Re: Schyns Response R1-R4

From: Dearns Rachael (rld195@soton.ac.uk)
Date: Thu May 14 1998 - 16:54:43 BST


Darling, Andrea <ald295@soton.ac.uk> wrote:

> R1. Introduction
>
> Objects form categories because they share a number of
> features and differ on other features from contrasting
> categories. One essential function of a feature is therefore
> to subserve the categorization and representation of objects.

This is a simplified view of featuring in object
categorization.

> The target article questioned the origin of features. Either
> people come equipped with a complete, fixed feature
> repertoire that accounts for all present and future
> categorizations,
>
ad> (THIS IS UNLIKELY AS BABIES ARE NOT NECESSARILY AS
ad> ACCURATE AS ADULTS IN CATAGORISATION, EG. THEY OFTEN CALL
ad> ANYTHING WITH FOUR LEGS ^A DOG^.)

Yes, and in addition, it seems non-sensical that the brain
is so advanced in terms of development from birth, that it
could contain such a complete, fixed feature repertoire.
Also, this view seems to ignore the fact that every
individual's experience is uniquely different, and for the
fixed feature repertoire to be able to account for all
future categorisations, it could only this on the basis of
past experience, and natural selection: which cannot take
into account future categorizations that have not yet
occurred.

> or they sometimes create flexible features to subserve
> new categorizations.
>
ad> IT IS PROBABLY BOTH: SIMPLE, INNATE REPERTOIRES (EG. HOW
ad> TO TELL APART DIFFERENT COLOURS) MAY LATER BE
ad> SUPPLEMENTED BY THE ABILITY TO CREATE NEW ONES; AS SEEN
ad> IN THE ^MARTIAN BLOBS EXPERIMENT^. IT MAY BE INTERESTING
ad> TO SEE IF A SIMPLER VERSION OF THIS TECHNIQUE COULD BE
ad> PRODUCED FOR USE WITH INFANTS EG. BY LOOKING AT THEIR
ad> GAZE, DUMMY-SUCKING RATES ETC (TO SEE IF THIS TOO IS
ad> INNATE OR LEARNED VERY EARLY). HOWEVER THIS WOULD BE
ad> VERY DIFFICULT AS THE TASK IS COMPLEX AND BABIES CAN^T
ad> TELL US HOW THEY CATAGORISE THINGS, WE CAN ONLY ASSUME.

This latter view seems much more likely, and is supported
by Schyns et al, who also discuss the implications of this
view:
 
>> The target article examined the implications of the
>> latter view that flexible features are not necessarily
>> derivable from a fixed set of primitives, can augment
>> the representational power of the feature repertoire,
>> and can change how an input is perceived.
 
These implications mean that the creation of flexible
features hold a great deal more value than fixed feature
repertoires could. The fact that they strengthen the
representational power of the feature repertoire
illustrates the importance of perceptual categorisation.

> R2. Componential and Holistic object representations.

> from the outset, it is important to emphasize that the
> flexible feature stance adheres to the reductionist
> enterprise of cognitive psychology which prescribes that
> complex object representations should be reducible to
> combinations of their building blocks.

ad> YES IT DOES, BUT WHAT ABOUT THE HOLISTIC VIEW (EG. GESTALT).
ad> IN EVERYDAY LIFE WE OBSERVE THAT OBJECTS ARE NOT SIMPLY A
ad> COMBINATION OF FEATURES THAT ARE PROCESSED BOTTOM-UP.

A fundamental example is that of face processing, and the
extent to which features are processed holistically and
configurally, rather than being reducible to the processing
of a combination of the individual components that
comprise the face. Certainly, beyond childhood, research
has shown that the face is processed in a configural
fashion, in terms of the overall arrangement of features in
relation to one another; rather than employing a piecemeal
approach with the processing of individual features.

However, although Schyns et al do not dispute a
reductionist approach, they:

>> do challenge the possibility that all object
>> representations can be reduced to a fixed set of
>> building blocks.

Schyns et al claim to hold the position of being
'Compositionalists'; that is, they assume that features,
once created, become the building blocks of future
encodings and representations.

> The main alternative to the compositional view proposes that
> objects are represented holistically,without a prior
> decomposition
> into their components. For instance, Edelman discusses a
> system which learns the appearance of entire objects and
> uses their memory traces as holistic features. In a
> related vein, Burgund and Marsolek suggest that specific
> brain systems could be dedicated to the extraction of
> holistic features that represent familiar objects.
> Holistic representations have two main advantages over
> componential representations: (1) they preserve
> the input in an unprocessed form and (2) they
> efficiently represent the input by compressing multiple
> sources of information into a single feature. Holistic
> and componential representations can co-exist in the
> flexible feature framework. It is therefore an augmented
> compositionalism.
>
ad> THIS VIEW IS VALID AND PROBABLY MORE REALISTIC, BUT MAKES
ad> IT MORE DIFFICULT TO STUDY AND UNDERSTAND CATERGORISATION
ad> AS IT COMPLICATES THINGS ( AS IN EVERY AREA OF PSYCHOLOGY
ad> FROM MENTAL ILLNESS TO COGNITION).

This view has the overall advantage of allowing holistic
and componential represenations to coexist in the flexible
feature framework. This account has the best of both
flexible and fixed feature approaches, since:

>> Compositions of flexible feastures retain all of the
>> strengths of the classical 'mental chemistry' of the
>> fixed feature approach, and many object descriptions can
>> be generated from a set of combination rules.

> As already discussed in section 1 of the target article,
> compositions of flexible features inherit all of the
> strengths of the classical mental chemistry of the fixed
> feature approach: Many object descriptions can be
> generated from a set of elements and a set of combination
> rules; feature combinations allow for structured
> hierarchical representations; similarity relations
> between different objects can be expressed in terms of
> their features and their combination rules.
>
ad> YES, BUT THIS DOESN^T ALLOW SUCCESSSUL GENERATION OF ALL
ad> CATAGORY MEMBERS AND MAY LEAD TO MISTAKES. SOME CATAGORY
ad> MEMBERS (THE NOTORIOUS TOMATO) ARE MORE SIMILAR TO
ad> MEMBERS OF OTHER CATAGORIES THAN TO THEIR OWN. FOR
ad> EXAMPLE, A TOMATO IS A FRUIT (DUE TO IT^S PIPS, I
ad> THINK!) YET IS OFTEN CONFUSED WITH VEGETABLES AS IT
ad> DOESN^T GROW ON TREES (LIKE MOST FRUIT), IS EATEN WITH
ad> SALAD ETC.

It is important not to forget that this alternative is
still a composition of flexible features; which can
account for category members that are less similar to a
prototypical example, and that move the category closer to
others.
 
> R3.1 Chunking.
>
> Chunking and perceptual unitization (see 3.1.) were
> discussed in the target article as processes that could
> synthesize new features from a set of more elementary
> components. Chunking therefore implies that the input is
> segmented into features before being chunked--much in the
> way in which a capital T may be initially be discretized
> into a vertical and a horizontal bar before being chunked
> into a holistic T.
> Phenomenologically, the perception of the chunk does not
> entail the perception of its components. A chunk is
> therefore a new, isolatable and independent information
> packet of psychological processing that has the
> properties of holistic representations: The recognition
> of a chunk will not necessarily prime the recognition of
> its components, and vice versa; response times to the
> chunk will be faster than predicted by response times to
> the components (Goldstone, Steyvers, Spencer-Smith, &
> Kersten, in press); and similiarity
> relationships will not necessarily be perceived between
> the chunk and its parts.

> Good examples of chunking producing new features have
> recently been described in the category expertise
> literature. For example, Tanaka discusses the case of
> faces which are initially analyzed into their consituent
> parts (e.g., nose, mouth, eyes and ears) before being
> chunked into configural, holistic representations with
> the progressive acquisition of expertise in face
> categorization.

ad> THIS APPEARS VALID AS IN EVERYDAY LIFE WE ARE QUICKER TO
ad> RECOGNISE FACES WE KNOW IN THE STREET THE BETTER WE KNOW
ad> THE PERSON. WE APPEAR TO BUILD UP HOLISTIC PROTOTYPES OF
ad> FACES WE KNOW WHICH AIDS PROCESSING. HOWEVER, FEATURE
ad> PROCESSING IS STILL IMPORTANT UNDER SITUATIONS OF
ad> AMBIGUITY (EG. WHEN PERSON IS WEARING SUNGLASSES THEY
ad> DON^T USUALLY WEAR).

This is a good point, and certainly makes sense according
to personal, everyday experiences. Expertise with the
categorization of faces appears to develop to such an
extent that response times to the chunk will be a great
deal faster than predicted by response times to the
components, with an increasing difference depending on the
level of expertise of the individual with the face
concerned.
 
> R3.2 Feature creation: Form-from-medium rather than form-from-
> form.
>
> We want to distinguish the form of feature newness involved in
> chunking (which we call form-from-form) from another kind
> (form-from- medium) which directly produces features from
> a medium, not from other features. By analogy, imagine a
> Martian whose visual medium (the output of its
> transducers) is very much like dough. On the first day of
> its existence, the outside world imprints a teddy bear
> into the dough. Of course, this object and its parts are
> unknown to the Martian. He or she cannot compose a new
> representation of the modelled teddy bear from an
> already-existing representation of its component parts.
> Feature creation is a process which can directly imprint
> on the visual medium: It can cut around the teddy-bears
> silhouette and represent the entire object as a new
> holistic feature. Note that this cutting process only
> separates the entire bear from the medium. At this stage
> of conceptual development, our Martian represents the
> bear as a unitary, holistic feature and its
> decomposition into subcomponents is unspecified.
>
ad> GOOD POINT, BUT THE ANALOGY OF AN ALIEN MAY BE MISLEADING-
ad> WE ARE MOST LIKELY BORN WITH INNATE PRINCIPLES FOR
ad> BREAKING DOWN HOLISTIC OBJECTS, ESPECIALLY HUMAN FACES,
ad> INTO SMALLER COMPONENTS! THIS HAS BEEN SHOWN NUMEROUS
ad> TIMES IN INFANT STUDIES. AN ALIEN WOULD NOT HAVE THIS
ad> ABILITY-OR IF IT DID IT MAY ONLY SUIT IT^S NEEDS ON THE
ad> HOME PLANET.

Yes, but isn't this example simply meant to be an analogy,
to illustrate the overall process of feature creation,
rather than a representation of the specific capacities of
human infants?
 
> R3.4 What is it like to be a feature?
>
> The discussion so far distinguished between chunking,
> imprinting and emergent perceptual properties as
> different forms of feature newness. These distinctions made
> implicit assumptions on what it is like to be
> a feature that we now highlight.
>
> Features are isolatable information packets. This first
> property stresses that features represent compressed
> information from the input to which the system is
> sensitive. This information can be parts
> (as suggested in Tanaka and Singh and Landau), but it is
> not limited to parts. Colored blobs, textural
> elements, and many other information packets can qualify as
> features as long as the systems psychogical response to
> the packet reveals that it is a discrete, holistic entity
> in psychological processing.

ad> INDEED, A FEATURE CAN BE ANYTHING, FROM A PERSON TO PLANET
ad> EARTH TO THE SOLAR SYSTEM TO THE GALAXY- IT DOESN^T
ad> MATTER AS LONG AS IT MAKES UP SOMETHING ELSE....

Can such a generalised definition of features really be
valid - Scyhns et al give several criteria that must be met
for something to be defined as a feature: (that
features are isolable information packets,a nd are
independently perceived).

> Features are independently perceived. Isolatable features
> share the property of being independently perceived.
> As opposed to feature conjunctions for which the perception
> of each component must precede the perception of the
> conjunct, isolatable features are perceived without such
> prior perceptions. Independently perceived features can
> evolve both from a chunking process which produces
> configural features, from an imprinting process which
> creates a new form from the visual medium, or from
> dimensionality reduction itself.

ad> YES, BUT MOST FEATURES ARE NOT PERCIEVED INDIVIDUALLY: THE
ad> FEATURE IN IT^S ENTIRE CONTEXT NEEDS TO BE CONSIDERED;
ad> WHEN ISOLATED, A ^MOUTH^ IS ONLY A WONKY LINE, YET WHEN
ad> ADDED UNDER TWO EYES AND A NOSE IT BECOMES OUR WELL KNOWN
ad> MOUTH!

I agree strongly with this point..surely contextual factors
play a fundamental role in object categorization on the
basis of features.

> R4.1 Does feature creation require more flexibility?
>
> Continuous versus discrete feature learning in time. One
> possible interpretation of our proposal is that
> features created to solve a categorization subsequently
> crystallize and become fixed. To this, several
> commentators objected that we should instill more flexibility
> in the feature repertoire (Hahn & Chater; Tanaka;
> Williams, Gauthier & Tarr) because feature creation and
> adaptation is a never ending process. We agree with the
> idea that the feature repertoire should continuously
> evolve in response to the flux of new categorizations
> facing the organism. One extreme view suggests that every
> encounter with the objects of a class could lead to a
> small amount of feature-space reorganization (Williams,
> Gauthier & Tarr; Hahn and Chater). The issue of
> continuous variability raises an important problem that
> was not discussed in the target article: How does the
> system store information about feature variations?

ad> YES, AND HOW DOES IT HAVE THE ROOM!

It seems likely that the feature repertoire should involve
a dynamic process, and thus should instill a significant
amount of flexibility in the feature repertoire. This does
raise the problem as Schyns et al noted, of how does the
system store information about feature variations?

> One possibility is that each time the system sees an
> object, it would replace its old feature representation
> with the representation of the new features.

ad> ECONOMICALLY MORE VALID

> However, such system is liable to catastrophic forgetting.
> Successive exposures to rare instances of a feature could
> lead to a transformation of the feature that would not
> match common instances anymore.

There are clear advantages as well as drawbacks of this
possibility.

> Another possibility for
> representing variations is to apply the template matching
> approach to features and store each transformation as a
> distinct feature exemplar in memory. However, given all
> the variations in illumination conditions, pose and
> occlusion that features can take, exemplar-based systems
> would rapidly be overwhelmed with the millions of feature
> exemplars it would need to store and organize.

ad> YES, BUT EXEMPLARS DON^T NEED TO BE THAT SPECIFIC. FOR
ad> EXAMPLE, AN EXEMPLAR OF A MINI (EG. ROUNDED BODY, SMALL
ad> MINI LOGO, SMALL ROUND HEADLIGHTS ETC..I M NO EXPERT ON
ad> MINIS!) ALOWS US TO RECOGNISE MOST MINIS (EG. IN MOST
ad> ILLUMINATIONS, SITUATIONS ETC). WE WOULDN'T NECESSARILY
ad> NEED AN EXEMPLAR FOR EVERY DIFFERENT TYPE OF MINI IF THEY
ad> SHARE ENOUGH BROAD FEATURES (WHICH THEY GENERALLY DO).

This example, even though it improves on the first in terms
of removing the risk of catastrophic forgetting, still has
the clear limitation of containing too many feature
exemplars, which would also overwhelm the system.

> Thus we favor an abstract representation (for parts,
> think, eg. of a parametrized elastic template) which
> would very slowly adjust to local feature transformations
> to preserve the a desirable inertia in the continuous
> adaptation to changes.

This compromise position therfore seems the most logical
feature representation.

Dearns Rachael
rld195@soton.ac.uk



This archive was generated by hypermail 2b30 : Tue Feb 13 2001 - 16:23:21 GMT