What Block calls a “mesh” can also be considered a “model” of the interrelationship between neural and psychological data, and of the relationship of both sources of data to the concepts of phenomenology and accessibility. This kind of model fitting is massively underdetermined, and there is a real danger of overfitting – massaging the data to create meshes that can support any assumption. In a Bayesian framework, competing quantitative models are evaluated by comparing their marginal likelihoods – how likely the observed data is under each model. The marginal likelihood incorporates an Occam's razor–like penalty, penalising complex models that have more parameters, and are thus more likely to overfit the data (Mackay Reference Mackay2004). The kind of “mesh” at issue here is clearly not yet a quantitative model for which a marginal likelihood can be computed, but the principles of model comparison should still be followed as closely as possible.
Block presents a single mesh whose explanatory power gives us “reason to believe” (target article, sect. 1, para. 2) the assumption embedded in it. We would like to emphasise the necessity of comparing fully developed competing meshes before we can have more reason to believe Block's assumption than its counterpart – a single entrant in a race will always be the winner. We will now use Block's example to demonstrate some of the issues inherent in evaluating whether a particular mesh really does provide the “best explanation” in this domain.
In his example, Block claims that without making the assumption that accessibility is constitutive of phenomenology it is not possible to build a mesh that provides a mechanism for overflow. The starting point is an argument from behavioural data that phenomenology overflows accessibility, which is then taken as a given for the mesh to explain. This conclusion is itself a kind of inference to the best explanation, which ultimately contains an assumption about the answer to the methodological puzzle. The importance of the puzzle is diminished by embedding an assumption about the answer within a larger explanatory structure. However, how much the structure relies upon such assumptions will affect the strength of the evidence it provides, and it is therefore essential to be extremely precise about which assumptions are made, and where.
Another issue that makes model comparison particularly difficult in this domain is the imprecise terminology. In his example, Block uses unitary concepts of accessibility and reportability to refer to different phenomena, in different contexts. This can lead to confusion. For instance, whatever is currently part of access-consciousness is reportable in the sense that it can be reported at any time, without any reorienting of attention or the sense organs. Objects in a Sperling task are reportable only if attended, but no terminological distinction is made between these two types of reportability even though the distinction is critical to Block's disentangling of phenomenology and accessibility. We propose that a more precise taxonomy of the different types of accessibility and reportability would be easy to develop and would resolve much of this confusion.
A significant source of difficulty in building and comparing meshes is how we should describe the neural data, and how it maps to psychological states. Block acknowledges that the neural data he invokes may be “wrong or at least highly incomplete” (sect. 13, para. 3). Therefore, great care should be taken at this stage because an error here can cascade through the mesh, severely limiting the strength of the evidence it provides. In demonstration of this point we suggest an alternative mapping between neural states and phenomenology which, in our opinion, would strengthen the mesh by better accounting for the data, or at least cast doubt over Block's version.
Block proposes that cortico-cortical (CC) recurrent loops from higher to lower areas are the core neural basis for phenomenology – for example, that feedback from V5 to V1 is constitutive of phenomenal motion. We propose that motion phenomenology should instead be viewed as being composed of different component phenomenal characters that correspond to the activity of independent but interacting cortical areas, and that recurrent feedback is then necessary for binding these phenomenal characters together. As an illustrative oversimplification, the phenomenal experience of a simple moving grating could be characterised as being composed of a phenomenal experience of a retinotopic array of contrast elements, and a phenomenal experience of motion within a particular retinotopic region. It has been argued elsewhere that the best candidate for the core neural basis of the phenomenal experience of contrast elements is V1 (Pollen Reference Pollen1999) and of motion is V5 (Zeki & Bartels Reference Zeki and Bartels1999). Under normal conditions the two attributes are bound to each other and one directly experiences the surface qualities of the thing that is moving – a modal experience. If the two are not bound, one will have an amodal experience of motion.
This hypothesis can be tested by looking at what happens when feedback from V5 to V1 is prevented. Block claims that V5 activity over a certain threshold in the absence of feedback to lower areas is not sufficient for the experience of motion (sect. 13). We would claim that it is in fact sufficient for an amodal phenomenal experience, supported by the observation (cited by Block in his Note 10) that patient G.K. can experience fast motion in his “blind field,” described as looking like “black moving on black” (Zeki & ffytche Reference Zeki and ffytche1998), despite having no possibility of recurrent feedback from V5 into V1/V2 in his damaged hemisphere (Semir Zeki, personal communication).
The same logic can be applied to the transcranial magnetic stimulation (TMS) evidence that disrupting feedback from V5 to V1 prevents the perception of a moving phospene (Pascual-Leone & Walsh Reference Pascual-Leone and Walsh2001). Subjects reported that the phosphene was “present but stationary,” which is equally consistent with the role of CC feedback in binding phenomenal contents. A true test would be whether subjects can distinguish between sham and real V5 stimulation in the context of disrupted feedback to V1, and whether this is correlated with an experiential report of amodal motion. Adopting this alternative mapping would potentially strengthen Block's mesh by better fitting the neuroscientific evidence.
Block's proposed approach compares meshes as a proxy for a direct comparison of competing hypotheses about the relationship between accessibility and phenomenology. To make this comparison meaningful, we must be very precise about the necessity of the hypothesis to the explanatory power of the mesh. From this perspective the explanatory gap may prove to be more like a crack in the pavement than a gaping hole, but we must still take great care in stepping over it.
