SE367 Introduction to Cognitive Science

Mid-Sem exam Two Hours. 84 marks

Please answer all parts of each question in the same place.

Q. 1. Philosophy [4x4 = 16]
a) Idealism argues that there is no reality – there is only one kind of thing, and that is the mind.
You can construct quite a valid argument based on this. What is its main weakness?

If everything is inside the mind, and there is no outside reality, how can different minds agree
as to whether a ball will fit into a hole or about any physical aspect.

b) Describe the Chinese dictionary mental experiment of Harnad.

Supposing a foreigner in Beijing wants to read a sign. He has a Chinese-to-chinese dictionary.
If he looks up a sign, he gets more Chinese signs. In the end, the definitions are circular.
Therefore, any system of symbols, without external (sensorimotor) reference will inevitably
lead to circularity of definitions or infinite regress. Hence symbols must be grounded in
something outside of the symbol space - e.g. sensorimotor structures.

c) What is “strong AI” and how is Searle's argument directed against it? Why doesn’t it affect
“weak AI”?

Strong AI: Simulating a mind is the same thing as building the mind itself.
Weak AI : Simulation is a model, not really the same thing.

Searle’s Chinese room argument tries to show that a simulation executed by Searle does not
“understand” Chinese in the same sense as a mind, though it can follow complex rules on a
complex database and appear to be conversing in Chinese.

d) What is cognitive toil and semantic theft? Which is more used for human learning?

Sensorimotor toil: direct interaction with the environment to acquire concepts
Symbolic theft : using the grounded meanings of existing symbols to learn new definitions and

Human learning – for most concepts in the adult mind – is overwhelmingly based on symbolic
theft. In their paper Cangelosi and Harnad show that in multi-agent simulations of a foraging
group, agents that learn from the utterances of others (theft) have some advantages over
those that learn directly - and that a population ultimately converges to nearly all members
adopting theft.

Q2. Tacit knowledge and Expertise [3x4 = 12]
a) Around the turn of the century, Edouard Claperede conducted an experiment with a pinprick.
How is this relevant to tacit knowledge?

Claperede was working with severely amnesic patients, who could not remember events even
immediately afterwards. With an elderly woman patient, he would have to introduce himself
every day. Once he concealed a pin in his palm as he shook her hand. Next time she was
reluctant to shake his hand, though she had no recollection of ever having met him or of the

b) What is a “chunk”? How is it related to a chess grandmaster’s expertise?

A chunk is an encoding of a pattern that is commonly associated with good (or bad)
performance on a function. By encoding (building models for) such chunks, a person can
reduce the conscious overload in many tasks. In chess, this enables the grandmaster to
effortlessly remember thousands of board games and results.

c) Discuss if this is correct: It is efficient for much of expert knowledge to be subconscious.

d) Conscious deliberation is slow and inefficient, owing to the extreme high dimensionality of the
input and decision spaces. Chunks encode the “good” answer regions in this mapping, and
automaticity enables us to execute these chunks effortlessly or fast decision making. Even
where conscious decisions are needed, they can work with these subconscious processes
either before or after to provide solutions.

Q. 3. [Attempt 10 out of these 12 questions. Pls do not attempt more than 10] [ 40 marks]

I. [Magnitude estimation]
a) It was well-known that musicians are better at temporal discrimination. What
position is argued for by [Agrillo etal 2012] when they also show that they are
also better at spatial and large number estimation tasks?

In 2003, an argument was made by Walsh based on neurological evidence that spatial,
temporal, and number estimates were all computed in the same region of the parietal cortex.
This bolstered earlier ideas that these three systems were related. However, no direct tests
were done for this. In this work, Agrillo et al attempt to establish that musicians, who are
known to have a better temporal sensitivity, are also better at spatial and numerical tasks,
thereby supporting the unified magnitude theory.

b) Walsh 2003] suggested the ATOM theory - what does ATOM stand for?

ATOM = A theory of magnitude. Though it was known that spatial and temporal magnitudes
were correlated, it was thought that countable numbers may be represented somewhat
differently from continuous quantities. However Walsh argued that these functions were
integrated in the parietal area so as to provide a faster sensorimotor response.

II. [Computational image description]
c) In [kulkarni & berg]’s image descriptions paper, for each object classified, there were 21
attributes. During text generation phase, how was it determined which attributes would be
generated and using what text?

The outputs of the classifiers served as nodes in the CRF. The main computation of the
attribute-object association was done through a potential function which used a large set of text
descriptions from Flickr (obtained by posting the attribute-object pair as query) to learn the
object-attribute potential map. These were later smoothed with an internet corpus for the same
keywords, obtained via google.

d) were spatial relations such as near learned from a training set or hand-coded?

They were hand-coded. Spatial relations were hand-coded. E.g. A near B was defined as a
function of a ratio - the distance between the centroids of the enclosing rectangles, divided by
the size of A.

e) in the erroneous captions, cite examples if you can to discuss which errors are most serious - i)
errors in object class, ii) errors in attribute, or iii) errors in relations?

Errors in object class detection, such as “furry road” for a cat, or “one person” where there are
none, are perhaps most disconcerting for humans. Spatial errors are not perceived that
critically. Attribute errors can be problematic – e.g. when “four persons” are counted for three,
but these are not as serious, at least in my view, as object errors.

III [Collaboration and Gaze]
f) what is the set up of the collaborative tangram puzzle task used in [kuriyama etal]? how is
gaze evaluated with respect to the other’s statement?

One subject can see the goal shape to be assembled, but can’t move the pieces; the other can
move the pieces but does not know the goal. Now A gives instructions to B to solve the
puzzle. These instructions contain specific references to objects. As A starts to utter the word
that refers to an object, the gaze or listener B and speaker A tend to converge – at least for the
successful trials.

g) what is "common ground" and how is it related to gaze?

Among two collaborating people, common ground is a shared understanding of the context for
the task. In tangram, it may be an understanding about which piece is to be moved where.
One test or this is that shared common ground results in shared gaze.

h) what may be the reason for gaze matches occurring earlier when a pronoun was used in the
description rather than a noun phrase?

NPs resulted in gaze match typically 2 seconds after the onset of utterance, whereas gaze
match was practically instantaneous for pronouns . Two possible explanations: i. pronouns
are more likely to be used only if the object is more salient (otherwise it may be hard to
disambiguate). Ii. Pronouns are shorter and are uttered more quickly in the speech .

IV [Attributing false beliefs]
i) It was been long known that children younger than 4 are not able to attribute false beliefs to
people who leave the room and come back while an object has been moved. What evidence
do Baillargeon et al present for an earlier age at which this is partially understood?

Baillargeon, working with Onishi and later with Song, showed that infants as young as 14
months look longer when such an adult demonstrates behaviour inconsistent with his
knowledge, which results in the correct reach. Yet when asked about which box the person
would go for, they give the wrong answers till 4 years or beyond.

j) What neurological processes may account for the failure of younger children in the false belief
task, despite being aware of it?

Though infants have some awareness of what actions are consistent with what the other
person knows, they may not be able to a) represent the fact of the others’ belief, b) compute
the right response, and c) inhibit their own awareness of the right action.

Baillargeon suggests a possible neurological process that may account for this. The region of
the brain where the false belief may be encoded is in the temporal-parietal boundary –
whereas the response areas are in the pre-frontal area. Connections between these two areas
are slow to mature, and this may account for the delay.

V. [Mirror Neurons]
k) Might mirrror neurons have some role in observations that action production primes action

Mirror neurons are in the pre-motor area, and fire when the agent does the action. It is known
that these areas backproject to the visual areas – particularly the superior temporal sulcus or
STS, where action recognition is thought to occur. Thus, doing the action may enhance
stimulation in the recognition areas thus helping in priming.
[a stimulus primes a later response if it is related in some way.]

l) Do you agree: Mirror neurons that differ based on the view of the action may have a role in
discerning intentionality.

Caggiano etal show that different neurons in the pre-motor area fire while observing differing
views of the same action. These view-dependent neurons thus inform the system whether the
action is occurring by them in their sensorimotor space (0 deg), across from them (180 deg), or
beside them (90 deg). This would help discriminate whether the action is being done by them
or by some other, and therefore intentionality of the other.

Q4. (Shorts) Explain the following terms and how they are used: (2x8=16)
i) Subitize

Subitize = Knowing the count without counting. A mechanism used for smaller numbers.

ii) VOE

VoE = Violation of Expectation. Used to assess an infant’s mental models of situations. If
they expect outcome A they will look significantly longer at B.

iii) Geon

Geon theory : 3D shapes are constructed out of volumetric units called geons (e.g.
sphere, cuboid, bent cylinder, etc.) .

iv) SNARC effect

SNARC effect : the right hand is faster to respond to addition tasks or tasks involving
somewhat larger numbers.

v) Allocentric

Allocentric view = view from outside, other’s view. (as opposed to “egocentric” – self

vi) Place cell

Place cell= neurons in the cerebellum that fire when the organism is in a specific location.

vii) F5

F5 = area in the frontal cortex (pre-motor area – just next to the motor area beside the
central sulcus). One of the sites for mirror neurons

viii) A-not-B

A-not-B error: Infant is habituated to toy being in box A. Now, in front of infant, toys is
shifted to B. But they persevere in looking for it in A. from abt 5 to 8 months,