Unit 2: the blind men and the elephant
the blind men and the elephant
the blind men and the elephant
Four case studies — four approaches to understanding the brain, and their limitations:
-
An ambitious broad-scope "computational" theory
(J. R. Anderson. ACT: A simple theory of
complex cognition. American Psychologist, 51:355-365, 1996).
-
A single-cell electrophysiology study (C. D. Salzman, K. H. Britten, and
W. T. Newsome. Cortical microstimulation influences perceptual
judgements of motion direction. Nature, 346:174-177, 1990).
-
An imaging study (J. V. Haxby, M. I. Gobbini, M. L. Furey, A. Ishai,
J. L. Schouten, and P. Pietrini. Distributed and overlapping
representations of faces and objects in ventral temporal cortex. Science,
293:2425-2430, 2001).
-
A "big data" / "deep learning" hack (V. Mnih et al., Human-level control through deep reinforcement
learning, Nature 518:529-518, 2015).
[NOTE: I did not include an example of a purely behavioral study
("experimental psychology") — criticizing something like that would be
like shooting fish in a barrel.]
case study #1: ACT-R (John Anderson)
"The ACT-R (Adaptive Control of Thought, Rational) theory rests upon three
important components: rational [= "effective + adaptive"] analysis
(Anderson, 1990), the distinction between procedural and
declarative memory (Anderson, 1976), and a modular
structure in which components communicate through buffers."
"The actual substrate of cognition is an interconnected network of
neurons. Whether or not this is a significant source of constraint is open
to debate."
— Taatgen, N.A. & Anderson, J.R. (2008).
ACT-R.
In R. Sun (ed.),
Constraints in Cognitive Architectures.
Cambridge University Press, 170-185.
ACT-R: a "simple theory of complex cognition"
"All that there is to intelligence is the simple accrual and tuning of
many small units of knowledge that in total produce complex cognition. The
whole is no more than the sum of its parts, but it has a lot of parts."
— Anderson, J. R. (1996). ACT: A simple theory of complex
cognition.
American Psychologist, 51, 355-365.
ACT-R: a "simple theory of complex cognition"
"All this knowledge creates a serious problem. How does one select the
appropriate knowledge in a particular context?"
"ACT-R has developed a two-pass solution for knowledge deployment. An
initial parallel activation process identifies the knowledge structures
(chunks and productions) that are most likely to be useful in
the context,
and then those knowledge structures determine performance."
— Anderson, J. R. (1996). ACT: A simple theory of complex
cognition.
American Psychologist, 51, 355-365.
ACT-R as a theory of cognition
"ACT-R is a hybrid architecture in the sense that it has both symbolic and
subsymbolic aspects. The symbolic aspects involve declarative chunks and
procedural production rules. The declarative chunks are the
knowledge-representation units that reside in declarative memory, and the
production rules are responsible for the control of cognition. Access to
these symbolic structures is determined by a subsymbolic level of
neural-like activation quantities."
— Anderson, J. R., & C. Lebiere (2003).
The Newell Test for a theory of cognition.
Behavioral and Brain Sciences 26:587-640.
ACT-R as a theory of cognition
"ACT-R consists of two key memories — a declarative memory and a
procedural memory.
The system has separate memories for each different
type of chunk — for example, addition facts are represented by one type
memory, whereas integers are represented by a separate type memory."
— Anderson, J. R., & C. Lebiere (2003).
The Newell Test for a theory of cognition.
Behavioral and Brain Sciences 26:587-640.
ACT-R: a "simple theory of complex cognition"
"Activation [of an item] in ACT-R theory reflects its log posterior odds of being
appropriate in the current context. This is calculated as a sum of the log
odds that the item has been useful in the past (log prior odds) plus an
estimate that it will be useful given the current context (log likelihood
ratio). [This is a Bayesian approach, more about which later.]
Thus, the ACT-R claim is that the mind keeps track of general
usefulness and combines this with contextual appropriateness to make some
inference about what knowledge to make available in the current
context. The basic equation is
Activation-Level = Base-level + Contextual-Priming
— Anderson, J. R. (1996). ACT: A simple theory of complex
cognition.
American Psychologist, 51, 355-365.
ACT-R as a theory of cognition
"In addition to the learning mechanisms that update activation and
expected outcome, ACT-R can also learn new chunks and production rules."
"New chunks are learned automatically: Each time a goal is completed or a
new percept is encountered, it is added to declarative memory. New
production rules are learned by combining existing production rules. The
circumstance for learning a new production rule is that two rules fire one
after another, with the first rule retrieving a chunk from memory. A new
production rule is formed that combines the two into a macro-rule but
eliminates the retrieval. Therefore, everything in an ACT-R
model (chunks, productions, activations, and utilities) is learnable."
— Anderson, J. R., & C. Lebiere (2003).
The Newell Test for a theory of cognition.
Behavioral and Brain Sciences 26:587-640.
ACT-R as a theory of cognition
"It might seem strange that neural computation should just so happen to
satisfy the well-formedness constraints required to correspond to the
symbolic level of a system like ACT-R. This would indeed be miraculous if
the brain started out as an unstructured net that had to organize itself
just in response to experience. However, as illustrated in the tentative
brain correspondences for ACT-R components [next slide] and in the [preceding]
description of ACT-RN, the symbolic structure emerges out of the structure
of the brain."
— Anderson, J. R., & C. Lebiere (2003).
The Newell Test for a theory of cognition.
Behavioral and Brain Sciences 26:587-640.
ACT-R architecture
The organization of information in ACT-R. Information in the buffers
associated with modules is responded to and changed by production
rules.
DLPFC: dorsolateral prefrontal cortex; VLPFC: ventrolateral
prefrontal cortex.
— Anderson, J. R., et al. (2004).
An Integrated Theory of the Mind.
Psychological Review 111:1036-1060
[but consider the octopus]
On the right: octopus central nervous system (Fig. 6B from The cephalopod nervous
system: What evolution has made of the molluscan design,
B. U. Budelmann, 1995).
[but consider the actual interconnection pattern of primate visual cortical areas]
Felleman, D. J., and D. C. Van Essen (1991).
Distributed hierarchical processing in the primate cerebral cortex.
Cerebral Cortex 1:1-47.
self-critique by Anderson: what ACT-R doesn't do
"Sometimes the suspicion is stated that ACT-R is a general computational
system that can be programmed to do anything. To address this issue, we
would like to specify four senses in which the system falls short of
that.
- Limitations: timing; serial bottleneck.
- Mechanisms not yet incorporated: e.g., speech.
- Domains not yet addressed: e.g., perceptual recognition (as opposed
to attention).
- Technical issues: e.g., temporal bounds for utility learning [cf. credit
assignment problem, due to Minsky].
— Anderson, J. R., & C. Lebiere (2003).
The Newell Test for a theory of cognition.
Behavioral and Brain Sciences 26:587-640.
case study #2: single-cell physiology (W. T. Newsome)
Localized brain stimulation for diagnostic purposes, pioneered by Wilder
Penfield, or: how to smell
burnt toast that's not there —
the cue chosen for Newsome's study: visual motion
Motion is a powerful cue to 3D structure and to "abstract"
features of the moving object:
inducing motion perception by microstimulation of cortical area MT
|
cortical area MT, which "specializes" in motion processing
|
microstimulation of area MT: behavioral tasks and neural data
Behavioral tasks:
- fixation
- saccade (gaze shift)
- reaching
The data: parallel recordings of —
- cell activity
- visual stimulus presence/absence
- gaze direction
microstimulation of cortical area MT: stimulus properties
Moving-dots stimuli at varying degrees of coherence:
The behavioral task is to detect and signal the direction of coherent
motion. Its difficulty (and therefore the performance) depends on the
degree of coherence.
degrees of motion coherence: four examples
typical responses of an MT cell to motion stimuli
 |
 |
| the typical response of a single neuron in area MT |
area MT is organized in columns by the neurons'
preferred direction of motion |
the outcome: inducing perception by cortical microstimulation
From William T.
Newsome's lab (Stanford):
-
(A) spatial layout of fixation point and eye movement targets
-
(B) tuning curve illustrating collective responses of neurons in a
single MT area column; data obtained from 100%-coherence trials
-
(C) behavioral data obtained by electrically stimulating the column
whose tuning is shown in (B);
these data are from
0%-coherence trials!
case study #3: fMRI (J. Haxby)
"Schematic diagram illustrating the locations of the fusiform face area
(FFA), which also has been implicated in expert visual recognition, and
the parahippocampal place area (PPA) on the ventral surface of the right
temporal lobe. In most brains, these areas are bilateral."
— J. V. Haxby, M. I. Gobbini, M. L. Furey, A. Ishai, J. L. Schouten, and
P. Pietrini (2001).
Distributed and overlapping representations of faces and
objects in ventral temporal cortex.
Science 293:2425-2430.
stimulus set
"Examples of stimuli. Subjects performed a one-back repetition detection
task in which repetitions of meaningful pictures were different views of
the same face or object."
stimulus set, continued
"Examples of stimuli. Subjects performed a one-back repetition detection
task in which repetitions of meaningful pictures were different views of
the same face or object."
brain response (average across categories)
"The category specificity of patterns of response was analyzed with
pairwise contrasts between within-category and between-category
correlations. These patterns were normalized to a mean of zero
in each voxel across categories by subtracting the mean response across
all categories. Brain images shown here are the normalized patterns of
response in two axial slices in a single subject. Responses in all
object-selective voxels in ventral temporal cortex are shown."
(C) Mean response across all categories relative to a
resting baseline.
classification based on multivoxel response
regarding multivoxel response
"All 66 regions were active in at least one domain; 65 (98.5%)
were active in two or more domains. [...] The 66 regions were active in an
average of 9.09 (SD 2.27) different domains. The 60 regions active in
action tasks were also active in an average 7.38 (SD 0.98) non-action
domains and 5.5 (SD 0.81) cognitive domains. The 64 regions active in
perception tasks were also active in 7.39 (SD 1.87) non-perceptual domains
and 5.34 cognitive domains. The 59 regions active in both perception and
action tasks were also active in an average of 5.53 (SD 0.80) other
domains, and the 7 regions not active in both perception and action tasks
were active in an average of 3.00 (SD 1.41) of the cognitive domains. Only
one region was active in only cognitive tasks, and that region was active
only in memory."
— Anderson, M. L. (2010).
Neural reuse: A fundamental organizational principle of the brain
Behavioral and Brain Sciences
33:245-266.
moral: ??? [the blind men and the elephant]
Four case studies — four approaches to understanding the brain, and their limitations:
-
An ambitious broad-scope "computational" theory (ACT-R): based on a
single top-down assumption.
-
A single-cell electrophysiology study (microstimulation): neat, but
leaves most questions unresolved.
-
An imaging study (fMRI): trying to pinpoint where in the brain
particular operations happen is not only not informative; localization
of function is a fool's hope.
-
Throwing machine learning at problems (Deep Learning etc.): even
if the hack succeeds, it doesn't reveal much about how the brain works.
