Theory of Intelligence




Hard to define. Is it the ability to label things and concepts? To make predictions? To generate adaptive behaviour?

Intelligence is a property of an organism and of its relationship with its environment. Intelligence emerges from the self-organization of a system such as the brain. Typical resulting organization has a hierarchical form.


An organism evolves in an environment, that may bring it good or bad eventualities. Natural environments are usually highly complex, full of hidden variables and not fully predictible. Abstraction helps make sense of a complex environment.


Labelling is a perceptual decision process: events, objects or general stimuli that verify a property are grouped into a discrete unit.

Labelling is typically involved in language, but is actually also at the core of action selection. It may also be used to consolidate representations into memory, or to build efficient representations of context or hidden state.


A representation of possible futures, usually weighted by probability values. A prediction is produced by a forward model. A prediction can also be an estimation of hidden states of the environment.


Inference (and especially Bayesian inference) is the process by which an organism evaluates the probability of different events or states.

Forward model

A forward model is an internal model the brain may hold that enables it to predict the future state of the environment under certain assumptions. It is a form of inference.

Adaptive behaviour & fitness

A behaviour is adaptive if it helps the agent thrive in its environment. Evolution leads to adaptive behaviour. Adaptive behaviour typically has representations associated with high value.


The value of an action is computed by organisms as an approximation of how efficient that action would be to maximize the organism’s fitness to its current environment. Organisms tend to select actions of highest value. Many different criteria may be involved in value calculation.


Perception is the process by which an organism builds representations of its environment. Perception leads directly to action in the case of automated responses.

Perception leads to labelling and the creation of concepts whose properties can be estimated. The process of perception typically leads to the construction of the self as a set of representations.


An action is a choice made in response to the environment, or as the result of an internal process. External actions lead to movement of the organism, whereas internal actions include things such as the manipulation of mental objects, the focus of attention on something, etc. Action selection is usually suboptimal, since the brain cannot compute a perfect prediction of all the contingencies of the environment.


Natural organisms behave in a way that could be described as objectively suboptimal, due to constraints in their perceptive and computationnal powers. However when taking these constraints into account, evolution usually finds a quite optimal solution.


The notion of self is typically an emergent representation that the organism holds that describes properties of itself. The self is associated to properties such as position or orientation of different body parts, which may be part of contextual metadata.


A system may self-organize into a given form. Example of a self-organizing system: a groups of neuron that organize to implement ICA using a learning rule such as Oja’s rule.

Learning rule


Evolution, evolution theory

Evolution is the framework in which natural beings develop their distinctive features and behaviours. Evolution leads organisms to being adapted to their environment through the process of natural selection.

Reinforcement learning


Many different notions of hierarchy are involved in the description of intelligence.

Perceptual systems are organized in a hierarchical fashion that map concrete stimuli to more and more abstract representations.


Corrolary discharge (aka efference copy)

A corrolary discharge, also known as an efference copy, is a copy of the signal sent to an effector that is directed to perceptive regions of the brain in which the perception may be affected by the action. The information that an action has been taken is used by the perceptive region to create a stable representation and not be surprised by the effects of the action.


Remapping is a phenomenon that occurs in visual perception during a saccade: visual neurons show activity predictive of the stimulus associated to the new location before the saccade even happens. Remapping happens thanks to corrolary discharge from areas involved in the preparation of a saccade. Remapping is a process leading to visual stability.


Surprise is a perceptive phenomenon that occurs when stimuli are not conform to predictions. In some cases an organism may attribute negative value to the eventuality of surprise, leading it to make as accurate as possible predictions.




Context is a state or representation that may influence brain processes such as perception or action selection. It is typically a representation of current contingencies of the external world or of the self.

Hidden state

The hidden state of the environment is the set of things of that environment that the organism does not directly observe but whose existence it may infer. It does not have any meaning to say that hidden states map to external realities, they are merely abstractions, conveniences of the mind.

Things that already exit in AI

Neural Turing Machines: read the paper.

Deep Q-learning

Hierarchical temporal memory: Le Numenta White Book n’est vraiment pas très bien fait, je vous conseille plutôt de regarder les vidéos de Numenta, et aussi de consulter leur site web.

A blog with nice ideas