Jean-Marc Fellous & Michael Arbib (eds.), Who Needs Emotions?, Oxford U. P. 2005. Chapter 12. [Comments in square brackets are my own thoughts.] Ch. 12: BEWARE THE PASSIONATE ROBOT (p. 334-383) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ Michael Arbib (USC) This is a wide-ranging chapter, with detailed commentaries on many of the earlier chapters (esp. Ortony et al., Rolls, Fellous & Ledoux), retrospectives on relevant past work in neuroscience, and his own ideas about animal, human, and robot emotion (& briefly, consciousness), taking a story in which he lost his temper as a case-in-point for studying human emotion (and for being cautious about endowing robots with human-like traits). As one might expect, he reaches few specific conclusions about the nature of emotions, but the thoughts strewn along the way are interesting. What I particularly noted was his concern with language: "I would argue that we are fully conscious in the human sense only because we have language" (p. 354). The most interesting suggestions made are 1. that language evolved via mirror neurons/ action-imitation/ pantomime; the "mirror system" is symmetrical with Broca's area; 2. that the function of emotion, abstractly, is to bias the agent towards rapid commitment to certain action strategies, in favor of more "cautious" appraisal and choice of strategies (p.376). In this abstract sense, we may equip robots with emotions as well, but aiming their activity more narrowly than human activity. (E.g., they will probably be produced in factories, rather than being self-reproducing.) We probably should not base robotic emotions in this abstract sense on any analog of the human basic drives, which lead to the "heated" and potentially risky form of emotions (e.g., losing their temper). Arbib (unlike some of the authors of other chapters) is a very good writer, though his style tends to be "stream-of-consciousness" writing with the structure of the argumentation remaining implicit and discontinuous, and he too tends at times towards "jargon- bombardment" (a temptation for any expert in a scientific domain). (A prime example: "In the mammalian brain, dopamine is contained in specific pathways, which have their origins in the substantia nigra pars compacta and the ventral tegmental area of the midbrain and ascend to innervate widespread regions of striatal, limbic, and cortical regions such as the striatum, prefrontal cortex, amygdala, and other forebrain regions." p.366) But he takes seriously the commonsense notions of emotion, passion, action, etc., trying to allow for them theoretically, and uses vivid examples. He critiques several earlier chapters and tries to relate ideas from them to his own comprehensive perspective. A cautionary tale (p.333-341) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ His story is about getting angry at schedule changes for a meeting with a (presumed) higher-up, and almost damaging a professional relationship with his impetuous reaction. He draws conclusions about how various emotions (such as annoyance, disappointment, pleasure, and shame) are triggered and dissipated. The observations are fairly obvious informal ones, such as "an emotional state may be 'terminated' when a plan is completed which addresses the source of the state; or it may simply dissipate when alternative plans are made. However, if some major goal has been rendered unattainable by some event, the negative emotion associated with the event may not be dissipated by embarking on other plans, since they do not approach the goal." To look for the further analysis he promises here, let's skip ahead 20 pages. An evolutionary approach to heated appraisals (p.360-70) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ He eventually (p.361) argues that Ortony et al. leave out a crucial facet of emotion -- its "heat" -- when they set aside visceral sensations, heart rate, the endocrine system, facial expressions, etc. [do they? Certainly their 3 levels allow for "interrupts" from "protoaffect" to higher levels]. He thinks the "heat" comes precisely from the more basic motivational system, which in humans is interconnected with the cortical systems for evaluating situations. He says much (p.361-366) about behavior control columns in the brain stem, the amygdala, orbitofrontal cortex, etc., in the functioning of emotion, referring to the other chapters, but none of this seems to address the "cautionary tale" in any of its specifics (except for the noted "heat"). He also says much about "neuromodulation" -- the broad and varied effects on brain systems due to variations in levels of dopamine, serotonin, and opioid peptides (endorphins, enkephalins, and dynorphins). He again speculates about the relation of these systems to "heat", and whether this evolutionary product would be at all appropriate for robots. Finally in this subsection he discusses consciousness and empathy. He really doesn't provide any theory of consciousness, just discussing various brain systems that seem to be involved, and arguing against the theory of Fellous & Ledoux that the content of consciousness is the content of working memory; he points out that some parts of the prefrontal cortex that are thought to be involved in working memory are not involved with consciousness. His own view, as mentioned at the outset and again below, is that consciousness crucially involves language -- though he never relates this claim to the concepts of consciousness (phenomenal consciousness, access consciousness) discussed by philosophers. Concerning Rolls, he thinks Rolls overemphasizes reward & punishment, important as these notions are for learning and adaptation, because basic drives, and a wide array of perceptual and motor systems are involved in emotional behavior (p.369). [Note that basic drives don't necessarily involve reward/punishment expectations. An infant suckles because that's part of a preprogrammed feeding drive, not because it expects rewarding feelings from the activity (though these of course occur as well); and an animal or human reacts angrily and aggressively to a rival for a mate, not because of fear of bad consequences (punishments) or loss of rewards (sex), but again because that is just a basic drive.] He also questions Rolls' claim that evaluation and correction of plans implies metalevel abilities -- he thinks "if ...then" rules are not language-like and don't necessarily involve language areas (like Broca's area). [I think some planning is unconscious -- e.g., choosing your path when crossing a room containing tables and chairs, while more challenging planning (e.g., planning a vacation, or in a risky situation) may involve conscious deliberation.] In commenting on empathy, he relates this back to the "simulation theory" of how we comprehend other minds (including emotions). He makes a remark about the evolution of empathy in relation to the evolution of language, but then sets this topic aside. His overall conclusion is just that all sorts of thinking in general involve both cognitive and emotional components [which to me seems pretty much consistent with what everybody else says, and not very illuminating concerning the "cautionary tale"]. Evolution of vision and language (p.344-354) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The idea is that - vision evolved from direct linkages between perceived features (moving dots, looming objects) to action (snapping, flight) to action-neutral representations of visual contents (in the visual cortex), allowing cognitively guided action; - but the action-oriented pathways evolved into a separate "how" system for performing actions, feeding directly into the premotor cortex (and this accounts for the capabilities of some "blindsight" patients); this action-oriented pathway computes "affordances", i.e., parameters like grasp width to grasp a particular object, based on a metric analysis of the object; a full diagram of both (combined) pathways to action is shown on p.351) - the evolution of manual dexterity aided by visual monitoring of the manual activity eventually led to a tendency to trigger the same neural circuits when observing *another* primate's actions as when carrying out the actions oneself: MIRROR NEURONS; these are symmetrically opposite Broca's area! - this in turn paved the way for communication by pantomime -- gesturing to indicate the action to be undertaken (allowing the observer to anticipate that action as well) - the pantomimes became more complex, and then started to be accompanied by vocalizations (these are not the same as primate calls, which involve brain regions other than Broca's area). - in this way language evolved, with specialization of Broca's area for that communicative function. [See the appendix at the end of this write-up for Arbib's own summary of the above evolutionary developments, in the form of an abstract for a talk he gave here in 2005.] [There's no mention of the role of Wernicke's area here, involved in speech understanding. Also what I was really missing here is any mention of the "language of thought" that must have developed concurrently with or prior to the development of linguistic communication. Many anthropologists say that symbolic thought and language appear to have emerged concurrently in homo sapiens.] At this point (p.354) Arbib stresses the role of consciousness in language, while at the same time he "strongly denies" that consciousness is "merely a function of language". He says we can be keenly aware of visual qualities and subtleties without being able to put them into words. [However, it's unclear how he reconciles these 2 claims. If language is crucial for human-like consciousness, how can I consciously appreciate "the shading and coloration of a face in great subtlety" *without* involvement of language?] From drives to feelings (p. 355-360) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ He starts by claiming to have discussed "awareness of both internal and external states" [did I miss something?]. He then discusses basic motivations (the 4 F's) and the toad (again). He discusses neural correlates of behaviors, but emphasizes that motivation should not be equated with behavior (such as fleeing or feeding) but rather that behavior is a reflection of the influence of basic drives. [sounds uncontroversial]. For rats, the connection between drives and actions is more indirect. He refers to his 1977 "world graph" model (with Lieblich) of an animal's knowledge of the world. This models "the process whereby the animal decides where to move next" (they seem to be mostly concerned with spatially guided behavior (rats in mazes?). The world-graph model, in its most recent form is shown on p.359. This indicates how various brain regions and systems feed into the world graph, which is presumed to reside in prefrontal cortex. [NB the references on pages 360 & 368 to Fig 12.1 are actually references to Fig. 12.2.] This reader couldn't get much insight into emotion from this section. One interesting point (p.360) is that the hippocampus is presumed to provide "a representation of context", especially the animal's location in space ("I am here"). Robot emotion (p.371-377) ~~~~~~~~~~~~~~~~~~~~~~~~~ He asks whether emotions are a side effect of a certain cognitive complexity (p. 373), or a byproduct of evolution that we could choose to impart -- or not impart -- to machines. He seems to lean towards the latter position, in particular, that a robot might be endowed with just appraisal capabilities (oriented toward whatever functions the robot was designed to serve), rather than "heated" emotions (as mentioned at the outset). These appraisal capabilities might be guided by emotions in the abstract sense: under certain conditions the robot's behavior may be biased towards rapid commitment to certain action strategies, in favor of more "cautious" appraisal and choice of strategies (p.376). But this should be done within the functional "niche" of the robot, and customers will demand this. [Well, if development of robots *smarter* than people is inevitable, as many AI researchers think, it's unclear how we could anticipate the effects of built-in constraints of any sort. I think we just have to ensure our machines are inherently benign, wishing to benefit rather than harm or coerce humans ...] [A final stray thought was prompted in connection with Arbib's discussion of rats that kept pressing a lever to stimulate a part of their hypothalamus, where an electrode had been implanted: If we build robots with a reward/punishment system, we may find that they'll discover how to access the reward variables internally and simply boost them! After all, internal access to data structures in computers is much easier than in brains!] Appendix to these notes: an abstract of a 2005 talk by Arbib ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ Dr. Michael Arbib, "Fletcher Jones Professor of Computer Science and Professor of Neuroscience" University of Southern California "How The Brain Evolved So We Could Have Language" The Mirror System Hypothesis (MSH) on the evolution of the language-ready brain posits several stages: action recognition grounds imitation of pragmatic actions of the hands, this in turn grounds pantomime as a new form of communication based on displacement of hand gestures to mimic other degrees of freedom; the repertoire is then extended as familiar actions are recursively varied and combined. Conventional codes then supplement and increasingly supplant pantomime to yield protosign, a combinatorially open repertoire of manual gestures, which provides the scaffolding for the emergence of protospeech (which thus owes relatively little to non-human vocalizations), with protosign and protospeech then developing in an expanding spiral. It is argued that these stages involve biological evolution of both brain and body. By contrast, it is argued that the progression from protolanguage (protosign and protospeech) to languages with full-blown syntax and compositional semantics was a historical phenomenon involving few if any further biological changes. [Among Arbib's many books, some of the best-known are Brains, Machines, and Mathematics (1987), An Introduction to Formal Language Theory (1988), The Metaphorical Brain (1989), and The Handbook of Brain Theory and Neural Networks (2002).]