 |
| Chloë FitzGerald |
This month, Chloë FitzGerald introduces chapter 9 of Jakob Hohwy's The Predictive Mind (OUP, 2013). Chloë is postdoctoral Fellow on the project Understanding Implicit Bias in Clinical Care at iEH2 (institut Ethique Histoire Humanités), Centre Médical Universitaire, Université de Genève.
Chapter 9 - Precision, Attention and Consciousness
Presented by Chloë FitzGerald
In Chapter 9, Hohwy describes how the PEM framework can be used to explain the functional role of attention and to illuminate the relation of attention to conscious perception. He claims that the PEM theory of attention is the best explanation we currently have of attention.
Hohwy’s account follows Karl Friston and colleagues’ proposal that ‘attention is just optimization of precisions in hierarchical prediction error minimization’ (p. 194). The sensory signals received by the brain vary in how reliable they are and this variability needs to be gauged to enable perceptual inferences to work well. This is why expected precisions are important for PEM. Howry explains that the brain needs to select where to put in most effort in prediction error minimization and that this effort should be spent where the prediction error signals are expected to be most precise. Attention thus ‘requires learning state-dependent patterns of noise and precision and then using such prior beliefs to set the gain on prediction error’ (p. 195).
As an example of how attention as optimization of precisions works, Hohwy cites the Posner paradigm, where subjects must fix their eyes on a central spot while they are flashed cues that correctly predict 80% of the time where a peripheral target stimulus will show up. Subjects are quicker to detect the target after they have seen a valid cue and the explanation is that subjects attend to the region prompted by the cue, a form of attention sometimes labelled ‘endogenous’ or ‘top down’ because it directed by the subject. According to Hohwy’s account, there is a precision regularity involving the cues correctly predicting the stimulus 80% of the time. Based on this regularity, the gain on prediction error from the cued region is increased and the prior probability of a target occurring there goes up. On his view, these increases are what explains the faster performance on the validly-cued trials. He also cites an experiment by Feldman and Friston (2010) that computationally modelled the Posner paradigm using a predictive coding model that replicated the psychological and neurophysiological results.
The other side of attention is when a loud noise or sudden movement attracts us to attend to it. This is often called ‘exogenous’ or ‘bottom up’ attention. PEM explains this kind of attention, according to Hohwy, by the regularity that ‘stronger signals are more precise’ (p. 197). When attention is caught by a strong signal, such as a loud noise, the brain will expect high precision in the signal, it will increase the gain on prediction error and will try to explain why the input is there.
Hohwy claims that an advantage of the PEM explanation of attention is that it just falls out of the PEM framework and the nature of regularities and precisions. There is no need to posit a special mechanism to explain attention. Another advantage is that the phenomenon of inattentional blindness can be explained by the PEM framework. The functional role of endogenous attention to keep one actively engaged in a particular aspect of the world runs the risk that sometimes other salient stimuli will be missed. To sum up, endogenous attention involves selecting certain stimuli and keeping the model of the world constant (prolonging active inference), while exogenous attention ensures that stimuli that have not been predicted are not ignored (prolonging perceptual inference).
Finally, conscious perception relates to attention in this way through PEM: conscious perception is the upshot of perceptual inference, the conclusion, while attention is what regulates in second order fashion the precisions of the predictions.
Questions and comments:
1. Attention and conscious perception can come apart according to Hohwy’s functional descriptions, and he refers to a Jiang et al.’s experiment as an example of attention without conscious perception. Some might challenge why we would want to call what happens in the Jiang et al. experiment ‘attention’, though. If one starts with an idea of attention as a particular mode of conscious perception, then these experiments don’t seem to be revealing attention, but something else.
2. If we think that there are at least two kinds of attention, conscious and non-conscious, what does PEM have to say about why and how attention becomes conscious?
3. I know that the next chapter will explore the conscious perception in more detail, but one worry is that conscious perception seems to have such a limited role on the PEM account. On p. 202, Hohwy makes the disclaimer that PEM can’t explain why perception is phenomenally conscious, but what is it that determines when the boundary (if we think of it as a sharp boundary at all) into consciousness is crossed?
Hi Chloë – thanks for these great comments. I’ll begin a response to the questions here, though I think there is lots to discuss for all of them.
ReplyDelete1. Can consciousness and attention come apart? This is a good and substantial question. There is a lot of work being done on this so there is probably no one thing to say to clinch it. With respect to the Jiang et al study, my argument is reasonably restricted: i) observe that the Posner paradigm is viewed by most to be a paradigmatic instance of an attentional mechanism; ii) Jiang et al find evidence that something resembling the Posner paradigm is in effect without consciousness (the cue is not seen consciously), so iii) a paradigmatic instance of an attentional mechanism is in effect without consciousness. There are several ways to resist this argument. One is, as Chloë mentions, that there is an intuitive pull to define attention in terms of conscious perception. My view is that it is best to keep the definitions open, not least because there is a lot of research in this area right now, which seems to be asking (somewhat) meaningful questions. Another way to take this objection is as a challenge to i) above. I have some sympathy for this, as the Posner paradigm doesn’t in a really obvious way engage truly endogenous, fully volitional attention. Nevertheless it is the dominant paradigm in attention research.
Related to this issue is the question what PEM says about the relation between consciousness and attention. The chapter suggests that there is always an element of attention in consciousness since consciousness is driven by perceptual inference and inference requires precision weighting, which is attention. So in one sense attention and consciousness cannot come apart. But this is consistent with attentional mechanisms being at play outside consciousness, namely through allocating weights to prediction errors that do not drive the winning hypothesis.
2. This is a nice question: what determines when attention becomes conscious? I think there might be two ways to approach this. First, in studies like Jiang et al, something like Posner attention is nonconscious. What happens when it becomes conscious, that is, when it is becomes a normal Posner paradigm? Described simply, it seems rather passive: the cue is then simply consciously perceived. This would be because accuracy and precision both align better than in the Jiang et al set–up, and the winning hypothesis concerns the cue. In this view, the transition to conscious attention means nothing in particular, beyond a slight change in the probabilistic landscape. Second, one might wonder about a transition to genuine volitional attention, where the agent decides to attend to the target. This is a different kind of process and, as I argue in the chapter, I think it is best viewed as an instance of active inference (selective sampling). One might speculate that active inference is hampered for cases of nonconscious attention, for example, that it is harder to act on the cue even if it can speed up detection of the target (the latter is just evidence accumulation).
3. Chloë is right that the next chapter will delve a bit more into questions of consciousness, and also that I do very little to address the metaphysics of consciousness (cf. the Hard Problem). I think it is right that this raises the worry that consciousness actually has very little role to play in PEM, that is, from the perspective of PEM it matters little if inference leads to consciousness or not, as long as prediction error is minimized. In the next chapter, I argue that the transition to consciousness happens when a hypothesis is chosen for active inference, which is a hugely important moment in a PEM mechanism. I think this view fits with a lot of consciousness science. I think it makes good sense that the function of consciousness is to flag which hypothesis is chosen to be acted upon – this sits well with unity of consciousness as one can only act upon one hypothesis at the time. However, this is purely a functionalist account, which tells us nothing about the metaphysics of consciousness (I can imagine a non-conscious creature with active inference). I am however pretty sure much more can and must be said about the relation between PEM and consciousness (including the relation to existing theories of consciousness).
ReplyDelete