Transcript Intelligence and Patterns

Intelligence and Patterns
The Brain is a Pattern matching Machine
[email protected], attributed copies permitted
1
His book is free on line for download http://hubel.med.harvard.edu/index.html
PREFACE download
1 INTRODUCTION download
2 IMPULSES, SYNAPSES, AND CIRCUITS download
3 THE EYE download
4 THE PRIMARY VISUAL CORTEX download
5 THE ARCHITECTURE OF THE VISUAL CORTEX download
6 MAGNIFICATION AND MODULES download
7 THE CORPUS CALLOSUM AND STEREOPSIS download
8 COLOR VISION download
9 DEPRIVATION AND DEVELOPMENT download
10 PRESENT AND FUTURE download
FURTHER READING download
SOURCES OF ILLUSTRATIONS
INDEX
[email protected], attributed copies permitted
2
Reverse-Engineering the Brain
At MIT, neuroscience and artificial intelligence are beginning to intersect.
www.technologyreview.com/read_article.aspx?id=17111
While AI's progress has been slower than expected, neuro-science has
gotten much more sophisticated in its understanding of how the brain
works. Nowhere is this more obvious than in the 37 labs of MIT's BCS
Complex. Groups here are charting the neural pathways of most of the
higher cognitive functions (and their disorders), including learning,
memory, the organization of complex sequential behaviors, the formation
and storage of habits, mental imagery, number management and control,
goal definition and planning, the processing of concepts and beliefs, and
the ability to understand what others are thinking.
One breakthrough example: Biological vision solves problems in several different ways.
One, according to Poggio's group, is to organize parallel processing around two simple
operations and then alternate these operations in an ordered way through layers of neurons.
Layer A might filter the basic inputs from the optic nerve; layer B would integrate the results
from many cells in layer A; C would filter the inputs from B; D would integrate the results
from C; and so on, perhaps a dozen times. As a signal rises through the layers, the outputs
of the parallelized processors gradually combine, identity emerges, and noise falls away.
Some of their assumptions turned out to predict real features, such as the presence of cells
(call them OR cells) that pick the strongest or most consistent signal out of a group of
inputs and copy it to their own output fibers. (Imagine a group of three neurons, A, B, and C,
all sending signals to OR neuron X. If those signals were at strength levels 1, 2, and 3
respectively, X would suppress A and B and copy C's signal to its output. If the strengths
had been 3, 2, and 1, it would have copied A's signal and suppressed those of B and C.)
When human subjects and Serre and Poggio's immediate-recognition program took the
animal presence/absence test, the computer did as well as the humans -- and better than the
best machine vision programs available. (Indeed, it got the right answer 82 percent of the
time, while the humans averaged just 80 percent.)
[email protected], attributed copies permitted
3
Processing without attention and
consciousness
Rapid visual categorization
Reverse
Engineering
the Brain
Visual input can be classified very
rapidly. As famously demonstrated by
Thorpe and colleagues (Thorpe et al.,
1996; Kirchner and Thorpe, 2006)
around 120 msec following image
onset, some brain processes begin to
respond differentially to images
containing one of more animals from
pictures than contain none. At this
speed, it is no surprise that subjects
often respond without having
consciously seen the image;
consciousness for the image may
come later or not at all.
Dual-task and dual-presentation
paradigms support the idea that such
discriminations can occur in the nearabsence of focal, spatial attention
implying that purely feed-forward
networks can support complex visual
decision-making in the absence of
both attention and consciousness.
Indeed, this has now been formally
shown in the context of a purely feedforward computational model of the
primate’s ventral visual system (Serre
et al., 2007).
www.technologyreview.com/printer_friendly_article.aspx?id=17111
www.scholarpedia.org/article/Attention_and_consciousness/
processing_without_attention_and_consciousness
[email protected], attributed copies permitted
4
One Example of Chunking
Explaining Rapid Categorization.
Thomas Serre, Aude Oliva, Tomaso Poggio.
http://cbcl.mit.edu/seminars-workshops/workshops/serre-slides.pdf
[email protected], attributed copies permitted
5
Chunking Hierarchy
The results by
Logothetis et al. are in
agreement with a
general computational
theory [Poggio, 2000]
suggesting that a
variety of visual object
recognition tasks
(involving the
categorization of
objects and faces at
different levels) can
be performed based on
a linear combination of
a few units tuned to
specific task-related
training examples.
Thomas Serre (2006), Learning a Dictionary of Shape-Components in Visual Cortex: Comparison with
Neurons, Humans and Machines, Ph.D. dissertation, Department of Brain and Cognitive Sciences,
Massachusetts Institute of Technology, April, http://cbcl.mit.edu/publications/ps/MIT-CSAIL-TR-2006-028.pdf
The organization of visual cortex based on a core of knowledge that has been accumulated
over the past 30 years. The figure is modified from[Oramand Perrett, 1994]mostly to include
the likely involvement of prefrontal cortex during recognition tasks by setting task-specific
circuits to read-out shape information from IT [Scalaidhe et al., 1999; Freedman et al., 2002,
2003; Hung et al., 2005].
[email protected], attributed copies permitted
6
[email protected], attributed copies permitted
7
[email protected], attributed copies permitted
8
[email protected], attributed copies permitted
9
[email protected], attributed copies permitted
10
[email protected], attributed copies permitted
11
96
[email protected], attributed copies permitted
12
[email protected], attributed copies permitted
13
[email protected], attributed copies permitted
14
[email protected], attributed copies permitted
15
[email protected], attributed copies permitted
16
Cortical column
Retrieved from "http://en.wikipedia.org/wiki/Cortical_column"
A cortical column, also called hypercolumn or sometimes cortical module,[1] is a group of neurons in the brain cortex which can be
successively penetrated by a probe inserted perpendicularly to the cortical surface, and which have nearly identical receptive fields.
Neurons within a minicolumn encode similar features, whereas a hypercolumn "denotes a unit containing a full set of values for any
given set of receptive field parameters"[2]. A cortical module is defined as either synonymous with a hypercolumn (Mountcastle) or as a
tissue block of multiple overlapping hypercolumns (Hubel&Wiesel).
Human cerebral cortex
The human cerebral cortex is composed of 6 somewhat distinct layers; each layer identified by the nerve cell type and the destination
of these nerve cell's axons (within the brain). The human cortex is a roughly 2.4 mm thick sheet of neuronal cell bodies that forms the
external surface of the telencephalon. The dolphin cortical column is composed of only 5 layers. The reptilian cortex has only three
layers.
The columnar functional organization, as originally framed by Vernon Mountcastle, suggests that neurons that are horizontally more
than 0.5 mm (500 µm) from each other do not have overlapping sensory receptive fields, and other experiments give similar results:
200–800 µm (Buxhoeveden 2002, Hubel 1977, Leise 1990, etc.). Various estimates suggest there are 50 to 100 cortical minicolumns in a
hypercolumn, each comprising around 80 neurons.
An important distinction is that the columnar organization is functional by definition, and reflects the local connectivity of the cerebral
cortex. Connections "up" and "down" within the thickness of the cortex are much denser than connections that spread from side to
side.
Hubel and Wiesel studies
Hubel and Wiesel followed up on Mountcastle's discoveries in the somatic sensory cortex with their own studies in vision. A part of the
discoveries that resulted in them winning the 1981 Nobel Prize[3] was that there were cortical columns in vision as well, and that the
neighboring columns were also related in function in terms of the orientation of lines that evoked the maximal discharge. Hubel and
Wiesel followed up on their own studies with work demonstrating the impact of environmental changes on cortical organization, and
the sum total of these works resulted in their Nobel Prize.
Size of cortex
From the size of the cortex and the typical size of a column, it can be estimated that there are about two million function columns in
humans [4]. There may be more if the columns can overlap, as suggested by Tsunoda et al [5].
References
1. Kolb, Bryan; Whishaw, Ian Q. (2003). Fundamentals of human neuropsychology. New York: Worth. ISBN 0-7167-5300-6.
2. Horton JC, Adams DL (2005). "The cortical column: a structure without a function". Philos. Trans. R. Soc. Lond., B, Biol. Sci. 360
(1456): 837–62. doi:10.1098/rstb.2005.1623. PMID 15937015.
3. "The Nobel Prize in Medicine 1981". http://nobelprize.org/medicine/laureates/1981/. Retrieved on 2008-04-13.
4. Christopher Johansson and Anders Lansner (January 2007). "Towards cortex sized artificial neural systems". Neural Netw 20 (1):
48–61. doi:10.1016/j.neunet.2006.05.029. PMID 16860539.
5. Kazushige Tsunoda, Yukako Yamane, Makoto Nishizaki, and Manabu Tanifuji (August 2001). "Complex objects are represented in
macaque inferotemporal cortex by the combination of feature columns". Nat. Neurosci. 4 (8): 832–8. doi:10.1038/90547. PMID
11477430.
[email protected], attributed copies permitted
17
Cortical minicolumn
Retrieved from "http://en.wikipedia.org/wiki/Cortical_minicolumn"
A cortical minicolumn is a vertical column through the cortical layers of the brain, comprising perhaps 80–120 neurons,
except in the primate primary visual cortex (V1), where there are typically more than twice the number. There are about
2×108 minicolumns in humans.[1] From calculations, the diameter of a minicolumn is about 28–40 µm.
Many sources support the existence of minicolumns, especially Mountcastle,[2] with strong evidence reviewed by
Buxhoeveden and Casanova[3] who conclude "... the minicolumn must be considered a strong model for cortical
organization" and "[the minicolumn is] the most basic and consistent template by which the neocortex organizes its
neurones, pathways, and intrinsic circuits". See also Calvin's Handbook on cortical columns.
Size – The minicolumn measures of the order of 40–50 µm in transverse diameter (Mountcastle 1997, Buxhoeveden 2000,
2001); 35–60 µm (Schlaug, 1995, Buxhoeveden 1996, 2000, 2001); 50 µm with 80 µm spacing (Buldyrev, 2000), or 30 µm with
50 µm (Buxhoeveden, 2000). Larger sizes may not be of human minicolumns, for example Macaque monkey V1
minicolumns are 31µm diameter, with 142 pyramidal cells (Peters, 1994) — 1270 columns per mm2. Similarly, the cat V1 has
much bigger minicolumns, ~56 µm (Peters 1991, 1993) .
The size can also be calculated from area considerations: if cortex (both hemispheres) is 1.27×1011 µm2 then if there are
2×108 minicolumns in the cortex then each is 635 µm2, giving a diameter of 28 µm (if the cortex area were doubled to the
commonly quoted value, this would rise to 40 µm). Johansson and Lansner[4] do a similar calculation and arrive at 36 µm
(p51, last para).
Facts
• Cells in 50µm minicolumn all have the same receptive field; adjacent minicolumns may have very different fields (Jones,
2000).
• Downwards projecting axons in minicolumns are ≈10µm in diameter, periodicity and density similar to those within the
cortex, but not necessarily coincident (DePhilipe, 1990).
• Thalamic input (1 axon) reaches 100–300 minicolumns.
• The number of fibres in the corpus callosum is 2–5×108 (Cook 1984, Houzel 1999) — perhaps related to the number of
minicolumns.
References
1.Towards cortex sized artificial neural systems, Christopher Johansson and Anders Lansner, Neural Networks, Vol. 20 #1,
pp48–61, Elsevier, January 2007
2.The columnar organization of the neocortex, Vernon B. Mountcastle, Brain, Vol. 20 #4, pp701–722, Oxford University
Press, April 1997
3.The minicolumn hypothesis in neuroscience, Daniel P. Buxhoeveden and Manuel F. Casanova, Brain, Vol. 125 #5, pp935–
951, Oxford University Press, May 2002.
4.Towards cortex sized artificial neural systems, Christopher Johansson and Anders Lansner, Neural Networks, Vol. 20 #1,
pp48–61, Elsevier, January 2007
[email protected], attributed copies permitted
18
Involuntary maybe, but certainly not random
12Feb09 – www.physorg.com/news153670434.html
Our eyes are in constant motion. Even when we attempt to stare straight at a stationary target, our eyes jump and jiggle imperceptibly.
Although these unconscious flicks, also known as microsaccades, had long been considered mere "motor noise," researchers at the
Salk Institute for Biological Studies found that they are instead actively controlled by the same brain region that instructs our eyes to
scan the lines in a newspaper or follow a moving object.
Their findings, published in the Feb. 13, 2009 issue of Science, provide new insights into the importance of these movements in
generating normal vision.
"For several decades, scientists have debated the function, if any, of these fixational eye movements," says Richard Krauzlis, Ph.D., an
associate professor in the Salk Institute's Systems Neurobiology Laboratory, who led the current study. "Our results show that the
neural circuit for generating microsaccades is essentially the same as that for voluntary eye movements. This implies that they are
caused by the minute fluctuations in how the brain represents where you want to look."
"There was a lot of past effort to figure out what fixational eye movements contribute to our vision," adds lead author Ziad Hafed,
Ph.D., Sloan-Swartz Fellow in the Systems Neurobiology Laboratory, "but nobody had looked at the neural mechanism that generates
these movements. Without such knowledge, one could only go so far in evaluating microsaccades' significance and why they actually
exist."
Wondering whether the command center responsible for generating fixational eye movements resides within the same brain structure
that is in charge of initiating and directing large voluntary eye movements, Hafed decided to measure neural activity in the superior
colliculus before and during microsaccades.
He not only discovered that the superior colliculus is an integral part of the neural mechanism that controls microsaccades, but he
also found that individual neurons in the superior colliculus are highly specific about which particular microsaccade directions and
amplitudes they command—whether they be, say, rightward or downward or even oblique movements. "Data from the population of
neurons we analyzed shows that the superior colliculus contains a remarkably precise representation of amplitude and direction down
to the tiniest of eye movements," says Krauzlis.
The Salk researchers, in collaboration with Laurent Goffart, Ph.D., a professor at the Institut de Neurosciences Cognitives de la
Méditerranée in Marseille, France, also temporarily inactivated a subset of superior colliculus neurons and analyzed the resulting
changes in microsaccades. They discovered that a fully functional superior colliculus is required to generate normal microsaccades.
"Because images on the retina fade from view if they are perfectly stabilized, the active generation of fixational eye movements by the
central nervous system allows these movements to constantly shift the scene ever so slightly, thus refreshing the images on our
retina and preventing us from going 'blind,'" explains Hafed. "When images begin to fade, the uncertainty about where to look
increases the fluctuations in superior colliculus activity, triggering a microsaccade," adds Krauzlis.
Microsaccades may, however, do more than prevent the world around us from fading when we stare at it for too long. Even when our
gaze is fixed, our attention can shift to an object at the periphery that attracts our interest. In an earlier study, Hafed discovered that
although we may avert our eyes from an attractive man or woman, microsaccades will reveal such objects of attraction because their
direction is biased toward objects to which we are unconsciously attracted.
By showing in the current study that the superior colliculus is involved in generating microsaccades, Hafed and his colleagues could
now explain why this happens. "The superior colliculus is a major determinant of what is behaviorally relevant in our visual
environment, so paying attention to one location or the other alters superior colliculus activity and therefore alters these eye
movements as well," says Hafed.
[email protected], attributed copies permitted
19
The Blue Brain project is the first comprehensive
attempt to reverse-engineer the mammalian brain, in
order to understand brain function and dysfunction
through detailed simulations.
In July 2005, EPFL and IBM announced an exciting new
research initiative - a project to create a biologically accurate, functional model of
the brain using IBM's Blue Gene supercomputer. Analogous in scope to the
Genome Project, the Blue Brain will provide a huge leap in our understanding of
brain function and dysfunction and help us explore solutions to intractable
problems in mental health and neurological disease.
At the end of 2006, the Blue Brain project had created a model of the basic
functional unit of the brain, the neocortical column. At the push of a button, the
model could reconstruct biologically accurate neurons based on detailed
experimental data, and automatically connect them in a biological manner, a task
that involves positioning around 30 million synapses in precise 3D locations.
In November, 2007, the Blue Brain project reached an important milestone and the
conclusion of its first Phase, with the announcement of an entirely new datadriven process for creating, validating, and researching the neocortical column.
[email protected], attributed copies permitted
20
About the Blue Brain Project
http://bluebrain.epfl.ch/page18699.html
The cerebral cortex, the convoluted "grey
matter" that makes up 80% of the human
brain, is responsible for our ability to
remember, think, reflect, empathize, communicate, adapt to
new situations and plan for the future. The cortex first
appeared in mammals, and it has a fundamentally simple
repetitive structure that is the same across all mammalian
species.
The brain is populated with billions of neurons, each connected to thousands of
its neighbors by dendrites and axons, a kind of biological "wiring". The brain
processes information by sending electrical signals from neuron to neuron along
these wires. In the cortex, neurons are organized into basic functional units,
cylindrical volumes 0.5 mm wide by 2 mm high, each containing about 10,000
neurons that are connected in an intricate but consistent way. These units operate
much like microcircuits in a computer. This microcircuit, known as the neocortical
column (NCC), is repeated millions of times across the cortex. The difference
between the brain of a mouse and the brain of a human is basically just volume humans have many more neocortical columns and thus neurons than mice.
This structure lends itself to a systematic modeling approach. And indeed, the
first step of the Blue Brain project is to re-create this fundamental microcircuit,
down to the level of biologically accurate individual neurons. The microcircuit can
then be used in simulations.
For an in-depth view of the project, read Henry Markram's Perspectives article in
the February 2006 issue of Nature Reviews Neuroscience.
[email protected], attributed copies permitted
21
Building the microcircuit
http://bluebrain.epfl.ch/page19092.html
Modeling Neurons – Neurons are not all alike - they come in a
variety of complex shapes. The precise shape and structure of a
neuron influences its electrical properties and connectivity with
other neurons. A neuron's electrical properties are determined to a large extent by a
variety of ion channels distributed in varying densities throughout the cell's
membrane. Scientists have been collecting data on neuron morphology and electrical
behavior of the juvenile rat in the laboratory for many years, and this data is used as
the basis for a model that is run on the Blue Gene to recreate each of the 10,000
neurons in the NCC.
Modeling connections – To model the neocortical column, it is essential to
understand the composition, density and distribution of the numerous cortical cell types. Each class of
cells is present in specific layers of the column. The precise density of each cell type and the volume of
the space it occupies provides essential information for cell positioning and constructing the foundation
of the cortical circuit. Each neuron is connected to thousands of its neighbors at points where their
dendrites or axons touch, known as synapses. In a column with 10,000 neurons, this translates into
trillions of possible connections. The Blue Gene is used in this extremely computationally intensive
calculation to fix the synapse locations, "jiggling" individual neurons in 3D space to find the optimal
connection scenario.
Modeling the column – The result of all these calculations is a re-creation, at the cellular level, of the
neocortical column, the basic microcircuit of the brain. In this case, it's the cortical column of a juvenile
rat. This is the only biologically accurate replica to date of the NCC - the neurons are biologically realistic
and their connectivity is optimized. This would be impossible without the huge computational capacity of
the Blue Gene. A model of the NCC was completed at the end of 2006.
In November, 2007, The Blue Brain Project officially announced the conclusion of Phase I of the project,
with three specific achievements:
1.A new modeling framework for automatic, on-demand construction of neural circuits built from
biological data
2.A new simulation and calibration process that automatically and systematically analyzes the biological
accuracy and consistency of each revision of the model
3.The first cellular-level neocortical column model built entirely from biological data that can now serve as
a key tool for simulation-based research
[email protected], attributed copies permitted
22
[email protected], attributed copies permitted
23