IEEE Circuits and Systems Magazine - Q2 2018 - 80

Cone τ
τ
Cone2
τ
Horizontal
Bip. Excit. τ
Desensitized
Bip. Excit. τ
Receptor
Amacrine τ
FB
τ
Ganglion

Input

λ
λ

-
λ
λ

τ
τ

Output of the Channel 1

-
λ
λ

τ
τ

τ Bip. Inhib.
Desens.
λ Bip. Inhib.
Rec.
λ Amacrine
λ

Output of the Channel 2

Figure 4. a two-channel retina model.

experimental setup the sensing elements were Si based
tactile sensors with 500 # 500 mm 2 taxel size, and the
signal processing unit was a 64 # 64 CNN-UM analog
VLSI chip. The size of the array is 2 # 2 taxel (tactile pixel). Each taxel contained four piezorezisitive elements.
The sensing elements were mounted on a two-fingered
robot hand. The actuator was controlled in closed loop.
An integrated sensing-processing-actuating system has
been developed. The analogic algorithm detected the
typical events. The motivation was to study the rising
forces between contacting surfaces. In different areas
of science (in robotics, automation etc.) and industry
(motor-car-, aeronautical-, construction industry), it is
essential to be able to read and process these 3D pressure fields.
IV. Simulators and Application Platform
The research teams within the CNN community synchronized their simulation and development efforts on
a way that they used the same kinds of simulation and
development platforms. This simplified the code and
template sharing and ultimately the communication between the laboratories. Multiple simulator families and
application platforms were developed according to
the actual CNN state-of-the-art and the changing available desktop or embedded computational platforms.
The simulators and development platforms, including
the CNN chip based test and application platforms are
discussed in this chapter.
A. Software Simulators and Development Platforms
The first simulators of the CNN were designed to be able
to select input and initial state (images or constant values) for the CNN and a cloning template, and simulate the
dynamics of the network using forward and backward
Euler or Runge-Kutta methods [11]. As it turned out,
forward Euler simulation of the CNN equations with rela80

IEEE CIrCuItS aND SyStEmS magazINE

tive large time constants (even 1 sometimes) led proper results for many templates. This led to the invention
of the discrete time CNN (DTCNN) and its chip implementation [12], because its discrete time state equations
practically the same as the forward Euler with time constant 1. DTCNNs implement the same dynamics and can
also be used in practice as it is described in [13] and [14].
Later on, the simulators were able to execute template
sequences on individual images or image streams by using script files to describe the operation.
The introduction of the CNN-UM made significant
changes, because the simulators were replaced with development platforms which had operation sets, low and
high level languages, compilers, function libraries, different kinds of memories, and IO options. The first low
level language was the Analogic Machine Code (AMC)
[15], which was used not just in simulators but for
programming CNN chips as well. Later on, a high level
language, called Alpha was defined, and a compiler
was developed.
Since these languages were designed for CNN operations, the input, initial state and the output of the CNN,
which are 2D arrays of continuous or binary value scalars, were handled as single variables. One could execute entire feedforward or feedback convolutions or logical operations on these arrays with a single instruction,
rather than writing a double for cycle for manipulating
an array. This was quite new at that time, especially in
an assembler type language. Naturally, the simulator resolved these operators as double for cycles, however,
the CNN chips could really execute them in one analog
operation, which took a few microseconds only, depending on the spatial-temporal transient decay of the analog
processor array.
The CNN related instruction set of the languages are
listed below. The execution time ranges of the instructions
are indicated in parentheses, when they are executed on
SECOND quartEr 2018

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