
One of the promising ways to solve this problem
seems to be the transition to the use of neuromorphic
computing systems based on the principles of the
human brain (Christensen, (2022). Their most
attractive features are the principles of biological
neural networks, such as highly parallel information
processing, processing procedures embedded directly
in data blocks, scalability, event-driven calculations,
etc. It is expected that a new generation of computers
based on these principles (sometimes called third-
generation neural – spiking networks) can be
effectively used both for storing extremely large
volumes of data and for processing it in an acceptable
time and at the same time with much less energy
consumption. In addition to energy efficiency,
neuromorphic systems are ideal for implementing
machine learning approaches and have enormous
potential for computing beyond the von Neumann
paradigm. These advantages will give them priority
in most information technologies.
Considering this, we have recently made attempts
to initiate the research on the development of new
methods for working with data streams based on the
principles of neuromorphic computing (Antsiperov,
2022). The proposed work presents some of the
results of the efforts undertaken. Namely, below we
discuss the possibilities of processing relatively large
volumes of streaming data using neuromorphic
methods in the problem of image encoding /
restoration. The proposed methods are focused on
representing input data in the form of a stream of
discrete events (counts), like the firing events of
retinal neurons. For its adequate statistical
description, a special representation was developed in
the form of sample of counts (sampling
representation). The probabilistic nature of the
representation naturally leads to a generative model
of the streaming data encoder, which can be
formalized as a parametric model of a set (mixture) of
components. We discovered that within the proposed
generative model the search for optimal encoding can
be reformulated as a statistical maximum likelihood
problem. We solved this ML problem under the
assumption that a set of components has a receptive
field (RF) structure that embody universal principles
(including lateral inhibition) of a biological neural
network. Issues of image decoding are considered in
the context of restoring spatial contrasts, which also
partly emulates the work of the so-called simple /
complex cells of the primary visual cortex. It is shown
that the coupled ON-OFF decoding model allows for
the restoration of sharp image details in the form of
edge-directed interpolation.
The main content of the work is grouped in the
following three sections. Section 2 contains a brief
overview of neurophysiological data on the structure
of RFs and methods for it modelling. Section 3 is
related to the substantiation of the statistical
description of the RF functions for processing the
input stream of samples. And in the last section the
results of the numerical procedure for image
restoration (decoding) are discussed based on the
results of encoding the input stream by the RF system.
The conclusion briefly summarizes the results and
outlines avenues for further research.
2 RETINAL RECEPTIVE FIELDS
AS STRUCTURAL UNITS OF
EDGE ENCODING
As mentioned above, the proposed encoding method
deals with equipping the image forming area with
some fixed “neuromorphic” structure. It is believed
that this structure is initially given and does not
depend, among other things, on the radiation intensity
focused by the lens of the eye on the retina, or by the
optics of the video camera on the CMOS-matrix.
Essentially, the structure mentioned is simple enough.
Namely, it models the structure of the receptors
(outer) layer of the human (or higher vertebrates)
retina, known as the receptive field (RF) system.
The general concept of RFs as structural units of
sensory neuronal systems of living organisms has
been known for a long time. As for the periphery of
the visual system, the beginning of systematic
research and analysis of the RF features is usually
associated with the work of Kuffler (Kuffler,1953) in
the early 50s. According to the tradition, that
followed Kuffler, receptive fields are understood as
small areas of the retina containing tens to hundreds
of input receptors (cones/rods), whose stimulation
leads to the activation of certain output neurons
(RGCs - retinal ganglion cells). It is important to note
that along the path of data propagation from receptors
to the RGC, visual information undergoes several
transformations and modifications carried out by
intermediate neurons (horizontal, bipolar and
amacrine cells) of the retina. As a result, in addition
to the spatial structure, the RFs also has a certain
functional arsenal. It is associated with the division of
the RF surface into two parts: a central region that
receives data directly from the retinal receptors,
which is called the RF-centre, and a peripheral region
concentric to the centre, which receives data through
horizontal cells and is called the RF-surround. It is
ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods
486