Definition
Free energy principle: The free energy principle is a theory, developed by Karl Friston, that self-organizing systems such as brains persist by minimizing variational free energy, a measurable bound on how surprising their sensory inputs are. They do this by updating internal beliefs (perception) and by acting to make inputs more predictable (action).
Also known as: FEP, active inference, variational free energy minimization
The free energy principle (FEP), developed by neuroscientist Karl Friston, holds that brains and other self-organizing systems persist by minimizing variational free energy, a quantity that bounds how surprising their sensory inputs are. They do this by updating internal models (perception) and by acting to make inputs match predictions (action).
Key takeaways
- The free energy principle frames perception, learning and action as one process: reducing the gap between what a system predicts and what it senses.
- “Free energy” here is an information-theoretic quantity, not thermodynamic energy. It is an upper bound on surprise that a system can compute from its own model.
- Active inference is the action half of the theory: organisms move, sample and intervene to make the world more predictable.
- The principle motivated the design of DishBrain, where cultured neurons learned a Pong-like game when feedback was predictable after hits and unpredictable after misses.
- It matters for organoid intelligence because it suggests a way to train living networks without dopamine-style rewards.
- The framework is influential but debated, mainly over how falsifiable it is.
What does the free energy principle actually say?
Friston set out the idea in a 2006 paper in the Journal of Physiology-Paris and summarized it in an influential 2010 review in Nature Reviews Neuroscience, provocatively titled “The free-energy principle: a unified brain theory?” The core argument runs like this:
- Any living system has to stay within a limited range of states (body temperature, ion concentrations, structural integrity) to keep existing.
- Staying in expected states means avoiding states that are, in a statistical sense, surprising.
- A system cannot measure surprise directly, because that would require knowing the true probability of every possible input.
- It can, however, compute variational free energy, which is always greater than or equal to surprise. Minimizing free energy therefore keeps surprise low.
To compute free energy, a system needs an internal generative model: a set of beliefs about the hidden causes of its sensations. The difference between what that model predicts and what actually arrives is, roughly, prediction error. This is why the FEP is closely tied to predictive coding, a theory in which cortical circuits pass predictions down and prediction errors up.
Perception and action as two sides of one process
A system has two levers. It can change its beliefs so they explain the data better, which corresponds to perception and learning. Or it can change the data by acting on the world, which Friston calls active inference. If you expect your hand to be warm and it feels cold, you can either revise your expectation or put on a glove. Both reduce the mismatch.
How does the free energy principle relate to machine learning?
Researchers in machine learning will recognize the mathematics. Variational free energy is the negative of the evidence lower bound (ELBO) used to train variational autoencoders and other approximate Bayesian models. The table below maps the vocabulary across fields.
| Concept in the FEP | Machine learning equivalent | Biological interpretation |
|---|---|---|
| Generative model | Probabilistic model of data | Internal beliefs encoded in neural connectivity |
| Variational free energy | Negative ELBO | Prediction error plus model complexity |
| Surprise | Negative log evidence | How unexpected an input is |
| Perceptual inference | Inference over latent variables | Updating neural activity to explain inputs |
| Learning | Parameter updates | Synaptic plasticity |
| Active inference | Model-based control | Movement, stimulation choice, sampling |
The parallel is one reason the FEP attracts computer scientists. It offers a single objective that, in principle, covers inference, learning and decision-making.
Why does the free energy principle matter for biocomputing?
Biocomputing with living neurons faces a practical problem: how do you tell a dish of cells what you want? Animals learn from rewards, often signaled by dopamine, but a culture on a microelectrode array has no reward system. The FEP suggests an alternative. If neural networks inherently prefer predictable input, then predictability itself can act as a teaching signal.
The DishBrain experiment
Kagan et al. tested this idea in Neuron (2022). Cultured human and mouse neurons were connected to a simulated Pong game through closed-loop stimulation: electrical stimulation encoded the ball’s position, and activity in designated regions moved the paddle. When the paddle hit the ball, the cells received predictable stimulation. When it missed, they received unpredictable stimulation. The authors reported that cultures improved their rally performance over minutes, which they interpreted as networks reorganizing to avoid unpredictable input. We cover the setup in detail in how neurons learn in a dish.
Testing the theory directly in vitro
Isomura et al. reported in Nature Communications (2023) that rat cortical neurons cultured on electrode arrays, exposed to mixed stimulation patterns, self-organized to separate the hidden sources in a way consistent with the predictions of free-energy minimization. They also reported that pharmacologically altering network excitability shifted the outcome in line with the model. Work like this turns the FEP from a broad philosophy into specific, quantitative predictions about neural encoding and decoding in a dish.
Free energy principle vs reinforcement learning
The FEP is often contrasted with reinforcement learning, the dominant framework for training agents with rewards.
| Free energy principle | Reinforcement learning | |
|---|---|---|
| Objective | Minimize surprise (via free energy) | Maximize cumulative reward |
| Teaching signal | Prediction error, predictability of input | Scalar reward |
| Role of exploration | Built in as information seeking | Usually added with heuristics |
| Biological analogue | Cortical prediction, homeostasis | Dopamine reward prediction error |
| Use in dish experiments | Predictable vs unpredictable feedback | Requires a reward channel cultures lack |
The two are not strictly opposed. Some researchers argue that reward can be folded into the FEP as “prior preferences” about which states are expected. In practice, for cultured neurons and brain organoids, predictability-based feedback is attractive because it only requires stimulation patterns, which an electrode array can already deliver.
What are the limitations and criticisms?
The free energy principle has strong supporters and serious critics. Common concerns include:
- Falsifiability. Because almost any behavior can be described as minimizing some free energy under some generative model, critics argue the principle is hard to disprove. Supporters respond that it is a framework, like the principle of least action, and that specific process theories derived from it are testable.
- Mathematical opacity. The literature is dense, and terminology (Markov blankets, generalized coordinates, expected free energy) has shifted over time.
- Mechanism versus description. Showing that a network behaves as if it minimizes free energy does not prove that it computes free energy. Other mechanisms, such as Hebbian learning or homeostatic plasticity, may produce similar results.
- Interpreting dish results. In DishBrain, the difference between predictable and unpredictable feedback might also change overall stimulation statistics, so alternative explanations have to be ruled out by careful controls.
These debates are healthy. For biocomputing they translate into a clear engineering question: which feedback designs reliably shape behavior in living networks, and why?
How is the free energy principle used in experiments today?
For anyone designing experiments with organoid intelligence systems, the FEP suggests concrete design choices:
- Encode the task state as structured, predictable stimulation.
- Deliver predictable feedback after desired outcomes and less predictable or random feedback after undesired ones.
- Measure whether network activity, not just task score, shifts in ways consistent with reduced prediction error.
- Include controls with no feedback or randomized feedback to separate learning from simple adaptation.
Running such protocols requires precise timing, a closed-loop runtime and careful logging of every stimulus. That infrastructure is part of what we are building at Organoid Programming, so that feedback rules inspired by the free energy principle can be written as code, versioned and compared across labs. For the wider context, see what is organoid intelligence.
Frequently asked questions
What is the free energy principle in simple terms?
The free energy principle says that any system that keeps itself organized over time, like a brain or a cell, must avoid surprising states. It does this by building an internal model of its world, using that model to predict incoming signals, and then either updating the model or acting on the world whenever predictions fail. Free energy is the mathematical quantity that tracks this mismatch.
Who proposed the free energy principle?
The free energy principle was developed by the neuroscientist Karl Friston at University College London. Key early statements include a 2006 paper in the Journal of Physiology-Paris, “A free energy principle for the brain”, and a widely cited 2010 review in Nature Reviews Neuroscience titled “The free-energy principle: a unified brain theory?”
Is free energy here the same as in physics?
Not exactly. The term borrows its mathematical form from statistical physics, but in the free energy principle it refers to variational free energy from information theory and machine learning. It is an upper bound on surprise (negative log probability of sensory data) that a system can compute from its own internal model, rather than a thermodynamic energy measured in joules.
How does the free energy principle relate to DishBrain?
In the DishBrain study (Kagan et al., Neuron, 2022), cultured neurons played a Pong-like game. When they hit the ball they received predictable stimulation, and when they missed they received unpredictable stimulation. The design was motivated by the free energy principle: if networks prefer predictable inputs, they should change their activity to hit the ball more often, and the authors reported that they did.
Is the free energy principle proven?
The free energy principle is a general theoretical framework rather than a single testable hypothesis, so it is not proven in the usual sense. Specific models derived from it can be tested. For example, Isomura et al. (Nature Communications, 2023) reported that in vitro neural networks self-organized in ways consistent with free-energy minimization. Critics argue the framework is so general it can be hard to falsify.
Sources
- Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
- Friston, Kilner & Harrison (2006). A free energy principle for the brain. Journal of Physiology-Paris.
- Kagan et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron.
- Isomura et al. (2023). Experimental validation of the free-energy principle with in vitro neural networks. Nature Communications.
