Stability and explicitness: In defense of implicit representation

Stability of activation, while it may be necessary for information to become available to consciousness, is not sufficient to produce phenomenal experience. We suggest that consciousness involves access to information and that access makes information symbolic. From this perspective, implicit representations exist, and are best thought of as sub-symbolic. Crucially, such representations can be causally efficacious in the absence of consciousness. While the hypothesis that information can be causally efficacious (i.e., influence behavior) yet not be available to consciousness is central in most theories of cognition, this assumption has seldom been questioned as directly as O’Brien & Opie (henceforth, O&O) do in their target article. We must point out right at the outset that it seems so intuitively obvious to us that we can do more than we seem to be aware of that the current widespread questioning about the role of unconscious cognition in general (i.e., Shanks & StJohn, 1994) is rather puzzling from our perspective. The basic issue is simple: Can knowledge be “in the system” and influence behavior without being available to consciousness? Even though providing an empirical answer to this question has proven far more difficult than previously thought, the theoretical answer has typically been to deny the problem altogether. This is simply because in the classical framework, representations are passive and only become causally active when they are accessed (interpreted, manipulated, etc.). From this perspective, therefore, the only possibilities to allow for unconscious cognition consist (1) of assuming that all the relevant knowledge is permanently embedded in the functional architecture, or (2) of assuming the existence of a powerful unconscious system that is basically the same as the conscious one, only minus consciousness (see Cleeremans, 1997, for further analysis). O&O successfully defend the claim that classical systems are therefore simply inadequate to conceptualize the implicit/explicit distinction, and offer the connectionist framework as an alternative — a perspective that we very much agree with for having defended it elsewhere (see Cleeremans, 1997, Cleeremans & Jiménez, submitted). However, O&O then strangely end up claiming (1) that most existing empirical evidence supporting implicit cognition is flawed based on an overly superficial review of the relevant literature, and (2) that we are phenomenologically aware of any stable activation pattern in our nervous system. In so doing, O&O thus paint a picture of cognition that again seems to rule out unconscious representation altogether (see also Perruchet & Vinter, 1997). We strongly disagree with these conclusions, while simultaneously espousing the connectionist approach as the framework of choice to understand implicit cognition. The main issue is that by relying so much on stable representations as a vehicle for conscious awareness, O&O's theory ultimately runs into deep conceptual problems. First, the assumption that stability generates phenomenal experience is wholly unsupported: Not only do stable patterns exist in our nervous systems that we seem to be incapable of ever becoming directly aware of (e.g., stable patterns of activation over the light receptors of our retinas), but we also fail to be convinced by the argument that stable patterns in artificial systems such as connectionist networks do not generate phenomenal experience because they are mere simulations, not the real thing. In answer to the first issue, O&O suggest that there is plenty of phenomenology that falls outside the beam of attention, but this argument only substitutes one problem for another: The “vehicle” theory of consciousness thus appears to be viable only at the cost of requiring a “process” theory of attention. The second issue is likewise problematic in that O&O borrow Searle’s (1992) mysterious “causal powers” of real biological systems to defend the claim that artificial stability, in contrast to biological stability, is incapable of producing phenomenal experience. In the absence of supporting arguments, however, this claim is merely a matter of belief. Note that the alternative perspective is no more satisfactory: If indeed one assumes that any stable pattern of activity in an artificial network is sufficient to generate phenomenal experience, the inevitable consequence is that one is then forced to accept panpsychism — a radical step that only few are willing to make (e.g., Chalmers, 1996). In short, stability does not appear to be sufficient to support phenomenal experience. Is it necessary? Let us start by asking how stability is involved in information processing. A basic principle of the connectionist approach (McClelland, 1979), namely cascaded and graded processing, is that a given module can start processing before its inputs have reached stability. In other words, unstable patterns of activation can be causally efficacious, as nicely illustrated by Mathis and Mozer (1995, 1997) or by Becker et al. (1997) through the formalism of attractor networks. Such patterns are no less representational than stable ones: The entire activation space at each layer of a connectionist network is thus both representational and causally efficacious. In such models, stable representations enjoy an enhanced status, and may thus provide the grounds for availability to consciousness, but by the same token, they also are merely specific points in an otherwise similarly causal and representational space. It is thus surprising to see O&O write that “prior to stabilization [...] there is no determinate pattern of activation, and hence no single, physically structured object that can receive a fixed interpretation” (p. 27). Such sentences only appear to restate the assumption, and leave it completely unspecified how static a pattern of activation should be in order to constitute a “physically structured object” or worse, who produces their “fixed interpretations”. Ultimately, by assuming that all patterns of activation are explicit and conscious, O&O end up adopting a position similar in some respects to the classical framework they otherwise reject, and as other commentators point out (Vinter & Perruchet), this perspective similarly results in deep problems with learning. O&O’s use of the language of process theories of consciousness in this context is also a hint that what makes a representation explicit is not mere stability, and in fact, O&O’s interpretation of the term “explicit” varies throughout the paper, by being sometimes taken to be equivalent with causal efficaciousness and consciousness, sometimes with “externally identifiable representation”, and sometimes with “interpreted representation” (as opposed to first-order patterns). In this respect, Dennet’s definition of explicitness as quoted by the authors on p. 14 appears to us to be the most convincing: Information is explicitly represented whenever it involves a pattern that is being interpreted by the system as a representation — a symbol. Crucially, this definition does not rule out causally efficacious sub-symbolic representations , which is exactly what we believe to be necessary to understand implicit cognition. The stability criterion may therefore be a necessary condition to support consciousness, but it does not appear to be sufficient to support metaknowledge, that is, to support a form of consciousness that involves access to knowledge as knowledge. This form of consciousness, however, is crucial for abstract thought and explicit learning, and is probably what specifically characterizes human cognition (Clark & Karmiloff-Smith, 1993). In our perspective, thus, patterns of activation in connectionist networks are continuously causally efficacious, whether stable or not, and do not in and of themselves generate phenomenal experience. Rather, they are potentially available to consciousness depending on other factors such as stability, strength, global coherence, access by some other structure, or their compositional and systematic character. Such patterns are best characterized as sub-symbolic. The genuinely hard problem is then to determine how such patterns can become symbolic, explicit, and conscious, that is, how they can be taken by the cognitive system as representations .