinspired materials that can be engineered to perform neuromorphic functions, which could be implemented in the future for sensing and brain-like computation. 

 

Over the past few years, researchers at Oak Ridge have created double lipid layers, or 鈥渂ilayers,鈥� that resemble the 鈥渨alls鈥� around neurons. They鈥檝e shown that one type of lipid bilayer can act as a memcapacitor, allowing ions to accumulate on either side, and another type as a memristor, allowing ions to flow through ion channels. 

Garc铆a-L贸pez was giving a talk at Oak Ridge when he met Charles Collier. Garc铆a-L贸pez had a chemical tool from his lab鈥檚 previous research at LSU that could switch a lipid bilayer between permeable and non-permeable states. 

As they talked, it quickly became clear that combining this molecule with lipid bilayers at Oak Ridge could open exciting new possibilities for a soft material that could switch between resistive and capacitive states. 

Garc铆a-L贸pez鈥檚 lab has created a unique molecule, somewhat like a tiny molecular machine, that changes its configuration in response to light. The molecular machine is called a rotaxane and has a ring-on-a-rod structure, in which a cyclic molecule is threaded onto an axle. When exposed to different wavelengths of light, the ring component changes shape. 

鈥淢y group conceived and developed this new class of molecules designed to modulate lipid membrane structure. We synthesized these compounds and how they interact with and reshape lipid bilayers in vesicular systems,鈥� Garc铆a-L贸pez said. 

The rotaxane can exist in two states. In one state, when embedded in a lipid membrane, the rotaxane alters membrane packing, creating holes that allow ions to pass through. When exposed to voltage, the membrane behaves as a memristor, a resistor with memory.

But when exposed to a particular wavelength of light, the rotaxanes change shape, inducing tight lipid packing and preventing ions from crossing the bilayer. Charges accumulate, and the system behaves as a memcapacitor. 

鈥淭he same membrane can be switched between these two neuromorphic behaviors simply by illuminating it with light, enabling dynamic control over the device function,鈥� Garc铆a-L贸pez said.

鈥淲e can reversibly go between two different electrical behaviors without having two different systems.鈥� 

Seeing the switchable behavior of these membranes is quite remarkable. In one video, Garc铆a-L贸pez demonstrates a vesicle鈥攁 circular enclosed structure with rotaxanes embedded in the lipid membrane鈥攖hat shrinks or expands when light is shone on it. 

In the newly published paper, the researchers demonstrated changes in electrical current across a switchable rotaxane-studded membrane. 

Biological membranes containing rotaxanes have many applications, thanks to the ability to control their thickness and permeability. But the ability to create soft resistor-capacitor devices is quite novel. 

鈥淲e are really the first ones to create an artificial or molecular switch for soft neuromorphic devices,鈥� Garc铆a-L贸pez said. 鈥淭his doesn鈥檛 solve all the issues needed to make soft material neuromorphic computing possible, but it is an important step.鈥� 

A switchable memristor-memcapacitor can save materials, space, and energy while improving performance and precision in an electrical circuit. Biomimetic membranes that can switch between memristor and memcapacitor modes open up many new opportunities in neuromorphic computing, reducing the complexity of the required architecture. 

鈥淥ur Oak Ridge collaborators are really pioneers in creating membrane-based neuromorphic materials, but in most cases, separate membranes are needed, or different voltage conditions for each functional state,鈥� Garc铆a-L贸pez said. 

鈥淏y integrating our light-responsive molecules with their platform, we can use light to reversibly switch between neuromorphic states in a single membrane.鈥�

This research was made possible through a collaboration with scientists at Oak Ridge National Laboratory and partial support from the LSU Provost Big Idea Grant. 

Read the study: