Deep Active Inference and Scene Construction

Abstract

14Adaptive agents must act in intrinsically uncertain environments with complex latent structure. Here, 15 we elaborate a model of visual foraging -in a hierarchical context -wherein agents infer a higher-order 16 visual pattern (a 'scene') by sequentially sampling ambiguous cues. Inspired by previous models of scene 17 construction -that cast perception and action as consequences of approximate Bayesian inference -we 18 use active inference to simulate decisions of agents categorizing a scene in a hierarchically-structured set-19 ting. Under active inference, agents develop probabilistic beliefs about their environment, while actively 20 sampling it to maximise the evidence for their internal generative model. This approximate evidence 21 maximization (i.e. self-evidencing) comprises drives to both maximise rewards and resolve uncertainty 22 about hidden states. This is realised via minimization of a free energy functional of posterior beliefs 23 about both the world as well as the actions used to sample or perturb it, corresponding to perception 24and action, respectively. We show that active inference, in the context of hierarchical scene construction, 25 gives rise to many empirical evidence accumulation phenomena, such as noise-sensitive reaction times 26 and epistemic saccades. We explain these behaviours in terms of the principled drives that constitute the 27 expected free energy, the key quantity for evaluating policies under active inference. In addition, we report 28 novel behaviours exhibited by these active inference agents that furnish new predictions for research on 29 evidence accumulation and perceptual decision-making. We discuss the implications of this hierarchical 30 active inference scheme for tasks that require planned sequences of information-gathering actions to infer 31 compositional latent structure (such as visual scene construction and sentence comprehension). Finally, 32 we propose experiments to contextualise active inference in relation to other formulations of evidence 33 accumulation (e.g. drift-diffusion models) in tasks that require planning in uncertain environments with 34 higher-order structure. 35

Realising Active Inference in Variational Message Passing: the Outcome-blind Certainty Seeker

“…Another interesting direction of research would be to design new generative models that can tackle more complex tasks, such as playing Atari games, human-machine interaction using natural language and automatic structure learning. Partial answers to these directions of research have already been provided with the use of deep active inference (Fountas et al, 2020;Ueltzhöffer, 2018;, deep temporal models Heins et al, 2020) and Bayesian model reduction Friston et al, 2017a;Wauthier et al, 2020). Nevertheless, we anticipate that additional work will pursue these avenues of research.…”

“…Active inference is also a form of planning as inference (Botvinick and Toussaint, 2012) consistent with Occam's Razor (Blumer et al, 1987) and can be seen as a generalisation of reinforcement learning (van Hasselt et al, 2015;Lample and Chaplot, 2016) and Kullback Leibler control (Rawlik et al, 2013). This framework has also been used to ground active vision (Ognibene and Baldassare, 2015;Heins et al, 2020;Van de Maele et al, 2021;Mirza et al, 2016Mirza et al, , 2018 within a strong theoretical framework.…”

Epistemic Communities under Active Inference

“…At first glance, it might appear difficult to model a phenomenon like confirmation bias using an active inference formulation, because action selection in active inference is guided by the principle of maximizing Bayesian surprise or salience, which requires constantly seeking out information that is expected to 'challenge' one's world model [85][86][87].…”

A mixed generative model of auditory word repetition

“…Under this framework, one can generate distinct behaviours using different generative models. For example, active inference has been shown to successfully simulate a wide range of complex behaviours, including word repetition with fully discrete models (Sajid et al, 2020a; Sajid et al, 2020b; Sajid et al, 2021b), dyadic exchanges (Friston and Frith, 2015; Friston et al, 2020a), active listening (Friston et al, 2020b), active vision (Parr et al, 2021) and scene construction (Friston et al, 2017d; Parr and Friston, 2017; Heins et al, 2020).…”