The brain-like memory chip that will never be forgotten: New memos can solve AI amnesia problem

Scientists have developed a new type of electronic component that can help artificial intelligence systems retain old knowledge while learning new skills, a fundamental challenge that limits the development of more powerful AI. This innovation details a study published this month Natural Communicationsrepresenting a significant advance in neuromorphic computing, aims to mimic the brain’s efficiency and learning ability.

Unlike traditional computer chips that process information sequentially, these specialized components (called memoirs) can process and store information simultaneously, similar to the role of neurons and synapses in the human brain. What makes new devices particularly promising is their ability to overcome ongoing problems in artificial intelligence known as “catastrophic forgetting.”

“Its unique properties allow the modulation of the memo to be controlled using different switching modes, so that the stored information is lost,” explained Ilia Valov of Peterson Glenberg College, who led the research team.

When artificial neural networks learn new tasks, they usually overwrite previously acquired knowledge, such as erasing old recordings to make room for new recordings. This phenomenon, that is, AI systems suddenly forget the information they have learned before, has been an important obstacle to the development of a wider range of artificial intelligence.

Humans usually don’t encounter this problem. Our brains can not completely forget old skills, not completely forget old skills. Neuroscientists call this property “chemistry” – basically, the brain’s ability to regulate how many specific neural pathways it may change.

The new memo aims to replicate this capability in electronic form, potentially allowing AI systems to accumulate knowledge over time without suffering from digital amnesia.

New mechanisms starting from the bottom of the

What distinguishes these new components from existing memos is their unique operating mechanism. “We found an electrochemical recall mechanism that was fundamentally starting with, which is more chemically stable,” Valov said.

Traditional memorandums operate through one of two mechanisms. The electrochemical metallization (ECM) conceptualizer forms a wire between the electrodes that can be dissolved when the voltage is reversed. Valence Change Mechanism (VCM) Memorandum Modifies the interface between the electrode and the electrolyte by oxygen movement.

The newly developed components adopt what researchers call the “fiber conductivity modification mechanism” (FCM). Instead of producing purely metal wires, they form stable metal oxide wires that never completely dissolve but are chemically modified.

“Our new memoir is based on a completely different principle: it uses filaments made from metal oxides, rather than pure metals like ECM,” Valov explained. “You can think of it as filaments that always exist to some extent, with only chemical modifications.”

Once digital and analog once

Perhaps the most important advantage of the new memos is their ability to operate in both binary (digital) and analog modes. This dual function simulates complex behaviors of biological synapses more accurately than previous electronic components.

The team has verified their method through computer simulations, implementing components in an artificial neural network model that enables high-precision pattern recognition in several image datasets.

The practical benefit is not just to solve the problem of forgetting. Compared to their predecessors, these memos show higher performance characteristics: they function over a wider voltage range, withstand higher temperatures, require lower operating voltages, and exhibit greater overall reliability and life.

This increased durability solves ongoing challenges in Memristor development – ​​despite theoretical advantages, higher failure rates and limited lifespan also slow down commercial adoption.

From research to reality

The journey from laboratory prototypes to commercial products remains faced with obstacles. While the new mechanism is an important step forward, researchers acknowledge that there is more work to optimize the materials and manufacturing process.

“So basic research is crucial to better control of processes at the nanoscale,” noted Valov, who has extensive experience in Memristor research. “We need new materials and switching mechanisms to reduce system complexity and increase functional scope.”

Looking into the future, Valov and his team plan to explore other materials that might play a more effective role in their memoirs. “Our results will further advance the development of electronic devices for “memory computing” applications,” Valov said confidently.

Potential applications are much more than simply creating more efficient computers. These brain-inspired components can enable entirely new computing paradigms for learning, memory and processing combined with traditional electronic devices, thus transforming the field from autonomous vehicles to scientific research and healthcare.

With its unique learning abilities without forgetting, these new memoirs may help usher in the AI ​​system that has become increasingly intelligent over time without losing previous knowledge, just like the human mind they aim to imitate.