Greg Detre
Thursday, April 03, 2003
In this
chapter, Minsky develops his ideas about how we deal with problems and mental
impasses, and employ different Ways to Think in response. The central idea is
of a Critic/Selector pairing, where the Critics are alerted by the fulfilment
of certain preconditions to some problem or difficulty, and they in turn
activate one of the Selectors they are paired to which they believe will be
most appropriate. Each Critic recognises a specific problem-type, and each
Selector activates the set of resources that correspond to a different way of
thinking.
When these
Selector-induced cascades are large enough, and externally visible, they can be
characterised as emotional states. This is one of the most satisfying
computational definitions of emotions I�ve heard. It encompasses the somatic
aspect of emotion that the James-Lange theory emphasises, while still leaving
room for recent results positing an important (and integrated) role for emotion
in cognition. Finally, it fits with our introspective intuition that there is
some real and clean distinction to be drawn between emotions and other
high-level or global cascades of activity (by defining emotions as the visible
cascades). Similarly, na� psychological distinctions, such as the difference
between a �mood� and an �emotion� can be distinguished in terms of the
time-frame that the cascade operates over. Sloman�s[1]
distinction between emotions at different levels fits too.
However,
I�m not sure that all emotions build as a response to cognitive difficulties. I
would have liked to have seen some mention of how selfish gene theory might
predict that some of our genetically hard-wired emotions exist because they
make our bodies behave in a way that leads the gene that encodes them to
proliferate. Such emotions could still be considered as externally visible
cascades, perhaps initiated by Selectors, but not necessarily in response to
activation of a Critic. Or if we take the example of �honour� (although I can�t
decide for sure whether this counts as an emotion), which� evolutionary psychologists[2]
describe as a kind of �tit for tat� mechanism, in that it enables us to avoid
costly conflict. This is because our �honour� would force us to act out on threats,
making them an effective deterrent, and adding binding force to contracts that
might otherwise be reneged on with impunity. Something like �disgust� is
another case where we learn what we should be disgusted by (probably from our
Imprimers), but I don�t see how Minsky could claim that such an emotional
cascade is initiated by a cognitive problem-type.
It was
initially unclear to me why the functions of Critics and Selectors are
separated. After all, why couldn�t the Critic that looks for the preconditions
indicating a particular problem-type do the responding itself? One reason for
separating them is that one problem-type might be best-suited to different ways
to think depending on the context/domain. By allowing the Critics to link to
multiple Selectors (presumably with varying strengths of weights), and each
Selector to be linked to by more than one Critic, the overall choice of
Selector can be influenced by intra-level competition and top-down influences,
allowing the system to make a decision that reflects a much wider range of
factors.
One
alternative to the Critic/Selector alarm/response paradigm is Hofstadter�s very
different approach to dealing with dead ends and impasses in his Copycat[3]
model. He doesn�t use multiple reflective levels[4]
or explicit Critic agents[5].
It involves the notion of a �temperature�, which is a broad, global variable
reflecting the system�s satisfaction with its current progress. As the system
builds representations that appear to work better (measured in terms of the
richness and coherence of their internal structure), the temperature goes down,
and the system continues with the approach so far. If the current path appears
to have reached a dead end, the temperature starts to rise again, causing the
system to behave in a more flailing, stochastic manner, dismantling current
structures and allowing more outlandish low-level processes a chance. This way,
when the system hits a snag or impasse, the temperature will rise slowly until
a new solution is found, without any explicit alarms, critics or higher-level
reflections being necessary. The self-criticism and initiation of a new way to
think is an emergent property of the system that becomes more likely and more
violent as the system becomes dissatisfied with its current progress.
Interestingly, because the temperature affects the activity of all the
low-level codelets in subtle ways, the system could appears to exhibit a
serial, centralised behaviour from outside. No alarm bells for particular
problem types ring, and no predefined responses are activated.
Another way of seeing the difference between the Critic/Selector and the global temperature/satisfaction variable approach is in terms of the distinction that Norman et al.[6] make between �strong� and �weak� methods. In short, strong methods require domain-specific knowledge to provide specialised solutions to particular tasks or problems, while �weak� methods provide more general, less efficient solutions that apply widely. We could see the Critics as providing strong methods, while temperature/satisfaction measures could provide a complementary weak method for problems where no available Critic would work well. I suppose we could see the idea of a temperature variable and the very highest-level, coarsest-grained Critics as converging though.
Considering
how high-level global temperature/satisfaction and Critic/Selector models are
different raises a related issue. It could be argued that the reason that
Hofstadter�s Copycat is able to use a single, global temperature variable is because
Copycat is only interested in one type of problem (that of finding a mapping
between pairs of strings and applying it to a new string). The central purpose
of Critics is to identify different problem types, and choose
appropriate approaches. So, we would have to allocate a temperature measure for
each sub-system, calculated using preconditions specific to that sub-system�s
problem-type, requiring the introduction of specialist agents that compare the
activity of the different sub-systems, and choose accordingly. Pretty soon, a
mechanism sufficiently general to do this might well end up looking like the
Critic/Selector model.
However, in
discussing this problem of how we could compare different measures of
temperature across sub-systems, I realised that this could be a problem for
Critics too. How can the system determine that one Critic should trump another
Critic that�s equally concerned? Minsky proposes that this could be resolved at
the same level by lateral inhibition, or by choosing the one that gives the
most specific advice, or at a higher-level. This goes some way towards
addressing the problem, but it assumes that the critics have already been
assigned valued weights that can be directly compared � my question is where
these weights come from. In a way, it�s a problem of credit assignment, but
within the system. If the mind tries various different ways to think, and
eventually makes progress, it�s not always the most recent one that has helped
the most, and it�s not possible to tell which of the earlier or lower-level
Critic-Selector activations was instrumental in shaping the representation into
a useful form.
In a way,
this worry is a veiled criticism of the lack of discussion of self-organisation
in the Society of Mind in general. I�m aware that this isn�t one of Minsky�s
central concerns, but since most of the learning mechanisms that are described
(since learning is an advanced form of self-organisation) rest on top of an
already fairly well-organised mind, we need to know how to get to this stage
first.
Admittedly, AI researchers aren�t required (like evolution) to build incrementally from the ground up, but can jump in by pre-building as much of the system as they like. However, my feeling is that even the minimum size for a mind interesting enough to bootstrap itself much further is probably too big and complex for us to be able to construct much of its componentry and inter-relations by hand. I�d be much happier if I could see a brief discussion of how the very early stages of the brain�s organisation might work. Presumably, any mind has to start with some hard-wired, basic survival goals and motivations, which can be broken down into sub-goals, which can then be elevated by attachment learning from Imprimers (since no discussion is given about how the starting goals are created). The A-brain could be largely pre-wired to see useful perceptual features as salient, and similarly, perhaps we start with some fundamental Critics looking for basic, common problem types in the A-brain, which become more finely attuned to the rest of the system as time goes by. However, at this early, disorganised stage, the internal credit assignment problem of which part of the system is responsible for success or failure is extremely hard. Moreover, since human programmers/scientists find it very difficult indeed to deal by hand with many small effects, might this be a stage where statistical methods will prove necessary?
The other half of the self-organisation problem relates to
learning at later stages. Minsky doesn�t seem to address the problem of how new
Critics are generated, as they surely must be to deal with new types of
problem, or to deal with the same fundamental problem types but at higher
levels or in different domains. A better theory of how the early system
self-organises would inform an understanding of how the later system could
re-organise. After all, presumably the same mechanism for generating new
Critics, Selectors and (maybe goals) would be the one that was involved in
building the higher reflective and self-reflective layers on top of the
instinctive and reactive ones. My speculative and half-formed suggestion would
be to emphasise the role of analogy-making within the mind between
problem-types in different domains and at different levels � new Critics could
be created by duplicating some currently functioning Critic and connecting it
to the Selectors and preconditions that correspond in the new situation.
Indeed, some enormous network of scouting agents working at a level below their
owner-Critics will be necessary to implement the current schema, since
something has to break down the preconditions that the Critics notice. In this
way, we might even have low-level Critics that aren�t connected to Selectors,
but are rather connected to higher-level Critics, and so on, until eventually
Selectors are activated[7].
Finally, I
would like to make one minor comment about naming. The Critics (unlike the too
similar-sounding �censors�) don�t necessarily inhibit or criticise activity so
much as redirect it, stimulating new cascades via the more
wide-effect-sounding �Cascaders�, or �Initiators�. As a result, Critics might
be better termed �Noticers�, �Remarkers�, �Observers�, �Informers�,
�Recognisers� or some other neutral term. Such a naming scheme would support my
initial reading of the chapter, where I ascribed a much more general role to
the concept of a Critic than merely passively noticing problem-types. On
this reading, they could be seen as sophisticated, probabilistic, interacting
if-then rules operating at every level in the Model Six architecture, i.e. it�s
tempting to see them as a lowest-level description of the Society of Mind in
general. If this was the case, Critics might be better termed �Ifs� and
Selectors �Thens�. I�m not now so sure that this is a useful way to see them
though.
In
conclusion, I feel that the Critic/Selector proposal raises issues that need
addressing further, but in principle, I�m persuaded that very similar
mechanisms must play a prominent role in any mind of sufficient power. This
because bolting on exceptions and inhibitions is the only way for a system to
build itself (i.e. learn) incrementally, and because I�m quite persuaded of the
necessity and value of heterogenous ways to think.
[1] Sloman (2001), �Varieties of affect and the CogAff architecture schema�.
[2] e.g. Pinker, How the mind works, chapter 6.
[3] Hofstadter, 1996, Fluid concepts and creative analogies, ch 5.
[4] Although the story is more complicated in later models, e.g. Metacat, indicating that this may have been a decision made for reasons of expedience.
[5] Again, this needs a caveat. We could see the low-level codelets that notice problems and continually evaluate the temperature in the background as being like low-level Critics.
[6] Norman, Ortony and Russell (2003), �Affect and machines design: lessons for the development of autonomous machines�, in IBM Systems Journal.
[7] Slightly off-topic, but perhaps a useful metaphor: this reminds me of the Signals and Slots mechanism in the Qt Gui framework. See http://doc.trolltech.com/3.1/signalsandslots.html