[From MK (2016.05.08.1545 CET)]
Readable intro re predictions vs intentions.
···
----
McGregor, Baltieri & Buckley, 2015: A Minimal Active Inference Agent
arXiv:1503.04187 [cs.AI]
http://arxiv.org/abs/1503.04187
1 Introduction
Active inference is the name given by Karl Friston to a process
of an agent’s simultaneously co-constrained inference regarding the world
combined and directed actions to change it, understood specifically as the
minimisation of an information-theoretic quantity describing the instantaneous
relation of a system to its "typical" sensory environment.
According to proponents of active inference, all mechanical implementation of
cognitive behaviour (including both the active and inferential aspects of
cognition) can be reduced to the minimisation of a single informational "free
energy" quantity; moreover, this quantity can be assigned an interpretation
using cognitive concepts resembling beliefs and intentions.
This theory has attracted a great deal of attention (both positive and
negative), since it claims to unify a wide variety of principles in cognitive
science, machine learning and neuroscience. However, a clear and balanced
discussions of the framework’s strengths and weaknesses is still lacking.
We suspect that one of the barriers preventing a wider understanding of the
theory is the language which been used to describe it. The language is
technically dense, using terms like ’surprise’, ’generative’, ’ensemble’,
’variational’ and can seem to beg philosophical questions in a way which is not
strictly mandated by the formalism of the active inference theory. For instance,
the framework is frequently characterised as claiming that cognitive agents act
so as to "confirm their predictions", or "minimise their surprise"; we will see
that alternative interpretations exist for the same formal assertions.
Furthermore one of the central strengths of the formalism is its potential to
unite an understanding of action and perception within a single
unified framework
[13]. However understanding the exact implications of this will likely require
the utilisation of agent based approaches which to date have largely been
absent.
This paper aims to clarify the foundations of the active inference or
"free-energy" framework in terms which can be understood by a wide audience. We
start by giving a brief conceptual overview of the information Free-Energy (IFE)
principle and active inference in an intuitive language.
[...]
One of the most interesting aspects of the active inference framework is that
two things which we generally regard as quite distinct are given the same formal
treatment: on the one hand, beliefs or predictions about sensory input and the
external world which underlies it, and on the other hand intentions regarding
the outcomes of our actions. In this sense, active inference blurs the
distinction between a prediction and an intention; both are treated as
conceptions-of-the-future, with the main difference being in what physical
variables change most when the reality turns out to be different from the
expectation [13].
Although the free-energy literature uses the terms "prediction" and
"expectation" interchangeably for this concept, it seems that only the verb
"expect" in English really captures the ambiguity between intention and
prediction which the active inference framework rests upon. It is unfortunate
that the word "expectation" also has a technical meaning which is relevant to
active inference, i.e. the mean of a probability distribution. Hence, we propose
the term "expectance" to refer to the superclass of probabilistic beliefs and
desires implicit in the active inference framework.
In this language, active inference asserts that agents constantly behave so as
to reduce the discrepancy between their expectances and reality; the resolution
of discrepancy by changing internal variables is known as "updating one’s
predictions", while the resolution of discrepancy by changing reality is known
as "acting on one’s environment".
Crucially, under active inference both of those dynamics are always occurring
simultaneously. It is not necessary to pre-judge which of those dynamics will
predominate by calling the expectance a "prediction" or an "intention" in
advance of observing the outcome. The typical presentation of active inference
states that agents behave so as to make their "predictions" come true. We
believe this wording erases the distinction between internal and external
resolution of expectance violation, in an unhelpful way. Intuitively, if an
agent "predicts" its arm to move to the left, and the arm instead stays in
place, we would expect it to change internally so as to make different
predictions; no action on the world would be presupposed by the discrepancy. On
the other hand, if the agent "intends" its arm to move to the left, and the arm
instead stays in place, we would expect it to exert a greater force on its arm
by contracting the relevant muscles.
Of course, it is entirely possible that the agent could begin by attempting to
move its arm and, in doing so, learn that its range of movement is restricted by
external forces. The expectance violation can be resolved by coming to a halfway
point in which the agent "updates its predictions" to reflect the fact that its
movement is restricted, "modifies its intentions" to be consistent with what is
achievable, and "acts on the world" by moving within its available range. Such
an equilibrium can emerge naturally from a continuous-time process where
informational free energy is minimised along distinct internal dimensions
simultaneously. ...
----
M