What is active inference? Acting to make the world predictable

Active inference is a theory of behavior in which an agent uses an internal generative model to both explain its sensations and choose its actions. Perception updates beliefs to reduce prediction error, and action selects policies expected to minimize future free energy, balancing information seeking with reaching preferred outcomes. Active inference is the action side of Karl Friston’s free energy principle.

Key takeaways

• Active inference extends the free energy principle from perception to action: agents change the world, not only their beliefs, to keep observations predictable.
• Actions are chosen by minimizing expected free energy, which combines pragmatic value (preferred outcomes) and epistemic value (information gain).
• Friston and colleagues’ 2017 process theory in Neural Computation described how neurons might implement this as gradient descent on free energy.
• Experiments with cultured neurons, including DishBrain and work by Isomura and colleagues, have tested free-energy ideas in a dish.
• For organoid intelligence, active inference offers a way to think about training without a reward channel, using the predictability of feedback as the teaching signal.

What is active inference? #

Most theories of the brain separate perception (working out what is out there) from action (deciding what to do). Active inference treats both as the same kind of problem. An agent has a generative model: a set of probabilistic beliefs about hidden states of the world, how those states change, and what observations they produce. It then has two ways to reduce the mismatch between model and data.

1. Perceptual inference and learning. Update beliefs about hidden states, and over longer timescales update the model’s parameters, so the model explains incoming observations better.
2. Action. Change the observations themselves by acting on the world, so that they match what the model expects.

The second route is what gives active inference its name. If you expect to see a word in focus and it is blurred, you can revise your belief about the word or move your eyes. Both reduce free energy, the information-theoretic quantity described in our entry on the free energy principle. That entry covers the general theory; this one focuses on the action and planning side.

How does active inference choose actions? #

The central quantity for planning is expected free energy. For each candidate policy (a sequence of actions), the agent predicts what observations would follow and scores them.

Pragmatic and epistemic value #

Expected free energy can be decomposed into two parts, a decomposition set out in Friston et al. (2017) and in the textbook by Parr, Pezzulo and Friston (2022):

• Pragmatic value measures how well predicted observations match the agent’s prior preferences. In active inference, goals are encoded as observations the agent expects to encounter, not as rewards.
• Epistemic value measures how much a policy is expected to reduce uncertainty about hidden states. Actions that resolve ambiguity, such as looking where information is, score well.

Because both terms sit in one objective, exploration does not need to be bolted on. An agent explores when it is uncertain and exploits when it is confident.

A process theory for neurons #

Friston and colleagues’ 2017 paper, “Active inference: a process theory”, asked whether neuronal responses can be described as gradient descent on variational free energy. Using a standard Markov decision process as the generative model, they derived the neuronal dynamics this implies and reported that simulations reproduced a range of known phenomena, including repetition suppression, mismatch negativity, place-cell activity, phase precession, theta-gamma coupling and the transfer of dopamine responses. In this account, belief updating is carried by message passing between neural populations, and dopamine is linked to the precision, or confidence, assigned to beliefs about policies.

Active inference vs reinforcement learning vs predictive coding #

Predictive coding can be seen as one way the perceptual half of active inference might be implemented in cortex. Reinforcement learning and active inference overlap substantially; some researchers treat reward as a special case of prior preference, while others argue the frameworks make different predictions about exploration and uncertainty.

How has active inference been tested in living neurons? #

Testing a broad theory on behaving animals is hard because the generative model inside a brain is unknown. Cultured networks on a microelectrode array offer a simpler setting where researchers control every input.

Blind source separation in a dish #

Isomura, Kotani and Jimbo (2015, PLOS Computational Biology) stimulated rat cortical cultures with mixtures of two hidden sources and reported that neurons came to respond selectively to individual sources, a result they interpreted in terms of the free energy principle. Isomura et al. (2023, Nature Communications) took this further, reporting that changes in effective synaptic connectivity reduced variational free energy as predicted, and that pharmacologically raising or lowering network excitability disrupted inference in ways consistent with altered prior beliefs. These studies mainly test perception and learning: the cultures did not act on their inputs.

Closing the loop with action #

DishBrain added the action side. Kagan et al. (2022, Neuron) used closed-loop stimulation to embed cultured human and mouse neurons in a simulated Pong game: stimulation encoded the ball’s position, and activity in designated regions moved the paddle. Hits were followed by predictable stimulation and misses by unpredictable stimulation. The authors reported that cultures improved their rally performance, and framed this as networks acting to make their sensory input more predictable. A no-feedback condition helped separate learning from simple adaptation. The full setup is covered in how neurons learn in a dish.

What are the debates and limitations? #

• Falsifiability. Critics argue that with a flexible enough generative model, almost any behavior can be described as active inference. Supporters reply that specific process theories make testable predictions, as the in vitro studies attempt to show.
• Explanation versus description. Showing that a culture behaves as if it minimizes free energy does not prove it computes expected free energy or plans. Simpler mechanisms such as Hebbian learning or homeostatic plasticity might produce similar results.
• Planning depth. A dish of neurons has no obvious way to evaluate multi-step policies. Most in vitro results are better described as learning from feedback than as explicit planning.
• Interpretation of DishBrain. Predictable and unpredictable feedback differ in their statistics, so controls are needed to rule out explanations that do not involve inference.
• Language. Terms such as “beliefs” and “preferences” are formal, not claims about experience. The DishBrain paper’s use of “sentience” drew discussion for this reason.

Why does active inference matter for organoid intelligence? #

Animals learn from rewards signaled by neuromodulators, but a culture or brain organoid on electrodes has no built-in reward channel. Active inference suggests that the structure of feedback, not a reward, can shape behavior. That translates into concrete design choices for living-network experiments:

• Make outcomes legible through predictability. Desired outcomes can be followed by structured, predictable stimulation and undesired outcomes by less predictable input.
• Give the network a real action channel. The culture’s activity must change what it senses next, which requires low-latency neural encoding and decoding.
• Measure beliefs, not only scores. Changes in evoked responses and effective connectivity can test whether a network’s internal model shifts as predicted.
• Run strong controls. Open-loop, random-feedback and no-feedback conditions separate inference from generic changes in stimulation.

Reviews of the field, such as Smirnova et al. (2023, Frontiers in Science), discuss feedback and closed-loop interfaces as part of the roadmap, which our guide to organoid intelligence covers in more detail. Writing encoders, decoders and feedback rules as versioned code, so that an active inference protocol can be rerun and compared across cultures, is part of what we are building at Organoid Programming.

Conclusion #

Active inference treats perception and action as one process: agents revise beliefs and act on the world to minimize free energy, choosing policies that balance preferred outcomes against information gain. It is ambitious, mathematically demanding and still debated, but it has already shaped how researchers design feedback for cultured neurons. For anyone training living networks, active inference offers a principled alternative to reward and a set of testable predictions to check against the data.

Frequently asked questions

What is active inference in simple terms?

Active inference says that an agent, whether a person, an animal or possibly a network of neurons, carries an internal model of its world and acts to keep its observations in line with that model. It does two things at once: it revises its beliefs when they fail to predict what it senses, and it chooses actions likely to bring about expected, preferred observations while also reducing uncertainty.

How is active inference related to the free energy principle?

The free energy principle is the broad claim that self-organizing systems minimize variational free energy, a bound on surprise. Active inference is the account of action and behavior that follows from it. In active inference, agents select policies, or sequences of actions, by scoring how much free energy they are expected to produce in the future, a quantity called expected free energy.

What is expected free energy?

Expected free energy is the quantity an active inference agent uses to rank possible courses of action. It combines two terms: pragmatic value, which favors actions likely to produce observations the agent prefers, and epistemic value, which favors actions likely to reduce uncertainty about hidden states of the world. This gives a single objective that covers both exploitation and exploration.

How is active inference different from reinforcement learning?

Reinforcement learning maximizes a cumulative reward signal, and exploration is usually added through extra mechanisms. Active inference replaces reward with prior preferences over observations and builds exploration into the objective through epistemic value. The two frameworks can produce similar behavior in many tasks, and researchers continue to debate how deeply they differ in practice.

Has active inference been tested in cultured neurons?

Related ideas have been tested. In DishBrain (Kagan et al., Neuron, 2022), feedback design was motivated by the free energy principle, and cultured neurons improved at a Pong-like game. Isomura et al. (Nature Communications, 2023) reported that rat cortical cultures inferring hidden sources behaved as free-energy minimization predicted. These studies support the framework in a dish but do not settle every claim of active inference.

Sources

1. Friston, FitzGerald, Rigoli, Schwartenbeck & Pezzulo (2017). Active inference: a process theory. Neural Computation.
2. Parr, Pezzulo & Friston (2022). Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. MIT Press.
3. Friston (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience.
4. Kagan et al. (2022). In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron.
5. Isomura, Kotani & Jimbo (2015). Cultured cortical neurons can perform blind source separation according to the free-energy principle. PLOS Computational Biology.
6. Isomura, Kotani, Jimbo & Friston (2023). Experimental validation of the free-energy principle with in vitro neural networks. Nature Communications.
7. Smirnova et al. (2023). Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science.