Real-Time Low-Power AI at the Edge: 2.5 Day Training Event for the Development of Mission-Critical Applications

RTML Neuromem Chip will be delivered at the training on a RasberryPi enabled board.

Training Program

Day 1 - Welcome

Real-Time Machine Learning closer to life:

• Mission statement: The need for pervasive generic vision

• Conceived the hardware architecture of NeuroMem

• Expert in using NeuroMem for real-time embedded artificial vision

• Merging different sensors (sound, seismic, chemical, etc) with vision sensors

• Artificial perception is “a must” for useful Artificial Intelligence.

NeuroMem is alive and ready to use!

• Fielded applications

• Demonstrations

• Embedded Vision

• Embedded Signal

• Instant access to parallel memory

Concept introduction - An alternative way

• Pattern recognition is supporting all biological activities

• Reproduction

• Survival

• Behavior

• Education

• Mobility

• Motricity

• Inspired by the neo-cortex

• Intelligence is about “memory”

• Memory access should be parallel

• Unrestricted Trainable Fuzzy Look Up Memory as a pattern recognition future?

• Matching output with inputs

Historical review

• Neural network an old concept?

• Early days

• Resurgence in 1988

• Back in the dark age

• IBM ZISC a venture with IBM and G. Paillet in 1993

• Derived from Batchelor and Leon N. Cooper models

• Compound classifier (Bruce Batchelor 1969)

• RCE (restricted Coulomb Energy (Cooper, Elbaum ,Reilly 1972)

• Innovative architecture inspired by massive parallel architecture (CERN – Geneva -1984)

• Semiconductor technology finally allows it (2005- )

Limitation of current technologies including deep learning

• Does not scale easily

• No incremental training capability

• Speed intrinsically decreases with the complexity of problems

• Needs thousands of examples if not more for a simple task.

• Multiples models lead to decreasing speed

• High power

• Medium to a large footprint

Bioinspired pattern recognition hardware complementary to Von Neumann model

• No software

• Non-linear classification

• Unlimited scalability

• Constant response time

• Miniature

• Ultra-low-power

• Unspecialized user trainable

Understanding the need for non-linear mapping

• What is a pattern

• The pattern as a raw data or derived signature

• Mapping a highly dimensional parameters space

• Comparison of useful metrics

• Features (parameters) space mapping

• Peculiarities in features spaces

• How to resolve them

• Phenomenon affecting pattern stability

• In vision

• In signal

• In dataset

NeuroMem is an automatic model generator

• Learning mechanism

• Dealing with uncertainty

Building a training and validation dataset

• Must be representative of the population to recognize

• Should be as large as practicable in order to increase detection probabilities (speed is not an issue for NeuroMem hardware)

• If no real data available computed models will work

• Annotating the data set

• What is annotation

• Some annotation needed for training

• More annotated data even needed for thorough validation

• Cross annotation if possible

• Using NeuroMem for detecting annotation inconsistencies

Day 2 - Beyond the silicon chip, the Intelligence…

Intelligence, can we understand it?

• Sensory stimulation

• Experience for building knowledge

• Developing inhibition (reward/penalty system)

What is an expert?

• Start with textbook knowledge (supervised training)

• Find unusual case, develop a unique expertise

• The “credit assignment” (reward/penalty)

• Expertise cloning

Rules-based systems

• Rules versus “intuition”

• Not everything can be formalized (limit of understanding)

• When a rule is accurate, use it!

• Traffic light analogy

Expert domains

• Designing experts with domain in mind

• Training of a single expert by example

• Type of expertise

• Shape

• Colors

• Others

• Multi-sensors experts

Unary decision-making process

• Throughput versus accuracy

• Cost of error

• Increase throughput or/and accuracy

• Unary uncertainty management

• Adjusting conservatism versus error cost function

Collective decision-making process

• Organizing multi-expert panel

• Rating individual expert knowledge

• Defining expert hierarchy versus skills

• Panel management

• Parallel experts

• Sequential expert

• Dealing with uncertainty

• Reaching unanimity or consensus

Day 3 - Workshop

General Vision Tools

Image Knowledge Builder

NeuroMem Learning System

General Vision supplied applications

• Vision

• Live

• In storage

• Signals

• Audio

• Bio

Users provided applications with GV staff assistance