IEEE Circuits and Systems Magazine - Q2 2018 - 79

non-linear dynamic operators and logic functions as
instructions [2].
To convert the CNN to a fully programmable microprocessor, some new elements were introduced both in
the cellular array (Fig. 3 left) and in the circuits around
it. Each cell was extended with analog and binary (logic)
memories, and pixel-by-pixel logic operation unit (Fig 3
right). Using these memories, entire m # n sized grayscale or binary images can be stored in the array. These
memories are similar to the registers of a conventional
microprocessor. The CNN-UM can execute image processing primitives on these images as the normal microprocessors execute additions or multiplications on
scalars stored in its internal registers.
The instruction set of the CNN-UM includes most of
the basic pixel manipulation operations of early image
processing. Each of them is represented by a template.
The templates are collected in the Template Library [6]
which is described in Section VII.
The extended cells operate in single instruction multiple data (SIMD) mode, therefore a control and programming unit (Fig. 3) is needed, which distributes
the instruction sequence for all the cells parallel. This
stores the templates (analog operations), the look-up
tables (binary, logical operations), and the description
of the instruction sequence. Among the instructions,
there are some branch and loop control ones, to be able
to write sophisticated program codes. A dedicated program language-called Alpha [5]-was developed for
supporting high level coding.
C. Universality of the CNN-UM in Turing Sense
Besides the stability, the Turing universality was also
studied. Universality was proved first with the implementation of the game of life [7]. Later on, it was proved
that the CNN-UM can implement any local Boolean function, which is equivalent with Turing universality as well
[8]. Finally, it was also proved that as a nonlinear dynamic operator, it can realize any local operator of fading memory (practically all reasonable operators) [5].
III. Biological Motivation
Mimicking and modeling biological structures is a good
practice in all engineering areas. Computer scientists
study the different sensory, cognitive, and understanding processes of the brain. Since CNN architecture has
an inherent similarity to the architectures of the different layers of the visual pathway, it is not surprising that
numerous neuromorph models were implemented using
different CNN architectures. Among these, we can find
5 layers retina model with delay type templates, LGN
models, and cortex models. Some of the models gives
explanation to visual illusions as well [5]. The visual sysSECOND quartEr 2018

LCCU:
Local Communication
and Control Unit

LAM:
Local
Analog
Memory

CNN
Nucleus
with Switches

LAOU:
Local Analog
Output Unit

LLM:
Local
Logic
Memory

LLU:
Local Logic
Unit

(a)
GAPU
APR: Analog Programming Instruction Register
LPR: Logic Program Instruction Register
SCR: Switch Configuration Register
GACU: Global Analogic Control Unit
(b)
Figure 3. Extended cell of the CNN-um (a), and the element
of global analogic Programming unit (gaPu) (b).

tem is the most important sensory organ for humans
as well as mammals. Its first and best known part is
the retina which is a preprocessor and it sends visual
information to the brain via several parallel channels.
The framework of mammalian retina modeling via multilayer CNN was published in [9]. The structure of the
retina modell can be seen on Fig. 4. Determination of the
model parameters requires very high computing power
and accurate solution. On the current multi-layer analog
VLSI chips only the basic building blocks of the retina
model can be implemented. Additionally, the size of the
cell array is relatively small. Emulated digital CNN-UM
architectures can be used to overcome the limitations
of the analog VLSI chips such as small number of layers,
small array size and limited accuracy. A real time operating retina model can be implemented on a six-million
gate equivalent FPGA.
A. Tactile Sensing and Preprocessing
The strength of the CNN processing infrastructure was
also shown in the area of tactile sensing. An efficient
and fast method was elaborated to detect and identify
the slipping and twisting motion of the touching objects
[10]. This kind of actions cannot be detected with sensors sensing only the normal (perpendicular) component of the forces acting between the surfaces. In this
IEEE CIrCuItS aND SyStEmS magazINE

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q2 2018