![](bg2.png)
object permanence, and temporal associativity to de-
velop invariant representations for variations of the
same pattern. To test and validate our cortical ar-
chitecture, we used a subset of handwritten digit im-
ages obtained from the MNIST database (Lecun and
Cortes, 1998). Our results show that our cortical ar-
chitecture learns to identify each of the unique digits
present in the sample set and it also pools variations
of the same digit together to develop invariant repre-
sentations.
The main contributions of this paper are as fol-
lows:
• We propose a cortical architecture that uses corti-
cal columns as its basic structural and functional
abstraction.
• We present detail modeling of feedforward and
lateral information processing algorithms that
columns used to identify independent features
from the patterns occurring in their receptive
fields.
• We hypothesizeand model how feedback process-
ing and temporal associations can be hierarchi-
cally utilized by the columns to learn invariant
representations for similar patterns.
• We hypothesize and model how the neocortex
might use feedback for better resource manage-
ment.
• Since in our model there is no separate training
and testing phase, it continues to evolve and learn
all the time.
• Due to its unsupervised learning rules, our model
contains an inherent resilience to permanent er-
rors (both in terms of hardware and software).
2 CORTICAL STRUCTURES AND
ORGANIZATION
The human brain can be divided into two main parts:
the old brain and the new brain. The old brain mainly
constitutes those parts of brain that developed early
in evolution. They include pathways from sensory
modalities to the new brain, spinal cord, and other
parts that deal with instinctual behavior. The new
brain, also referred to as the neocortex, is part of the
brain which is unique to mammals and is highly de-
veloped for humans; it accounts for about 77% of the
human brain (in volume) (Swanson, 1995). The neo-
cortex is responsible for perception, language, imag-
ination, mathematics, arts, music, planning, and all
the other aspects necessary for an intelligent system.
It contains virtually all our memories, knowledge,
skills, and experiences.
A very intriguing property of the neocortex
is its apparent structural and functional unifor-
mity (Mountcastle, 1978; Mountcastle, 1997). Be-
cause of this property, the regions of the neocor-
tex that process auditory inputs, for instance, ap-
pear very similar to the regions that handle visual
and other inputs. This uniformity suggests that even
though different regions specialize in different tasks,
they employ the same underlying algorithm. In
essence, the neocortex is a hierarchy of millions of
seemingly-identical functional units that are called
cortical columns. The concept of cortical columns
was introduced by Mountcastle in his seminal paper
in 1978 (Mountcastle, 1978). Since then, this concept
has been widely accepted and studied. Later studies
showed that cortical columns could further be classi-
fied into minicolumns and hypercolumns (Hubel and
Wiesel, 1962; Calvin, 1998; Johansson and Lansner,
2004; Ringach, 2004; Hirsch and Martinez, 2006). A
hypercolumn contains about 50 to 100 minicolumns,
and each of these minicolumns consists of around 200
to 300 neurons. The term cortical column is some-
times used for both types of columns, though, in lit-
erature, it usually refers to hypercolumns. The mini-
columns within the same hypercolumnshare the same
receptive field and are strongly connected with each
other through inhibitory lateral connections. Studies
(Hubel and Wiesel, 1962; Hubel and Wiesel, 1968)
hypothesize that the minicolumns use these paths to
learn unique/independent features from set of inputs
they are exposed to. The hypercolumns are then ar-
ranged in the form of a hierarchy throughout the neo-
cortex. Information flows up this hierarchy via excita-
tory feedforward paths and flows down the hierarchy
through feedback paths. Figure 1 shows the typical
structure of a hypercolumn.
The arrangement and functionality of the hyper-
columns and minicolumns has been studied in detail
in the visual cortex – the part of the neocortex respon-
sible for processing vision (Hubel and Wiesel, 1962;
Hubel and Wiesel, 1968; Binzegger et al., 2004; Sil-
lito et al., 2006; Peissig and Tarr, 2007). These stud-
ies suggest that minicolumns at the lower levels in
the hierarchy learn to identify very basic features like
edges of different orientation and communicate their
response to minicolumns at the upper levels. It is be-
lieved that cortical regions operate by progressively
abstracting and manipulating increasingly complex
notions throughout the neural hierarchy (Peissig and
Tarr, 2007). For instance, from the set of pixels of an
image, the visual cortex will first identify segments,
then elementary shapes such as angles and intersec-
tions, and increasingly complex combinations, such
DISCOVERING CORTICAL ALGORITHMS
197