91av

Neuroscience seems an unlikely place to find fundamental truths that could apply to everything in the universe. Brains are specific objects that do things that few, if any, other objects in the universe seem capable of. They perceive. They act. They read magazine articles. They are usually the exception, not the rule.

That is perhaps why the free-energy principle (FEP) has garnered so much attention. What began in the early 2000s as a tool to explain cognitive processes like perception and action began to be presented as a “unified brain theory”. Then the FEP outgrew the brain, being put forward as a definition of life and, inevitably, as the basis for a new kind of artificial intelligence that can reason. Today, some proponents argue that the FEP even encapsulates what it means for something in the universe to exist at all. “You can read the free-energy principle as a physics of self-organisation,” says its originator, at University College London. “It is a description of things that persist.”

Yet some researchers are sceptical that the FEP can live up to many of its loftiest promises, having grown frustrated with its shifting scope. “It has been a moving target,” says , philosopher and cognitive scientist at Tilburg University, the Netherlands.

All of which has made the FEP a source of both fascination and frustration. Its dizzying breadth is key to its enduring appeal, even while it remains famously difficult to get your head around. So, given the claims that it can be used to explain everything in one go, does the FEP really explain anything at all?

Friston, a psychiatrist by training, is by many accounts one of the most influential neuroscientists alive. In 1990, he developed a computational technique called statistical parametric mapping that allows researchers to meaningfully compare images of different brains to each other, despite their varied shapes and sizes. This alone would have earned Friston a mention in the scientific history books. But he went on to develop other tools that made the brain ever more transparent for our digital eyes.

However, for all that neuroscientists were learning about the brain in the 1990s, they were left with a stubborn, slippery question: how and why, exactly, do the lumps of wrinkly, fatty tissue sitting in our skulls actually do most of the things we think of as, well, thinking?

Although researchers had some success understanding individual cognitive processes, such as perception or action, the study of the mind had largely remained fragmented, says philosopher and cognitive scientist at the University of Copenhagen, Denmark.

What is the free-energy principle?

In the early 2000s, Friston looked to physics and mathematics for a new way to understand cognition. His solution was the FEP.

“The FEP is an extremely universalist approach to the mind,” says Bruineberg. Unlike prior ideas, it sought to unite cognition under one principle: the minimisation of surprise.

The FEP does this by casting the brain as a probability-estimating engine. The name of the brain game, the thinking goes, is to develop beliefs about the world that get as close to reality as possible. Together, those beliefs constitute what is known as a generative model – a set of beliefs that can be used to make guesses about the world. This intuitive process can be rephrased with mathematical rigour using “Bayesian inference”, a statistical method in which pre-existing beliefs are updated based on new information.

Precisely how a gooey blob of neurons with the consistency of warm butter imprints itself with a generative model isn’t important to the FEP – it doesn’t care about the mess of biology. What is important is that the brain updates its beliefs when it receives new data via the senses.

Another key aspect of the FEP is that it is focused on the definition of what it means to be something, says Friston. To exist, a thing needs to be distinguished from everything else, he says. In other words, an object must have a boundary.

To divide the brain from the world it models, Friston implemented another mathematical tool: the Markov blanket. This acts as a sort of causal go-between, determining the relevant information that defines a particular brain state (see What is a Markov blanket?, below). Depending on the scale you are interested in, a brain state could be something as granular as whether a particular neuron is firing or as enormous as depression.

In an abstract sense, you can view the universe as composed entirely of nested Markov blankets, says Friston. “There are blankets within blankets within blankets all the way down and all the way up. You have to pick the level at which you want to apply the free-energy principle.”

Defining the boundaries of objects by Markov blankets allowed Friston and his colleagues to show that systems that maintain a steady state in time – for instance, staying organised into a cell, brain or human – can be mathematically described as performing Bayesian inference on their environment.

However, doing perfect Bayesian inference is impossible, so Friston proposes that systems use an approximation. This is where the “free energy” of the FEP finally comes in. This isn’t the thermodynamic free energy of your high school physics class. It is a quantity from information theory called variational free energy, which you can think of as a measure of surprise. The brain tunes its generative model so that what it perceives lines up with what it predicts, reducing surprise – mathematically, it minimises free energy.

Imagine, for example, you catch the scent of chocolate chip cookies wafting from down the hall. Your brain would then tune its model to include cookies baking in the kitchen.

The FEP wasn’t the first framework to cast perception in Bayesian terms. But it went a step further by uniting perception and action within the same conceptual framework. It was also used to explain diverse cognitive processes like attention and learning. All of which eventually led Friston to ask if the FEP was a in 2010.

The FEP explains action as another way that the brain can minimise the dissonance between its model and its reality. Instead of changing its expectations to match the world, the brain changes the world to match its expectations. It acts to avoid surprise.

For instance, if someone is starving, their brain will receive sensory information from the body that it can use to update its model to know it is starving. But that is a surprising state for a brain to be in. If brains expected to receive starvation signals from the stomach, brains wouldn’t exist for long. So the brain acts, via the body, to try to change the situation by getting some food, ending the surprising sensation.

Action is the FEP’s explanation for how organisms – including their brains – persist. By taking action to avoid the surprise of starvation, a brain persists. Likewise, a fish’s model of itself in the world would expect water, but deem air rather surprising – and a fish out of water would take action to change that by flip-flopping back into the ocean. “It effectively says that survival can only happen on the basis of action,” says philosopher at the University of Wollongong, Australia.

A definition of life

The FEP goes further still. Although it was originally explored as a brain theory, in 2013, Friston published a Walling off a system behind a Markov blanket, he suggested, leads to self-organisation, which could then lead to life, or at least lifelike behaviour.

The border between the living and non-living worlds has long remained frustratingly fuzzy – and there is still no widely accepted definition of life. So the idea that the FEP might provide a universal account of biological self-organisation was a tantalising one. “It was very explicit around that time that the point is to get at the differences between living and non-living systems,” says philosopher

Indeed, much of the principle’s enduring appeal derives from these lofty claims, says Colombo. But he, Nave and others doubted that the FEP’s formal, mathematical definition of what it is to be a thing that persists was up to the task of describing life. Now, some of the hype seems to be fading, says Colombo.

In “Life as we know it”, Friston drew a parallel between the cell membrane and a Markov blanket. After that, scientists and philosophers began to discuss Markov blankets as physical boundaries between objects in the real world, says Bruineberg. In 2021, he and his colleagues and that Markov blankets aren’t able to delineate the natural boundaries of real objects.

“The Markov blanket stuff is a real mess,” says Nave, as it isn’t clear what it applies to. For example, it has been argued that a candle flame lacks a Markov blanket because its boundary changes more quickly than its bulk. Nave argues the same is true for living things. “The parts that make them up are in continual turnover, just as much as a candle flame.”

Moreover, the FEP rests on assumptions that might not apply to life, says complexity scientist at the Basque Center for Applied Mathematics in Bilbao, Spain. In 2022, he and his colleagues found that only they tested could satisfy the FEP’s assumptions. These include that a system will visit every possible configuration of its states through time. Keeping in mind how fussy biology must be for living processes to continue working, it isn’t clear how something alive could satisfy that assumption without destroying itself.

Faced with criticisms like these, Friston and other proponents of the FEP have backtracked on some of the grander claims, says Bruineberg. Today, Friston as a general description of what it means for a thing to exist. That sounds quite staggering, but it actually dilutes the FEP’s original appeal as a specific description of what is special about the mind, and later as a potential theory of life. Nave says there is a tension between the broader, weaker interpretations of the FEP that accommodate life but don’t distinguish it from non-life, and stronger, more specific claims about the nature of biology that don’t stand up to counterexamples. In other words, if the FEP can apply to anything, it is questionable whether it is useful to apply it to life.

This shifting scope has been a source of frustration for Nave and Bruineberg. They describe a pattern in which FEP proponents put forward radical claims only to retreat to more defensible, but less-interesting positions when challenged. “[Others working with the FEP] kind of feel they haven’t been taken seriously,” says Bruineberg.

“The compass of the FEP has certainly increased over time,” says Friston. However, he notes that the underlying mathematics has remained the same and that “from its inception, the FEP makes a careful distinction between living and non-living things”.

Bruineberg also takes issue with the FEP’s original application as a unifying explanation of cognition. One problem, he says, is that the FEP assumes that brains are optimists because they deem anything harmful to be surprising. If you put your hand on a hot stove, your brain could minimise its free energy by updating its model to expect hand-burning sensations. Yet we clearly choose to act instead and pull our hands away from the heat.

“There’s a kind of tension there between this optimism bias and learning from experience, because our everyday environments are very unlikely to be optimistic,” says Bruineberg.

Friston doesn’t see a problem. The FEP is tautological, he says: it assumes that things exist, then describes what things do if they exist. If your brain were somehow wired to expect sensations like burning, freezing, starvation or thirst, it wouldn’t persist for very long, and the FEP wouldn’t describe it. The same goes for brains that easily learn to expect starvation after going without food for a few days. The FEP assumes that brains – and all persisting things – don’t act in ways that would cause the dissolution of their very being.

So the optimism bias is there, but Friston would argue that is sort of the point. “The FEP, in and of itself, gets you absolutely nowhere,” he says. To apply the FEP to a brain you also need to know “the kind of thing” that brain is – you need to know what the brain expects about itself and its environment. “That’s where all the hard work is,” he says.

A new scientific language

The perceived dissonance between what the FEP might seem to offer and its real limitations may point to a larger pattern. a philosopher of science at the University of Cincinnati, Ohio, wonders whether the confusion surrounding the FEP might boil down to a case of unrealistic expectations: perhaps proponents and critics of the FEP have all simply mistaken it for something it isn’t.

“It’s not right to ask whether it’s true in general or true of specific systems – as in true of nature,” says Andrews. “By fixating on this question, ‘Is it true?’, it’s generated a lot of literature around this framework that says effectively nothing.”

Andrews doesn’t believe the FEP is a grand theory explaining life, the universe and everything. It isn’t a theory at all, they say, nor is it a hypothesis to be tested. Instead, the FEP is more accurately described as a set of self-consistent assumptions and mathematical tools that offer a scaffold for research – a kind of language for building new theories.

Although Friston has presented the FEP as a theory in the past, he now says that it is best thought of as a tool to apply rather than a truth to debate. “It is not a theory, it’s not a hypothesis, it’s a principle,” he says. The word “principle” is rather slippery, meaning different things to different scientists. But Friston is clear about what it means to him: it is a truism. “That means all you can do with the free-energy principle is apply it. You can’t talk about it, you can’t admire it, you can’t falsify it, you can’t critique it.”

Every human language smuggles extra layers of meaning and assumptions about what is worth paying attention to into the messages it carries. The same is true for the languages of science. Classical mechanics, general relativity, organic chemistry, genetics – each has its own self-consistent mathematical and conceptual grammar and lexicon that we can use to describe the world. Some are better suited to certain tasks than others. Phrasing the swing of a simple pendulum in the language of general relativity would be clunky overkill. It is possible to make false statements using any of these languages, but, when used carefully, they reveal new facets of the truth.

The free-energy principle is more accurately described as a language for building new theories

The FEP’s dialect casts existence in terms of information exchange between an observed world and an observing agent. The question researchers should be asking, says Andrews, is what this buys us – if the FEP is new scientific language, is it a useful one?

The many, eclectic applications of the FEP in recent years suggest that it is useful to researchers across disciplines. Over the past year, the FEP has been referenced in papers that study how police officers can , why and how our Yet on closer inspection, while the FEP may have been an inspiration to this research, the extent to which it actually offers explanation is debatable, says Bruineberg. “The question is how much good research the FEP really sparked that couldn’t have been done without it.”

Friston, for his part, thinks the FEP is “extremely useful”. In particular, he points to active inference, a concept in machine learning and cognitive science founded on the FEP that is being used to build AIs. “There is a small industry of people that you probably won’t find in the philosophy literature. They’re just people getting on with the job of applying the free-energy principle.”

This “small industry” includes a company seeking to build computers that simulate human thought processes, where Friston is chief scientist. “We’re basing everything on the root of this principle, the free-energy principle,” says Verses chief product officer Hari Thiruvengada. The idea is that by replicating the workings of the mind – or at least the mind according to the FEP – its AIs could form hypotheses about the world and, to some extent, reason.

Verses is now assessing its models against AI image-recognition benchmarks as well as the Atari 100k challenge, which tests an AI’s ability to play video games. So far, the firm can’t announce anything specific, says Thiruvengada. But by Friston and his colleagues that is yet to be peer-reviewed suggests that models like these need significantly less training data to learn to accurately classify images.

Next to these real-world applications, Friston considers wranglings over the FEP’s meaning a spectator sport – one from which physics offers safe retreat. “If one stays close to the physics, there should be no need for proponents or defenders,” he says. “Applications of the FEP may or may not be useful. Time will tell.”

91av audio
You can listen to many articles – look for the headphones icon in our app