Why ‘Neuromorphic’ Chips Are the Key to Making AI Think Like a Human

Neuromorphic computing is still relatively new outside of the research facilities in California and Zurich. No grandiose product introductions. There are no throngs of people waiting outside electronics shops. Just quiet labs with engineers staring at wafer samples and oscilloscopes, attempting to make silicon behave a bit more like biology.

However, researchers are beginning to suspect that something strange is going on. Artificial intelligence has become more potent—sometimes in a startling way—but it is also growing hungry. GPU-filled data centers run nonstop, consuming electricity at a rate that is, to be honest, unsustainable. As this develops, it’s difficult to ignore a paradox: while the human brain silently uses about 20 watts, machines designed to simulate human intelligence need massive amounts of power. about a lightbulb’s energy. Neuromorphic chips are gaining a lot of attention because of this contrast.

Neuromorphic hardware, in contrast to traditional AI processors, aims to replicate the communication between neurons in the brain. These systems use spikes, which are short electrical pulses that move through networks of artificial neurons, to process information rather than continuous streams of numbers. The structure resembles the messy dynamics of living brains rather than neat software models, and it is more akin to biology than mathematics. This distinction may prove to be more significant than most people think.

Deep neural networks powered by graphics processors are a major component of traditional AI systems. Because GPUs are so good at processing large numbers of calculations at once, they have become the foundation of contemporary machine learning. However, intelligence was not their original purpose. They were designed to display textures, shadows, and explosions on a screen in video games. The outcome is effective. But it’s uncomfortable.

Data travels thousands of times per second on tiny electrical highways between memory and the processor in many modern AI systems. This slowdown is referred to by engineers as the “von Neumann bottleneck.” The machine moves data for nearly as long as it considers it.

Neuromorphic chips adopt an alternative approach. Computation and memory are often housed together on the same piece of hardware. It is similar to how biological neurons process and store information in one location, making connections stronger or weaker as signals pass through the network. That change might seem subtle. However, it modifies the physics of computing.

IBM Research researchers have been testing chips that use phase-change memory devices to store AI model weights. To put it simply, when electricity flows through these materials, they physically alter their structure, changing the strength of the signals. A slight change in the substance may indicate the strengthening of a synapse, which is similar to the formation of a memory.

When these chips are demonstrated, it gives the odd impression that the computer is acting more like a nervous system than a calculator.

When it comes to energy efficiency, things start to look particularly intriguing. Certain tasks, such as handwritten number recognition, can operate hundreds or even thousands of times more efficiently on Intel’s neuromorphic processors than on conventional AI hardware, according to experiments. Such numbers seem to catch the attention of investors.

However, there is still a sense of hesitancy. For many years, neuromorphic computing has only existed in lab prototypes and scholarly publications. Electrical engineer Carver Mead started experimenting with circuits that behaved like neurons in the 1990s, which is when the field got its start. Since then, scientists have developed ever-more-advanced prototypes, but there is still little commercial use on a large scale. It seems as though technology is still looking for its pivotal moment.

Software is a part of the problem. Today’s AI is powered by algorithms that were created for traditional computers. Neuromorphic systems employ spiking neural networks, which exhibit distinct behavior. Researchers are still figuring out how to rethink decades of machine-learning techniques in order to train them.

In the meantime, conventional AI continues to face challenges in the real world. Autonomous vehicles provide an excellent illustration. They are amazing devices, but sometimes they overlook clear dangers. A deer on the road. A cyclist stepped out of the shadows. These instances highlight the fragility of pattern-recognition systems.

Because of their event-driven behavior and adaptive learning, neuromorphic architectures may be better able to deal with unpredictability. Or so scientists hope.

The edge computing angle is another. Imagine information being processed locally by thousands of tiny devices, such as wearable monitors, sensors, and drones, rather than being sent back to remote cloud servers. Power efficiency is more important for those systems than raw processing speed. Compared to a chip that acts like a data center, one that acts like a brain might be far more useful.

The conversations in the hallways of AI conferences these days seem different. Naturally, GPUs continue to rule the stage. But in between the espresso machines and posters, scientists are once again discussing neurons. Not precisely biological ones. However, it was close.

It’s unclear if neuromorphic computing will actually enable machines to “think” like people. Over millions of years, the brain underwent evolution. Only a few decades have passed since silicon chips were invented. That is a huge disparity.

Nonetheless, a subdued interest is growing throughout the sector. Perhaps artificial intelligence in the future will resemble a nervous system etched into silicon rather than software, since intelligence in nature is based on tiny electrical spikes moving through networks of neurons.