Reconfigurable AI Device Shows Brain-Like Promise
Artificial intelligence will never be able to replace the human brain. The fact that the human brain adapts when it learns something new is proof of this.
Researchers at Purdue University have developed computer chips that can constantly reconfigure themselves to receive new data, much like the brain works. This allows AI to keep learning over time. The circuits on a computer chip do not change, unlike the brain, which constantly forms new synaptic connections to facilitate learning.
The circuit a machine used for years is identical to the one created for a factory engine. The goal is to create a platform on which machines will learn throughout their lives.
There are stumbling blocks to making AI more portable, for example for autonomous vehicles or space robots that must make decisions on their own in remote locations. These machines would be more efficient if AI could be built directly into the hardware rather than into the software, as is done now.
According to the new study, the developed adaptable device can transform into all the key electrical components needed for artificial intelligence hardware, for potential use in robotics and autonomous systems.
Brain-inspired or "neuromorphic" computer hardware aims to mimic the human brain's remarkable ability to adaptively learn from experience and process information quickly in an extremely energy-efficient way.
These features of the brain are largely due to its plastic nature, its ability to change its structure and function over time through activities such as the formation of neurons, or "neurogenesis."
"We hypothesized that if we could mimic the behavior of neurogenesis in electrical hardware, we could create machines that learn throughout their lives," said senior author of the study Shriram Ramanathan, an electrical engineer and materials scientist at Purdue University in West Lafayette, Indiana.
Ramanathan and his colleagues developed highly adaptable neuromorphic devices using perovskite nickelate. Perovskites are crystals that have driven phenomenal progress in solar cell performance over the past decade.
The scientists incorporated protons into the perovskite nickelate. Electrical pulses applied to this material can move protons within the material's lattice, changing its electronic properties. The researchers could electrically reconfigure a device made from this proton-doped perovskite nickelate into a resistor, a memory capacitor, a neuron or a synapse on demand.
“We can use one device to perform several neuromorphic functions. In addition, we can switch between these functions with simple electrical pulses on the nanosecond scale”
"The versatility of this device can simplify the design of AI circuitry for complex computing tasks, avoiding the clustering of different functional blocks that take up a lot of space and consume a lot of energy," said study co-author Michael Tae Joon Park, an electrical engineer and materials scientist at Purdue. He said potential applications include robotics and autonomous systems.
When simulated using the new device in an artificial neural network that mimics the structure of neurons in the biological brain, the scientists found that the reconfigurable nature of the new device allows the neural network to "make decisions more efficiently compared to conventional static networks in complex and constantly changing conditions."
The new device performed consistently through 1.6 million switching cycles between states. "In addition, hydrogen ions remain in the device for a long period of time after its initial processing, more than six months, which is encouraging," Park said.
The researchers suggest their device could find use in expandable networks, which are neural networks capable of increasing their computing power on demand. Similarly, these networks could shrink in size if the device detects nodes that are regularly inactive, in order to become more efficient.
The scientists note that they fabricated their devices using technologies compatible with semiconductor foundry manufacturing, suggesting they could readily find application in the electronics industry.
“However, our research is in its infancy. Much more work is required to fabricate large-scale integrated test circuits with these devices”
The scientists detailed their findings in the journal Science: "Using our reconfigurable artificial neurons and synapses, simulated dynamic networks outperformed static networks for gradual learning scenarios. The ability to form building blocks of brain-inspired computers on demand opens new directions in adaptive networks."
