Greg Detre
Sunday, March 30, 2003
cognition � �interpretet, understand, reflect upon, and remember things about the world�
affect � �rapidly evaluates events to provide an initial assessment of their valence or overall value with respect to the person�
�positive or negative, good or bad, safe or dangerous, hospitable or harmful, desirable or undesirable etc.�
they mention Damasio�s evidence that patients with neurological damage to their affective systems are unable to operate on a day-to-day basis, despite appearing to perform normally on standardised cognitive tests
affect as essential for intelligent behaviour by altering goal priorities and generating interrupts
they�re thinking about systems that have a variety of tasks + goals, must run attended and autonomously, and need high reliability
three levels:
reaction, routine, reflection
processing at each level serves both a cognitive and affective function
higher levels: greater depth and slower processing
control information (activation/inhibition) flows downwards
Reaction
lowest level, perhaps genetically determined in animals
immediate responses to state information coming from the sensory systems � function is rapid reaction to the current state
fast, hard-wired detectors
interrupts higher-level processing
responses at the reaction level can be potentiated or inhibited by inputs from these higher levels, and they can habituate, reducing sensitivity to expected signals
Routine
skilled and well-learned, largely �routinised� behaviours
this is the home of most motor skills, incl language generation
considerable processing to select and guide behaviour
control signals from above (inhibition and activation), it can both inhibit and activate reaction level responses and can pass affective information up to the reflection level when confronted with discrepancies from norms or routine expectations
performs assessment � values on three dimensions
positive
negative
(energetic) arousal
default expectations � when these aren�t met, the system can make adjustments and learn
Reflection
mind deliberating about itself � performs operations upon its own internal representations of its experiences, its physical embodiment, its current behaviour, and the current environment, along with the outputs of planning, reasoning and problem-solving
input only from lower levels and neither receives direct sensory input nor is capable of direct control of behaviour
interrupts from lower levels can direct and redirect Reflection-level processing
don�t see how this fits with the downward-only control???
negative affect
vigilance
especially with high arousal appears to lead to more focused an deep processing
positive affect
curiousity
broad, more widely spread processing
humans have more enhanced creativity in a pleasurable state
current systems do not distinguish between affect (evaluation) and cognition (understanding)
the 3 levels in artificial systems
alarms
deviations from norms
restructuring queues, priorities or resource allocation
e.g. virtual memory in modern OSs
considers the example of a RAID system that starts to take precautions, and prioritise extra reliability over performance when it becomes �anxious�
�anxiety� here amounts to the system�s appraisal of its own trustworthiness
important question:
why are they proposing the introduction of affect as a whole new system, rather than just writing appropriate algorithms for each potential problem?
strong methods:
exploit specific domain knowledge + structure
weak methods + heuristics:
designed to be much more general, making them slower, less efficient and not guaranteed to succeed � trade efficiency for generality (e.g. hill climbing)
strong methods are always preferable when you know the situation,the environment is predictable and limited in scope � otherwise you have to rely on weak methods (just as biology employs (in combination with strong methods, e.g. hard-wiring) to cope with an uncertain, world) � affect is just such a weak method
only 3 dimensions of affect???
where do they fit with Ekman�s 6???
is there a way of breaking them down further, or are they orthogonal to human emotions???
cf autonomic computing� references???
isn�t an entirely downward control information flow very rigid/awkward??? how does it deal with interruptions???
3+ reasons for the interest in computer models of emotions
1. an interest in emotions s something to be modelled + explained
2. giving machines an understanding of emotions in order to make them better for HCI
3. new kinds of computer entertainments that employ convincing emotional behaviour
architecture-based concepts:
�starting with specifications of (virtual machine) architectures for complete agents and then finding out what sorts of states and processes are supported by those architectures�
e.g. thrashing and deadlock (i.e. time spent swapping and paging virtual memory) as architecture-based concepts
�defined in terms of causal interactions between states and processes within mechanisms in a virtual machine architecture, and in that sense they involve a functional perspective�
in contrast with philosophical functionalism, which �defines mental states in terms of the relationships between inputs/outpus of the whole system without mention of the internal architecture�
various definitions of emotion emphasise:
brain processes
peripheral physiological processes
patterns of behaviour
eliciting conditions
functional roles
introspective qualities
some include all motives + desires (e.g. hunger and curiosity), others include surprise (or treat as just a �cognitive state in which a belief expectation has been found to be violated, which may or may not produce an emotional reaction�)
emotions are rich in semantic content + direction (e.g. �being angry with a particular person about a particular action performed by that person etc.), as opposed to continuously variable global state
should not necessary represent intensity of emotions as single numerical variables � the intensity may be emergent, like the degree of thrashing
�important not to assume that the forms of representation that are useful � to use when describing a complex system or predicting its behaviour are to be found in the system itself�
ways to deal with these conceptual confusions:
ignore
seek operational definitions of various states in terms of measurable aspects of behaviour
do surveys of linguistic usage
study the role of emotions in literature
conceptual analysis � iterations of conjecture, testing, modification
try and define different emotions in terms of their biological functions (and adaptiveness)
but some, e.g. grief + embarrassment might be emergent states
CogAff schema
pillars
perception
central processing
action
layers
reactive mechanisms
deliberative reasoning
meta-management (reflective processes)
alarms
purely reactive and pattern-driven, stupid, capable of mistakes, may be trainable
the nature of each component in the 3x3 grid is defined by its functional connections to all the others
architecture-based concepts (e.g. of pain) make the concepts precise, and allow questions to be formulated
reactive systems
lack the ability to represent, evaluate and compare possible actions + their consequences (i.e. counterfactuals, relativity of time + place)
combinatorial explosion
normal vs threat/opportunity instances � proto-emotions
where different kinds of needs dominate innate processing � proto-motives
more global states, less goal-directed but change the quality of processing in some general way � proto-moods
deliberative
able to �represent, analyse, compare, evaluate and react to descriptions of hypothetical future scenarios or possible explanations of previously observed phenomena�
use structured representations with a compositional semantics
it may not be possible to tell reactive proto-caution from deliberative proto-caution from the outside, but only by looking at the internal processing architecture
meta-management
�self-observation or self-monitoring of a wide variety of internal states, along with categorisation and evaluation of those states, linked to high-level mechanisms for learning and for controlling future processes�
higher-order emotions
depend on which layer they relate to
virtual machines???
i.e. multiply realisable???
swapping vs paging???
the difficulty of categorising surprise (and indeed the motivational states) is that they can�t be cleanly differentiated from cognitive
surely, different numbers of layers exist(ed) in different modules, depending on complexity, need + ecological niche � they needn�t/wouldn�t have emerged complete/distinct
is there a clean distinction between reactive + deliberative???
both need a kind of LTM
hypothetical scenarios though�
how is CogAff a space of architectures???
he used �qualia�! presumably just to mean sense-data though�
more than 3 (levels???)??? nonsensical to abstract this far???
ignores the �hard problem�
do Sloman�s levels correspond to Norman et al.�s 3, and do they correspond to Minsky�s A-, B- and C-brains???
what do Minsky�s extra 3 add???
look at Minsky�s 6-level model
look at �what is it like to be a rock� again
look up references
when a stranger rebukes us, or when we experience negative reinforcement for some action, we alter our methods
pride + shame � �attachment-based learning�
attachments help us learn new kinds of goals
�in the face of a parent�s blame or reproach she learns that her goal was not good to pursue�
�when we�re praised or rebuked by the people we love, we don�t just feel pleased or dissatisfied; instead, we tend to feel proud or ashamed�
the way we have evolved to deal with living in complex communities that potentially vary enormously (in climate, culture, language, lifesteyle etc.) is to rely less on survival skills that we possess from soon after birth, and more on learning from those around us
cf reputation-based appraisal systems, which have to solve the similar problem of sifting through an enormous amount of data in real-time, evaluate it coarsely then at a more fine-grained level, but using algorithmic processes � you start by conjecturing who�s worth listening to (moderators), then you pay attention to their evaluations, then you reevaluate your moderators
for this reason, our parents� and elders� views are even more important, not just because we are so dependent on them as children, but because of the need to �download� their wisdom
other theories: smaller heads to facilitate childbirth, or that huge brains just take longer to learn � he thinks we mature in well-spaced stages to make multiple levels of representation
�it is one thing to learn how to get what you want � and another, to learn what you ought to want�
the popular theory about children building their characters during their formative years doesn�t answer:
to whom do our children become attached?
how does attachment help establish our values?
what is the span of those �formative� years?
when, if ever, do we outgrow them?
Goal-Cloud model � when a goal is satisfied by some action, then you connect the goal to the sub-goal of achieving that action (or avoiding some action that didn�t work)
but how do you learn new high-level goals, i.e. values?
cf Aristotle rhetoric quote about shame
doesn�t like �caregiver� as a term for people you�re attached to, because the attachments can form without physical care
Imprimer: a child�s Imprimer is one of those persons to whom that child is attached
Impriming: a special way to learn new goals that works when a child�s Imprimer is present
he�s happy to admit that there�s no neuroscientific evidence for Impriming, but that there hasn�t been much of a search for it
pride from Imprimer praise elevates her present goal to a higher kind of priority
talks about controlling the �level of detail�
3 types of learning when Carol plays with mud:
aversion learning � when the stranger scolds her, she learns to avoid such situations
attachment learning � devalues her current goal when scolded by mother
subgoal (�reinforcement�) learning � when succeeds by using an action, the child learns that this is a useful subgoal
learning is a suitcase word
which aspects of what hse happened to do should take credit for her final success?
�Which features of those recent events should
Carol�s brain decide to record?
Should she record where they occurred, or which other
persons were present?
Should she remember which shoes she wore or whether
the weather was cloudy or clear?
Which of the thoughts she was thinking then should
be included in that description?
Where will she store the records she makes�and how
will she later retrieve them?
How will Carol represent all those kinds of information? (See Representations.)�
feelings �reflect our attempts to describe complex cascades of reactions�
conscience, ideals values: when we use those Imprimer-based emotions on our own to evaluate unfamiliar goals
rants about mysticism, and debunks other theories about where our moral sense comes from
Bowlby: "The variables which determined most clearly the figures to whom the children would become attached were the speed with which a person responded to an infant and the intensity of the interaction in which he engaged with that infant."
the presence and praise of an Imprimer has come to serve as an �innate releasing mechanism�
(Tinbergen) for a special system we call the �attachment elevator�
discusses Imprimer problems
having more than one Imprimer, may become muddled, inconsistent (see point re trusting yourself below)
the danger of having your goals altered by strangers
discusses cult membership
�we must be wary of changing ourselves in ways that prevent us from ever changing again
Bowlby argued that attachments are mainly for physical safety � as opp to food provision
there�s no point being able to put together a 10-step plan unless you�re going to have the future self-consistency to carry it out
��the more you�re able to trust yourself, the more you can simplify yourself � until you become your own caricature� - hmm
see self-caricature point
tells the story of Arthur Samuel�s Checkers program that had to have its learning turned off when playing worse players � that�s because it didn�t know enough about what/why it�s learning�
discusses celebrities
rhetoric � �an oration that has the right kinds of timing can seem like a �virtual interaction� by entraining some of the listener�s thoughts, by first raising questions in their minds � and then swiftly and aptly answering them�
the self-caricature bit is laid on too thick
�In the course of each person's development, they tend to evolve certain policies that are so systematic and consistent that we (or our friends) can recognize them in the forms of what we call features or traits�and these become parts of our self-images. Then when we formulate our plans, we can use those traits to predict what we�ll do (and thus reject schemes that we know we won't do). Whenever this works we�re gratified, so then we continue to train ourselves to behave in accord with these simplified descriptions. Thus, over time our imagined traits proceed to make themselves actual!�
does he explain how the species creates new ideas???
an element of trial and error, accident, fortuituous misassignment of credit, I spose
For instance, while chimps may be
able to discuss counterfactual situations in � , and that this allows us to
perform some meta-management level thinking about, say
cf reputation-based appraisal systems, which have to solve the similar problem of sifting through an enormous amount of data in real-time, evaluate it coarsely then at a more fine-grained level, but using algorithmic processes � you start by conjecturing who�s worth listening to (moderators), then you pay attention to their evaluations, then you reevaluate your moderators
�Inborn, Instinctive
Reactions: Joan hears a sound and turns her head. Many infant animals do such
things; they�re born with just enough �instincts� to help them survive.
Learned Reactions: She sees a
car. Joan has learned that certain conditions demand new, specific reactions to
them. How does she manage to learn such things?
Deliberative Thinking: She is
thinking about what to say at the meeting. Here she imagines alternative
futures, and various ways to choose among them. What kinds of resources might
work in her brain, to envision and reason about such things?
Reflective Thinking: Joan
reflects upon what she has done. To do this Joan must use some internal paths
through which her resources can �react� not only to her environment, but also
to traces and records of things that recently happened inside her brain.
Self-Reflective Thinking:
Uneasy about arriving late. Some monitor watches Joan's temporal progress, and
insists that she must not delay her decision�or something bad might happen to
her.
Self Conscious Emotions: �What would my friends have thought of me?� Now Joan thinks about what she ought to have done. Did her actions live up to her self-image of how she should have or ought have behaved?�
Differences
role of alarms
how integrated the affective component is
how much the low-level does
whether you divide up the high levels
Gregory discusses the importance of allowing top-down and bottom-up to intermingle and pressure each other
For instance, while chimps may be
able to discuss counterfactual situations in � , and that this allows us to
perform some meta-management level thinking about, say
cf reputation-based appraisal systems, which have to solve the similar problem of sifting through an enormous amount of data in real-time, evaluate it coarsely then at a more fine-grained level, but using algorithmic processes � you start by conjecturing who�s worth listening to (moderators), then you pay attention to their evaluations, then you reevaluate your moderators
see notes � curious machines project 1