Embedded Files
12 June 2015 | Krzysztof Dołęga (Ruhr-Universität Bochum)
Action-oriented Predictive Processing (PP for short, also known as Prediction Error Minimization or Predictive Coding) is an exciting conceptual framework emerging at the crossroads of cognitive science, statistical modelling, information theory, and philosophy of mind
Aimed at obtaining a unified explanation of the processes responsible for cognition, perception and action, it is based on the hypothesis that the brain’s architecture consists of hierarchically organized neural populations performing statistical inference. Rather than accumulating and compounding incoming information, the neural hierarchies continuously form hypotheses about their future input.
Thus, the traditional bottom-up approach to explaining cognition is subsumed by a top-down organisation in which only sensory information diverging from the predicted patterns of activation is propagated up the cortical hierarchies. Due to its divergence from the ‘predicted’ patterns, this information is often referred to as “prediction error” (Clark, 2013). Minimising error is postulated to be the main function of the brain; by accommodating unexpected information in its predictions the brain can fine-tune its future hypotheses regarding sensory input and track the states of the world causing this input more accurately.
However, most of the framework’s appeal lies with the relatively recent proposal that the brain can minimize error not only by revising and constructing new hypotheses about the input patterns, but also by interacting with the environment in order to erase the source of mismatch between the best hypothesis and patterns of sensory activation. This feature, referred to as “active-inference” (eg Hohwy, 2012), is primarily responsible for much of the framework’s appeal and its promise of a unified theory of brain organization, explaining how information about many different cognitive functions is encoded and processed in the brain (Friston, 2010).
In my poster presentation at the first iCog conference, I tried to draw similarities between Predictive Processing and another radical proposal about the nature of perception and cognition – enactivism. Because of the relative novelty of the PP framework, its relationship to the embodied and sensorimotor approaches to cognition and perception has not been well defined (at least at the time of the conference, see the bibliography below for several recent articles tackling these issues).
What struck me as an interesting avenue for research was the similarity between the notion of active-inference on the PP framework and the enactive focus on the role sensorimotor contingencies and possibilities for action play in shaping perception and phenomenology. Pursuing this correspondence is especially valuable for PP, as it does not yet offer a clear account of how phenomenology fits within its probabilistic architecture. Due to perception’s breadth as a topic, I decided to focus on a very particular case of Sensory Substitution Devices.
Sensory substitution devices emerged from the laboratory led by Paul Bach-y-Rita in the ’60s. Bach-y-Rita (1983) set out to prove the extent of lifelong brain plasticity (a highly contested thesis at the time) by devising gadgets that would help handicapped people restore lost senses by substituting them with inputs coming from different sensory modalities.
The idea behind the project was to use the brain’s natural ability to adapt to the inputs it receives from different sensory channels in order to train it to recognize information specific to the lost modality in patterns delivered through a different sensory modality.
Bach-y-Rita’s work focused on tactile-visual sensory substitution (TVSS for short), in which visual information from a video camera was translated into vibro-tactile input on the subjects skin. Despite its limitations, this method proved to be a huge success, as subjects were able to learn to extract vision-like information (e.g. a presence of a white X in front of the camera) from tactile stimulation after a surprisingly short adaptation time. This discovery jumpstarted a whole new field of research. Below is a video of a recent TVSS device:
TVSS proved to be a fertile study ground for enactivism due to limitations and problems inherent to the project of sensory substitution. Very early into his research, Bach-y-Rita discovered that substitution is mostly unsuccessful when subjects do not have control over the camera movements.
The critical importance of exploration and active sampling for TVSS fits well with the core enactive claim that perception consists in ‘exercising a mastery of sensorimotor contingencies’ (O’Regan & Noë 2001: 85), understood as practical know-how about the possible changes in perception of objects caused by our actions (O’Regan & Noë 2001: 99). Moreover, O’Regan & Noë have argued that the contingencies of how our sensory modalities sample and interact with particular objects explain certain features of the brain’s plasticity.
For example, it is because of the dynamics particular to our visual involvement with the world that the blind TVSS subjects show increased activity in the visual cortex and can be said to genuinely see (although in some impoverished way, TVSS does not allow for colour perception). To support this controversial claim they point to neurological data, as well as experiments demonstrating subjects’ gullibility to distinctively visual illusions exploiting the basic properties of visual engagement with the environment, such as making perceptual contact only with the surfaces facing the observer (Hurley & Noë 2003: 143).
Let us now return to Action-oriented Predictive Processing and how the framework can accommodate sensory substitution. PP assumes that the main function of the brain is minimization of prediction error resulting from comparing actual sensory inputs with predicted patterns of activations generated by the system.
To predict the sensory states efficiently, the brain tracks patterns of statistical regularities present in the incoming signals and tries to infer the causal structure of the world responsible for these regularities. Thus the brain constructs and maintains a model of the world, which it uses to predict (i.e. generate) its own sensory states.
The particulars of this proposal are much more complex (I recommend Jakob Hohwy’s 2013 monograph for details), but these core ideas are sufficient to understand how PP can explain sensory substitution.
On Action-oriented Predictive Processing, what happens during TVSS is a result of the hierarchical system recognizing and subsequently tracking a set of statistical regularities specific to the visual modality in sensory patterns delivered through tactile stimulation. The subjects need to have control over the input device (a camera, in this case) in order to learn how the regularities in the sensory stimulus change with the sampling of the environment. Having done this, the brain can predict how the stimulus will change in response to particular actions, updating its generative model accordingly.
The core of the PP explanation of TVSS is thus very similar to the enactive treatment. In both cases it is the system’s ability to recognize possibilities for action and the ability to predict how these actions will change the states of the sensory input that make sensory substitution possible.
One could try and push this similarity further by saying that in PP, just like in enactivism, it is the sensorimotor contingencies that shape the phenomenal quality of the substitution. After all, the system tracks distinctively visual regularities obtaining between the body and the world. From the perspective of the brain, the manner of their presentation (via tactile stimulus vs ocular nerves) plays a secondary role to their contents (in both cases the brain has to infer the causes behind the patterns of sensory activations).
Despite these similarities, one should be careful about casting PP as subscribing to an enactive understanding of perception and sensory substitution. Though the views in question do overlap in their explanatory ambitions, they are built on diametrically opposing assumptions.
In a previous paragraph, I tried to speak about ‘the system’ rather than the brain or agent as a whole. This is because PP is usually understood as a neurocentric view (Hohwy, 2014), while enactivism instead stresses the situated and embodied nature of cognition (Noë, 2005). Moreover, PP is based on an inferential architecture, often associated with rich representational contents – something widely eschewed by enactivists.
The divide between the two positions is not unsurmountable and much of present work by Andy Clark is focused on bridging the gap between these radical views (see Clark’s forthcoming book). This post does not allow me to dive into the nuances of both views and how similarly they treat TVSS and analogous cases; however, I hope I managed to spark some interest in these radical views about perception and how they may be related. Below is a list of references, some of which were unavailable at the time of the original presentation of this material.
References
References
Bach-y-Rita, P. (1983). Tactile Vision Substitution: Past and Future. International Journal of Neuroscience 19: 29–36.
Briscoe, R. (forth.). Bodily Action and Distal Attribution in Sensory Substitution. [online] Available from: http://philpapers.org/archive/BRIBAA [Retrived: 21, Nov. 2013].
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Science 36(3): 181–204.
Friston, K. (2010). The free-energy principle: A unified brain theory?. National Review of Neuroscience 11(2):127–138.
Friston, K. (2008). Hierarchical Models in the Brain, PLoS Computational Biology 4(11) doi:10.1371/journal.pcbi.1000211.
Hohwy, J. (2014). Self-evidencing Brain. Nous 48(1). doi: 10.1111/nous.12062
Hohwy, J. (2013). The Predictive Mind. Oxford: Oxford University Press.
Hohwy, J. (2012). Attention and conscious perception in the hypothesis testing brain. Frontiers in Psychology 3(96). doi: 10.3389/fpsyg.2012.00096.
Hurley, S. & Noë, A. (2003). Neural Plasticity and Consciousness. Biology and Philosophy 18: 131–168.
O’Regan, J.K. & Noë, A. (2001). What is it like to see: A sensorimotor theory of perceptual experience. Synthese 129(1): 79–103 .
Noë, A. (2005). Action in Perception. Cambridge, MA: MIT Press.
Pepper, K. (2013). Do Sensorimotor Dynamics Extend The Conscious Mind?. Adaptive Behavior 22(2): 99–108.
Pickering, M., Clark, A. (2014). Getting Ahead: Forward Models and their place in Cognitive Architecture. Trends in Cognitive Sciences 18(9): 451–456.
Prinz, J. (2009). Is Consciousness Embodied? in Robbins, P. & Aydede, M. (Eds.) Cambridge Handbook of Situated Cognition. Cambridge: Cambridge University Press.
Rietveld, E., Bruineberg, J. (2014). Self-organization, free energy minimization, and optimal grip on a filed of affordances. Front. Hum. Neurosci 8(599). doi: 10.3389/fnhum.2014.00599.
Ward, D. (2012). Enjoying the Spread: Conscious Externalism Reconsidered. Mind 121(483):731- 751.
Report abuse