>
CogniMem www.recognetics.com
1
CogniMem neuron bank
Networks
Signal
Sensor
Communication
CogniMem
Image recognition
Unstructured data mining
Signal recognition
Actuators
Image
Sensor
CogniMem
Communication
Bio Informatics
Internet content filtering
IP routing and addressing
Smart cache for hard disk
Industrial automation
Video surveillance
Medical imaging
Intelligent Transportation Systems
Motion analysis and target tracking
Machine condition monitoring
(Vibration, noise, acceleration,
etc)
Human condition monitoring
(Heart condition, Blood flow, etc)
Voice recognition
CogniMem
CogniMem is a high-performance network of trainable neurons arranged in parallel and
capable of classifying one input vector in 10 microseconds regardless of the number of
neurons in use (i.e. equivalent to 100,000 recognitions per second). Its unique parallel
architecture achieves two technological breakthroughs: a recognition time independent of
the complexity of the knowledge and the ability to cascade multiple chips to size the
network at will.
CogniMem is the practical solution to pattern recognition applications in domains ranging
from high-speed miniature sensors to large data mining server systems. www.recognetics.com
2
A parallel neural network
Vector is broadcasted to all neurons in parallel
Best match of all neurons, 36 clock cycles later
Neuron
#1
Neuron
#2
Neuron
#n
Influence
Field**
Distance*
Category
Context
Prototype
Internal neuron architecture
*calculated during
each vector broadcast
**updated during learning
The true significance of a neuron is
its arrangement into a parallel
network which can decode the
collective response of all the
neurons in a constant amount of
time. This response can be a list of
categories and distances
automatically sorted per
decreasing confidence level. An
empty list means that the vector is
not recognized.
A neuron is a reactive memory
which can autonomously evaluate
the distance between an incoming
vector and a reference vector
stored in its memory. If this
distance falls within its current
influence field, it returns a
positive classification. www.recognetics.com
3
Powerful adaptive learning and modeling
0,0
Accurac
y
Initial
Knowledge
Further
Training
m
o
de
rat
i
on
ct
l
liberal
conservative
Evolutive knowledge curve
- Learn by examples (supervised or unsupervised)
- Automatic model generator
- Map decision spaces by aggregate instead of hyper planes
- Cope with non-linear, convex, disjoints and embedded spaces
- Modulation of throughput versus accuracy
- Multiple space generation using different context for data
fusion and hypothesis generation
- Novelty detection
- Save and restore the contents of the neurons.
A
A
B
C
A or B or C
Transfer human expertise by teaching examples of vector data. The neural network builds
the corresponding knowledge on its own. Add more training at any time to expand or
complete a knowledge base. The neural network will adapt to fit any example adding
novelty to an existing knowledge. The throughput and accuracy of the knowledge can be
tuned to produce a recognition engine with conservative or moderate behavior. www.recognetics.com
4
Unique recognition capabilities
- Constant recognition time after vector broadcast to the
neurons, independent from the number of neurons in use
- Global response status: Positively identified, Identified
with uncertainty or Unknown
- Detailed response of the firing neurons: Distance value
between input and prototype, Category value of the
prototype. This data is retrieved per firing neuron per
increasing distance value (i.e. decreasing confidence level)
- Recognition under multiple independent contexts for data
fusion and hypothesis generation
- Anomaly detection and predictive maintenance through
the detection of an Unknown classification status
Recognition
Time
Number of prototypes
CogniMem
CPU or DSP
A
A
B
C
Unknown
Identified (A)
Uncertain (C,B,A) www.recognetics.com
5
High-speed recognition engine
The following diagram illustrates the implementation of a real-time recognition engine,
where the sensor data is broadcasted directly, or after a signature extraction, to the
neurons. Upon receipt of the last component of the vector, the status of the classification
is readily available as identified, uncertain or unknown. The next inquiry can retrieve the
category of the firing neuron with the closest match. Following inquiries can read the
categories of the other firing neurons to consolidate a weighted global response.
The usage of the high-speed recognition engine requires that a knowledge be previously
loaded into the neurons.
Neural Network
(chain of neurons working in parallel)
Video signature
extraction
Recognition
Of best match
Controller
Sensor input bus
(signal,
video
, else)
Communication www.recognetics.com
6
Single and multi-chip configurations
N CogniMem chips connected
together make a network of N
times more neurons. A knowledge
built on smaller network size can
be loaded as is and expanded by
learning new examples.
Parallel bus control
CogniMem
CogniMem
Sensor
input
I2C bus control
Parallel bus control
CogniMem
CogniMem
Sensor
input
Parallel bus control
CogniMem
CogniMem
CogniMem
A sizable neural network
Multi-chip configuration for High speed image or signal recognition
The first chip of the chain receives the sensor input and run the recognition engine.
The neural network is distributed between the 1
st
chip and the additional chips
I2C serial bus interface
parallel bus interface www.recognetics.com
7
Status of the Technology as of Apr-07
CogniMem on FPGA
1 cluster of 16 neurons
Available now
CogniMem ASIC_1024
64 clusters of 16 neurons
Prototypes expected Aug 2007
Production batch Q4, 2007
Common specifications
Neurons=1024
Prototype or vector=256 bytes
Contexts= 127
Distance norm= L1 or Lsup
Category= 32768
Choice of
- K Nearest Neighbor (KNN) model
- Restricted Coulomb Energy (RCE) Compound
Classifier model
Classification status after 1 clock cycle
Category and distance inquiry = 36 clock cycles
Distance inquiry = 16 clock cycles
Video signature extraction: average intensity of
up to 256 user-defined blocks of pixels
CogniMem simulation (Available)
- 2000 neurons
- Non parallel architecture and therefore
recognition time based on the CPU clock and
knowledge size www.recognetics.com
8
CM_1024 chip, Electrical and Mechanical Specs
Communication
Parallel control bus (23 pins)
I2C slave control bus (2 pins)
Mechanical/Electrical
30-pins connectivity
Up to 27 Mhz clock in single chip
Up to 15 Mhz for multiple chips
Up to 27 Mhz sensor clock
1.2v for the core; 3.3v for the I/O
TQFP 100 package, 16 x16 mm
CogniMem pin out
> RECO_EN
> S_CHIP
> V_EN
> V_DATA
> V_CLK
> V_FV
> V_LV
> I2C_EN
> I2C_SCK
> I2C_SDA
B_BSY >
CAT_VALID >
> G_CLK
G_RESET
> DCI
DS ><
R_W_ ><
REG ><
DATA ><
ID_ ><
UNC_ ><
RDY ><
DCO >
Direct vector
data input
I2C port
parallel port for
external control
and inter-chip
communication