Neuro-Inspired AI: The Path to Sustainable Intelligence

Dr. Javier Baladrón, an academic and researcher in the Department of Computer Engineering, is leading a Fondecyt Regular project to develop more efficient and sustainable artificial intelligence through neuromorphic computing. This initiative, supported by the Office for Scientific and Technological Research (Dicyt), combines computational simulation, neuroscience principles, and reconfigurable hardware to design algorithms inspired by the human brain. These innovative algorithms will be capable of operating on lightweight, low-energy devices, laying the groundwork for a new generation of AI that mimics the brain's remarkable efficiency.

Imágenes tomográficas.

Today's AI advancements, from ChatGPT to facial recognition, rely on massive, energy-intensive data centers that process millions of calculations per second. This model, however, incurs high costs and significant environmental impact, while the ability to further increase AI's processing power is starting to stagnate.

This presents a critical challenge: to sustain innovation and broaden access to intelligent systems, we must explore new processing paradigms that decouple AI from its reliance on large data centers and high energy consumption. This need is particularly urgent given the current rapid expansion of artificial intelligence into diverse applications.

In response to these challenges, Dr. Javier Baladrón, an academic in Usach's Department of Computer Engineering, is leading the NeuroMetaEvo project. This initiative champions neuromorphic computing, a technology that mimics the brain's energy-efficient operation where neurons activate only when needed, using brief electrical signals. This approach hinges on a core belief: that AI's progress can be significantly enhanced by directly drawing inspiration from how humans learn.

"I always wondered how our brains could solve incredibly complex tasks that machines struggled with," the academic explains. "For me, the path to artificial intelligence always involved understanding the brain's mechanisms so we could replicate them in machines. The human brain, remarkably, uses far less energy than AI systems—it can perform many of our daily tasks on just 20 watts."

To achieve this, the project will develop novel training methods for these brain-inspired networks. Moving beyond traditional AI's complex mathematical calculations, they will employ metaheuristics, techniques that mimic natural processes like biological evolution. This involves the system "learning" by trying various solutions, discarding ineffective ones, and retaining the most successful, without explicit instructions. Additionally, the project will integrate principles of synaptic plasticity, the brain's capacity to adapt its connections based on experience: reinforcing effective pathways and seeking alternatives for unsuccessful ones.

"The challenge lies in finding a form of learning that doesn't rely on precise calculations," he explains. "Instead, it should allow the network to adapt autonomously, just as we learn from experience."

Learning by Iteration: Experiment, Adjust, Validate

The project unfolds across two complementary workstreams. First, the algorithms will be developed and rigorously tested within ANNarchy, a simulation environment. This platform acts as a virtual laboratory, enabling researchers to experiment with, adjust, and validate the learning capabilities of these brain-inspired networks before their physical implementation, eliminating the need for immediate hardware manufacturing. 

Once the algorithms are refined, a second team will focus on implementing them in FPGAs (Field-Programmable Gate Arrays). These reprogrammable electronic chips will enable the creation of physical neuromorphic systems—machines designed to process information much like the human brain. The goal is to integrate the software and hardware, ensuring seamless functionality as it would in a real-world device.

"First, we simulate everything digitally to observe network behavior and make necessary adjustments," Baladrón explains. "Then, we transfer those models into the physical world using reprogrammable chips, essentially building an artificial brain on a circuit. The project's ultimate goal is to integrate both parts, ensuring our algorithms function efficiently within the hardware we design."

One of the key advantages of this technology is its potential use in devices that require AI but cannot incorporate complex systems due to size or power consumption, such as smart watches, drones, or portable sensors.

"We need something lightweight and portable that still boasts significant computing power," he states. "Such applications are perfectly suited for neuromorphic computing."

Beyond its technological aims, the project fosters active collaboration with Chemnitz University of Technology in Germany, a leader in neuromorphic AI. This partnership seeks to bolster the development of this field in Chile through student training, scientific publications, and open-source software. "We want to elevate neuromorphic computing from its current standing to at least the level that present artificial intelligence systems have achieved," Baladrón concludes.

Categoría