Can AI Work More Like the Human Brain?

Twin Things Podcast #11

Can computers learn to process information more like the human brain? In episode 11 of Twin Things Podcast, Jan Argasiński from Sano joins Katarzyna Baliga-Nicholson to explore neuromorphic computing, spiking neurons and the potential for building more energy-efficient AI systems inspired by the nervous system.

Jan Argasiński had a shake for lunch, and his brain can run on it for a very long time. The human brain needs only about 20 watts, roughly the power of a small light bulb, to see, remember, learn and make decisions. Today’s AI models need entire data centres to do far less. This contrast opens episode 11 of Twin Things Podcast. Jan leads the Computational Neuroscience team at Sano and works at the Department of Human-Centred Artificial Intelligence at Jagiellonian University. In conversation with Katarzyna Baliga-Nicholson, he explores whether a fundamentally different approach to computing could bring machines closer to the way the brain works.

A different path to computing

Ever since the first computers were developed, researchers have wondered whether machines could one day think more like humans. Periods of great enthusiasm for artificial intelligence have alternated with so-called AI winters, and today the field is experiencing another remarkable surge. Throughout this history, nature has remained an important source of inspiration.

In 1990, American engineer Carver Mead proposed an idea that was ahead of its time. Rather than simply making conventional processors faster and more powerful, he suggested studying how biological systems process information and using those principles to build electronic circuits. He called this approach neuromorphic engineering.

As Jan explains, the idea initially struggled to gain momentum. Conventional computer chips continued to become smaller, faster and more powerful, leaving little incentive to explore alternative architectures. Today, however, energy consumption is becoming a major concern. The enormous computing demands of large AI models are prompting technology companies to look for new energy sources and more efficient ways to power their infrastructure.

Neuromorphic computing offers a different direction. As Jan emphasises, it is not simply the next generation of AI, but a separate, parallel computing paradigm. Rather than replacing conventional computers, it could complement them, with the two approaches working together in hybrid systems.

When hardware does the thinking

In conventional computers, memory and processing are physically separated. Data constantly travels between the two, while a clock coordinates each operation. Moving information back and forth takes both time and energy. Jan compares this arrangement to having to retrieve information from a USB stick every time you want to think about something.

The brain operates differently. Neurons receive signals from other neurons and generate brief electrical pulses, known as spikes, when their activity reaches a certain threshold. Connections between neurons can become stronger through repeated activity. Neuroscientists often summarise this principle with the phrase: “What fires together, wires together.”

In the brain, information processing and memory are closely intertwined. Signals are transmitted through networks of interconnected neurons, and activity is distributed across the pathways involved in a particular task.

Neuromorphic systems aim to reproduce some of these principles in hardware. Instead of treating processing and memory as separate functions, they bring them closer together and use event-driven computation, in which activity is triggered by incoming signals rather than by a continuous sequence of clocked operations.

“In neuromorphic systems, the hardware is the algorithm,” Jan explains.

Several neuromorphic computing platforms have already been developed, including IBM’s TrueNorth, Intel’s Loihi and SpiNNaker, a system developed at the University of Manchester. Some neuromorphic approaches also use memristors – electronic components whose resistance changes depending on the electrical current that has passed through them. In this way, they can retain information about their previous states.

Jan compares this behaviour to a path through fresh snow. Once someone has walked along a particular route, it becomes easier for the next person to follow. Paths that are rarely used gradually disappear. In a similar way, memristive devices can retain traces of previous electrical activity.

From drones to medical devices

Neuromorphic computing could be particularly useful for processing information that changes over time, such as visual and audio signals.

Consider a drone navigating through a forest. A conventional vision system might analyse a complete image many times per second, even when very little in the scene has changed. A neuromorphic system, by contrast, can focus on detecting changes and responding to new events, much as our visual system does. When there is little activity, it can consume very little energy.

This event-driven approach could make it possible to build compact, energy-efficient devices capable of processing information directly where it is collected, without constantly sending data to a central computer or the cloud.

Potential applications range from sensors in robots and autonomous vehicles to wearable medical devices. One example is an EEG-based wearable that could help detect signs of an approaching epileptic seizure. Neuromorphic systems could also support autonomous vehicles that need to adapt continuously to their surroundings, potentially making real-time learning more practical in situations where conventional approaches are computationally demanding.

Learning from biology again

Jan points out that the term neuromorphic has gradually become an umbrella covering a wide range of technologies and approaches. Some researchers now argue that the field needs a kind of soft reset: a renewed focus on what neuroscience can teach us about the brain.

After all, the brain is much more than a network of neurons and synapses. It also relies on chemical neuromodulators, diverse types of neurons and highly complex architectures. Our understanding of these mechanisms has advanced considerably since the early days of neuromorphic engineering in the 1990s.

Bringing more of this biological complexity into computing could open up new possibilities, including approaches that move beyond conventional binary representations of information towards analogue computation.

The relationship between neuroscience and neuromorphic computing works both ways. Neuromorphic models of particular brain structures can also serve as experimental platforms. Researchers can manipulate or switch off selected components and observe the effects, in ways that can help them investigate brain function, damage and neurodegenerative processes.

Who gets the final say?

The conversation also turns to questions about the future of human-AI interaction.

Among the rapid-fire questions posed during the episode are several generated by Katarzyna’s AI assistant. One imagines a future in which, by 2080, a smartwatch equipped with a neuromorphic chip has learned so much about its owner’s personality that it starts disagreeing with their decisions.

Jan believes that people should always have the final say. At the same time, he admits that he would welcome an internal voice capable of warning him when he is about to make a poor choice.

The question becomes more complicated when the technology moves from a smartwatch to a chip implanted directly in the brain. If its suggestions become indistinguishable from our own thoughts, how can we tell where our decisions really come from?

Jan notes that the issue is not entirely hypothetical. Some medicines already influence how we feel, think and make decisions. In that sense, the boundaries between external intervention and our own mental processes may be less clear than we think.

Asked whether the future of AI lies in teaching machines to take a good nap, Jan agrees with the underlying idea: “The point is not to do more, but to do better.”

What comes next?

Looking ahead, Jan wants to bring more insights from neurobiology into neuromorphic platforms, making them more biologically plausible and useful for scientific experiments. He is also exploring how these principles could be applied beyond neuroscience, including in systems designed to support medical decision-making.

His team is involved in EBRAINS, the European research infrastructure that grew out of the Human Brain Project and supports collaborative research into the brain. The initiative is now preparing its roadmap for the next ten years.

The team has also recently submitted a grant proposal for a neuromorphic benchmark based on data provided by neurobiologists. The goal is to help evaluate and advance neuromorphic approaches using biologically grounded evidence.

Neuromorphic computing is unlikely to replace conventional computing altogether. Its promise lies in offering a different way to process information – one that draws on the remarkable efficiency and adaptability of biological systems. And as AI continues to demand more computing power, learning from the brain may become an increasingly important part of the search for more sustainable technologies.

Watch the full episode on the Sano Science YouTube channel: https://www.youtube.com/watch?v=tLJjDx79JnA

DIFFICULT WORDS

• Neuromorphic computing / neuromorphic engineering — computers and chips designed to imitate the way the nervous system processes information. The term was coined by Carver Mead around 1990 and now covers hardware, learning methods and spiking neural networks.
• Spiking neural network (SNN) — an artificial neural network in which neurons exchange short pulses (spikes) at specific moments in time instead of continuous numbers.
• Spike (action potential) — a short electrical pulse that a neuron sends when the signals it receives cross a threshold.
• Synapse — the connection between two neurons, where the signal passes from one to the other.
• Plasticity / STDP (spike-timing-dependent plasticity) — the ability of connections between neurons to strengthen or weaken. In STDP, the change depends on the exact timing of spikes (“what fires together, wires together”).
• Von Neumann architecture — the classical computer design in which the processor and memory are separate and data travels between them. The cost of this traffic is called the von Neumann bottleneck.
• Memristor — an electronic component whose resistance depends on the current that has passed through it, so it keeps a “memory” of its history. Used to imitate synapses.
• TrueNorth, Loihi, SpiNNaker — examples of neuromorphic chips and systems, developed by IBM, Intel and the University of Manchester (SpiNNaker within the Human Brain Project).
• Backpropagation — the standard method for training artificial neural networks. The error of the result is sent back through the network to adjust its connections. Powerful, but computationally and energetically expensive.
• Surrogate gradient learning — a method that allows spiking neural networks to be trained in a way similar to backpropagation by replacing the “all-or-nothing” spike with a smooth approximation.
• Event-based (event-driven) processing — computation that happens only when something changes, e.g. event cameras that record changes in brightness instead of full frames.
• Edge computing — processing data directly on the device (a sensor, a camera, a wearable) instead of sending it to a remote server.
• EEG (electroencephalography) — recording the electrical activity of the brain with electrodes placed on the scalp.
• Neuromodulators / three-factor learning — chemical substances in the brain (e.g. dopamine) that regulate how neurons fire and learn. In three-factor learning, a neuromodulator-like signal is added to the two neurons involved in plasticity.
• Analogue computation — computing with continuous physical quantities (e.g. current) instead of zeros and ones.
• Biologically plausible — consistent with how real biological systems work.
• Lesion — damage to part of the brain. In models, a part of the network can be switched off on purpose to study its role.
• EBRAINS — a European digital research infrastructure created as a legacy of the Human Brain Project (EU flagship, 2013–2023). It offers brain data, modelling and simulation tools and access to neuromorphic computers.