Whether a computer could ever pass for living things remains one of the key challenges for researchers in artificial intelligence. There has been a lot of progress in AI since Alan Turing first created the Turing test—whether a machine can exhibit intelligent behavior equal to or different from a human.
However, machines still struggle with one of second nature, a skill fundamental to humans and other life forms: lifelong learning. Learning and adapting while performing tasks without forgetting past tasks or instinctively transferring knowledge gathered from one task to another.
Now, with support from the DARPA Lifelong Learning Machine (L2M) program, researchers at USC Viterbi have collaborated with colleagues from institutions in the United States and around the world on a new resource for the future of AI learning, defining how How artificial systems can successfully think, act and adapt in the real world, just as living beings do.
Co-authored by Alice Parker, Dean’s Professor of Electrical and Computer Engineering and Professor of Biomedical Engineering and Biokinesiology and Physical Therapy Francisco Valero-Cuevas and their research teams, was published in nature machine intelligenceIn collaboration with Professor Dhirsha Kudithipudi at the University of Texas at San Antonio, along with 22 other universities.
The research team also included Ali Marzinejad, a postdoctoral researcher at ValeroLab, Dario Urbina, a biomedical engineering PhD candidate, and Suraj Chakravarti Raja, an electrical engineering PhD candidate. Valerolab’s previous work in the bio-inspired AI field involved an AI-controlled robotic limb operated by tendons similar to animals that can learn to walk without prior knowledge.
Valero-Cuevas said the purpose of this publication was to provide a resource for researchers in AI for decades to come, focusing on the concept of lifelong learning in artificial systems. The blueprint will outline the development of next-generation technology such as self-driving cars, autonomous robots and drones, and a new type of machine intelligence for intelligent prostheses, exoskeletons and wearable devices.
Lifelong learning is a collection of biologically inspired abilities that we take for granted, such as acquiring new skills without compromising our old skills, adapting to changes, and applying previously learned knowledge to new tasks.
“The way you and I learn is through experience, imitation, and self-exploration, and just because you learn a new task doesn’t mean you forget prior tasks,” Valero-Cuevas he said. “Humans learn on the job. We show up and start learning, and then we know how to do it. Machines aren’t there yet.”
Valero-Cuevas said that machines can be programmed through algorithms. For example, a self-driving car might use algorithms that are fed data from other vehicles to learn how to operate it before being deployed to the world.
“But there is a clear distinction between training and deployment. When a machine is deployed, it is not necessarily learning. If you want to teach it something new, you generally have to clear memory, and It forgets how to do the last thing,” said Valero-Cuevas.
The issue, known as “catastrophic forgetting,” is one of the significant shortcomings of current AI systems that the new blueprint highlights.
“As a human, you can show me how to play ping pong, and then I can use that skill and transfer it to learning to play tennis or badminton,” Valero-Cuevas said. “At the moment, a machine can only be taught to play a racquet sport, such as ping pong. If you want to play tennis that requires you to use your body differently, you have to change the program.”
As well as overcoming the catastrophic forgetting factor, the researchers’ blueprints outlined a range of other competencies needed for artificial systems that can achieve lifelong learning in the same way that living things can. This includes:
- Transfer and Adaptation – Ability to transfer and reuse knowledge and adapt to new environments.
- They exploit task similarity—the ability to learn multiple tasks and transfer that knowledge between them.
- Task-agnostic learning – understanding when to switch from one task to another during training without being told.
- Noise Tolerance – Completion of the same task in a real-world scenario when conditions differ by training duration (such as operating during the night instead of the day).
- Resource efficiency and sustainability – ensuring that lifelong learning does not negatively impact storage capacity and slow down its response time.
An expert in biologically inspired circuits, Parker has a long research history looking at the human brain to understand how it works and how this knowledge can be applied to artificial systems. For this latest paper, Parker contributed to the understanding of neuromodulation: systems in the brains of living things that help improve learning, overcome catastrophic forgetting, adapt to uncertain environments, and understand changes in context .
Parker focuses on building electronic structures that support neuromorphic circuits (electronics that model the brain) by adding biological features not found in existing neural network hardware, but that support lifelong learning. Features include astrocytes—other brain cells that interact with neurons to support learning—and neural codes—spikes and bursts that convey information in a coded form.
Valero-Cuevas said the collaboration was a comprehensive resource for future research into machine learning and AI.
“Biology is motivation. The ultimate example of lifelong learning is a biological system,” Valero-Cuevas said. “What we are doing is looking at all the things that biological systems can do, of which we are extremely envious, and asking what the machines would have to do to do it.”
“We think that at least some of the mechanisms from biology — lessons learned over millions of years of evolution — can be translated and adopted by AI,” he said.