[1]

Parts of this section condense and draw upon materials from Clark (in press).

[2]

There is now a large, and not altogether unified, literature on enaction. For our purposes, however, it will suffice to consider only the classic statement by Varela et al. (1991). Important contributions to the larger space of enactivist, and enactivist-inspired, theorizing include Noë (2004, 2010, this collection), Thompson (2010), and Froese & Di Paolo (2011). The edited volume by Stewart et al. (2010) provides an excellent window onto much of this larger space.

[3]

Such a process repeats at several organizational scales. Thus we humans do not merely sample some natural environment. We also structure that environment by building material artifacts (from homes to highways), creating cultural practices and institutions, and trading in all manner of symbolic and notational props, aids, and scaffoldings. Some of our practices and institutions are also designed to train us to sample our human-built environment more effectively – examples would include sports practice, training in the use of specific tools and software, learning to speed-read, and many, many more. Finally, some of our technological infrastructure is now self-altering in ways that are designed to reduce the load on the predictive agent, learning from our past behaviors and searches so as to serve up the right options at the right time. In all these ways, and at all these interacting scales of space and time, we build and selectively sample the very worlds that - in iterated bouts of statistically-sensitive interaction - install the generative models that we bring to bear upon them.

[4]

I have engaged such arguments at length elsewhere – see Clark (1989, 1997, 2008, 2012). For sustained arguments against the explanatory appeal to internal representation, see Ramsey (2007), Chemero (2009), Hutto & Myin (2013). For some useful discussion, see Sprevak (2010, 2013), Gallagher et al. (2013).

[5]
For reviews and discussions, see Bengio (2009), Huang & Rao (2011), Hinton (2007), and Clark (in press).
[6]

For a sustained discussion of these failings, and the attractions of connectionist (and post-connectionist) alternatives, see Clark (1989, 1993, 2014), Bechtel & Abrahamsen (2002), Pfeifer & Bongard (2007).

[7]

Bayesian perceptual and sensorimotor psychology (see for example, Rescorla 2013; Körding & Wolpert 2006) already has much to say about just what worldly and bodily states these may be.

[8]

The point about multiple areas (not just multiple levels within areas) is important, but it is often overlooked in philosophical discussions of predictive processing. Different neural areas are best-suited – by location, inputs, structure, and/or cell-type - to different kinds of prediction. So the same overarching PP strategy will yield a complex economy in which higher-levels predict lower levels, but different areas learn to trade in very different kinds of prediction. This adds great dynamical complexity to the picture, and requires some means for sculpting the flow of information among areas. I touch on these issue in Clark (this collection). But for a much fuller exploration, see Clark (in press).