Neuromorphic Intelligence: Brain-inspired Strategies for AI Computing Systems

Summary

Giacomo Indiveri from the University of Zurich and ETH Zurich discusses brain-inspired strategies for developing low-power artificial intelligence computing systems. He highlights the limitations of current AI algorithms, which consume significant energy and are less versatile than natural intelligence. Indiveri introduces neuromorphic engineering as a promising approach, emphasizing the importance of emulating the brain's structure and function to create efficient, compact, and intelligent devices. He showcases the Dynamic Neuromorphic Asynchronous Processor as an example of this technology and its applications in fields like industrial monitoring and machine vision.

Takeaways

• 🧠 The talk by Giacomo Indiveri from the University of Zurich and ETH Zurich focuses on brain-inspired strategies for low-power artificial intelligence computing systems.
• 📈 The success of AI algorithms and neural networks, which originated in the late 80s, has recently surged due to advancements in hardware technologies, availability of large datasets, and improvements in algorithms.
• 🔋 A significant challenge with current AI algorithms is their high energy consumption, with estimates suggesting the ICT industry could consume about 20% of the world's energy by 2025.
• 💡 The high power usage is largely due to the extensive data and memory resources required, particularly the energy spent moving data between memory and processing units.
• 🌐 The narrow specialization of AI networks is highlighted as a fundamental issue, contrasting with the general-purpose capabilities of natural intelligence found in animal brains.
• 🚀 Neuromorphic engineering, inspired by the structure and function of the brain, is presented as a promising approach to overcome the limitations of current AI systems.
• 🏫 The term 'neuromorphic' has been adopted by various communities, including those designing CMOS circuits to emulate brain functions, those developing practical devices for problem-solving, and those working on emerging memory technologies.
• 🔬 The biological neural networks differ significantly from simulated ones, utilizing time dynamics and the physics of their elements, with memory and computation co-localized within each neuron.
• 💡 The key to low-power computation lies in parallel arrays of processing elements with co-localized memory and computation, avoiding the need for data transfer between separate memory and processing units.
• 🌟 The potential of neuromorphic systems is demonstrated through various applications such as ECG anomaly detection, vibration anomaly detection, and intelligent machine vision, showcasing their potential for practical, energy-efficient solutions.

Q & A

What is the main focus of Giacomo Indiveri's talk?

The main focus of Giacomo Indiveri's talk is on brain-inspired strategies for low-power artificial intelligence computing systems.

Why have artificial intelligence algorithms and networks only recently started outperforming conventional approaches?

Artificial intelligence algorithms and networks have only recently started outperforming conventional approaches due to advancements in hardware technologies providing enough computing power, the availability of large datasets for training, and improvements in algorithms making networks more robust and performant.

What is the estimated energy consumption of the ICT industry by 2025 in relation to the world's total energy?

By 2025, it is estimated that the ICT industry will consume about 20 percent of the entire world's energy.

Why are current AI algorithms considered to be power-hungry?

Current AI algorithms are power-hungry because they require large amounts of data and memory resources, and the energy cost of moving data from memory to computing and back is very high.

How do neuromorphic computing systems differ from traditional artificial neural networks?

Neuromorphic computing systems differ from traditional artificial neural networks by emulating the brain's structure and function more closely, using parallel arrays of processing elements with computation and memory co-localized, and leveraging the physics of the devices for computation.

What is the significance of the term 'neuromorphic' in the context of Giacomo Indiveri's work?

In the context of Giacomo Indiveri's work, 'neuromorphic' refers to the design of systems that mimic the neural structure and computational strategies of the brain, aiming to create compact, intelligent, and energy-efficient devices.

What are the three main strategies Giacomo Indiveri suggests for creating low-power artificial intelligence systems?

The three main strategies suggested for creating low-power artificial intelligence systems are using parallel arrays of processing elements with co-localized computation and memory, leveraging the physics of analog circuits for computation, and matching the temporal dynamics of the system to the signals being processed.

How does Giacomo Indiveri's approach to neuromorphic computing address the issue of device variability and noise?

Giacomo Indiveri's approach addresses device variability and noise by using populations of neurons and averaging over time and space, which can reduce the effect of device mismatch and noise, and by exploiting the variability as an advantage for robust computation.

What are some of the practical applications of neuromorphic computing systems discussed in the talk?

Some practical applications of neuromorphic computing systems discussed include ECG anomaly detection, vibration anomaly detection, industrial monitoring, intelligent machine vision, and consumer applications.

What is the Dynamic Neuromorphic Asynchronous Processor (DNAP) and what is its significance?

The Dynamic Neuromorphic Asynchronous Processor (DNAP) is an academic prototype built at the University of Zurich and ETH Zurich. It is significant as it demonstrates the feasibility of neuromorphic computing with a thousand neurons organized in four cores, showcasing the potential for edge computing applications with low power consumption.

How does Giacomo Indiveri's research contribute to the field of neuromorphic intelligence?

Giacomo Indiveri's research contributes to the field of neuromorphic intelligence by developing new architectures, packaging systems, and memory devices that are inspired by the brain's computational principles, aiming to create more efficient and powerful computing systems.

Outlines

00:00

🧠 Introduction to Brain-Inspired AI and Energy Efficiency

Giacomo Indiveri from the University of Zurich and ETH Zurich introduces the concept of brain-inspired strategies for low-power artificial intelligence computing systems. He discusses the history of AI, highlighting the resurgence of interest due to hardware advancements that provide the necessary computing power. The talk emphasizes the unsustainable energy consumption of current AI systems, which is projected to reach 20% of the world's energy by 2025. Indiveri points out the inefficiency of data movement between memory and processing units as a significant contributor to this energy consumption.

05:00

🌟 The Emergence of Neuromorphic Computing

The concept of neuromorphic computing is explored, which was first coined by Carver Mead in the 1980s. Neuromorphic computing aims to emulate the brain's efficiency in processing information. Three main communities are identified: those designing CMOS circuits to emulate brain functions, those building practical devices for problem-solving using spiking neural networks, and those developing nanoscale devices for non-volatile memory. The talk suggests that by combining new materials, architectures, and theories, neuromorphic computing could lead to more efficient and powerful AI systems.

10:01

🔄 The Temporal Dynamics of Biological vs. Artificial Neural Networks

Indiveri contrasts the temporal dynamics and physical computation of biological neural networks with the algorithmic simulation of artificial neural networks. He explains that while artificial networks simulate neuron properties, biological networks use the physical properties of neurons and synapses for computation. The talk emphasizes the importance of understanding these differences to develop more efficient computing systems, suggesting that the structure of biological networks is inherently the algorithm, unlike in artificial networks.

15:03

🔋 Strategies for Low-Power Neuromorphic Systems

The talk outlines strategies for creating low-power neuromorphic systems, focusing on the co-localization of computation and memory, the use of analog circuits for efficient computation, and the importance of temporal dynamics matching the processing needs. Indiveri argues for a paradigm shift in computing, moving from traditional CPU-memory architectures to brain-inspired architectures that interweave memory and computation in a distributed system, leading to significant power savings.

20:05

🤖 Practical Implementations and Applications of Neuromorphic Systems

Indiveri discusses practical implementations of neuromorphic systems, including the development of analog circuits with slow, non-linear dynamics that operate in parallel. He highlights the benefits of device mismatch and variability, which can be harnessed for robust computation. The talk also covers the advantages of using analog circuits over digital ones, such as avoiding the need for clock circuits, reducing the need for data conversion, and leveraging the physics of devices for complex operations. Indiveri presents examples of neuromorphic systems developed at the University of Zurich and ETH Zurich, including chips that can perform edge computing applications with low power consumption.

25:06

🚀 Advancing Neuromorphic Intelligence and Its Applications

The talk concludes with a look at the future of neuromorphic intelligence, emphasizing the potential to transfer academic knowledge into practical technology through startups. Indiveri mentions applications such as industrial monitoring, machine vision, and consumer electronics as promising areas for neuromorphic technology. He also discusses the Dynamic Neuromorphic Asynchronous Processor, an academic prototype that demonstrates the potential of neuromorphic systems for edge computing. The talk ends with a call to action for further development and application of neuromorphic intelligence in society.

Keywords

💡Artificial Intelligence (AI)

Artificial Intelligence refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the context of the video, AI is a central theme, with a focus on how AI algorithms and networks have evolved over time and their current impact on various fields, particularly in machine vision. The script mentions the historical development of AI, dating back to the late 80s, and the recent advancements that have led to AI systems outperforming traditional methods.

💡Convolutional Neural Networks (CNNs)

Convolutional Neural Networks are a class of deep neural networks most commonly applied to analyzing visual imagery. The video script highlights the breakthroughs in CNNs starting from 2009, where they began to achieve impressive results in the field of machine vision, leading to a surge in AI's popularity and application.

💡Backpropagation

Backpropagation is a computing technique used to train artificial neural networks through the gradient descent optimization. It is mentioned in the script as a fundamental algorithm that, despite its success, has limitations when compared to the natural intelligence found in biological systems. The speaker suggests that incremental improvements to backpropagation may not lead to breakthroughs in achieving natural intelligence.

💡Neuromorphic Computing

Neuromorphic Computing is an approach to computing that mimics the neuro-biological architectures present in the nervous system. The video discusses neuromorphic computing as a promising strategy for developing low-power AI systems, drawing inspiration from the brain's efficient use of resources and its ability to perform complex computations with minimal energy.

💡Energy Consumption

Energy Consumption in the context of the video refers to the significant amount of power required to train and run AI algorithms, particularly neural networks. The script estimates that by 2025, the ICT industry's energy consumption could account for 20 percent of the world's total energy use, highlighting the need for more sustainable and energy-efficient computing strategies.

💡Memory-Computation Co-location

Memory-Computation Co-location is a design principle where memory and computation occur in the same physical location, reducing the need to transfer data between separate memory and processing units. The video emphasizes this concept as a key strategy in neuromorphic computing to reduce energy consumption, drawing an analogy to how information is processed within neurons in the brain.

💡Spiking Neural Networks (SNNs)

Spiking Neural Networks are a type of artificial neural network that mimic the way biological neurons communicate using 'spikes' or brief electrical pulses. The script discusses SNNs as part of the neuromorphic approach, where digital circuits simulate the spiking behavior of neurons to create more efficient computing systems.

💡In-Memory Computing

In-Memory Computing is a paradigm where data is processed in the memory subsystem itself, reducing the need for data movement between memory and processing units. The video script describes how in-memory computing technologies can be used to implement long-term, non-volatile memories, contributing to the development of neuromorphic systems.

💡Analog Circuits

Analog Circuits are electronic circuits that process signals in a continuous and variable manner, as opposed to digital circuits that process signals in discrete levels. The video discusses the use of analog circuits in neuromorphic computing to emulate the physics of biological neurons and synapses, enabling more efficient and low-power computation.

💡Temporal Dynamics

Temporal Dynamics refers to the time-dependent behavior of a system. In the video, the importance of matching the temporal dynamics of computing systems to the characteristics of the signals being processed is emphasized. For instance, the video mentions how the brain's temporal dynamics are integral to its low-power and efficient computation.

💡Fault Tolerance

Fault Tolerance is the ability of a system to continue operating properly in the event of hardware or software faults. The script discusses how neuromorphic systems, with their distributed and parallel nature, can be inherently fault-tolerant, allowing them to continue functioning even if some components fail, which is a stark contrast to traditional computing systems.

Highlights

The talk focuses on brain-inspired strategies for low-power artificial intelligence computing systems.

Artificial intelligence algorithms and networks have their roots in the late 80s but only recently started outperforming conventional approaches.

The success of AI is attributed to advancements in hardware technologies, availability of large data sets, and improvements in algorithms.

AI algorithms require significant memory and energy resources, leading to concerns about sustainability.

The ICT industry is projected to consume about 20 percent of the world's energy by 2025.

The movement of data between memory and computing units is a major factor in energy consumption.

AI algorithms are narrow and specialized, unlike the general-purpose nature of animal brains.

The backpropagation algorithm is a fundamental component of AI but differs from the brain's computational principles.

Neuromorphic engineering aims to emulate the brain's efficiency and could lead to breakthroughs in AI.

Neuromorphic systems use analog circuits and in-memory computing to reduce power consumption.

Biological neural networks use time dynamics and the physics of their elements, unlike simulated networks.

The brain's structure and architecture are intertwined with its computational processes, unlike traditional computer systems.

Neuromorphic systems use parallel arrays of processing elements with computation and memory co-localized.

Analog circuits in neuromorphic systems can perform complex operations more efficiently than digital counterparts.

The temporal domain is crucial in neuromorphic systems, with dynamics matched to the signals being processed.

Neuromorphic systems can achieve fast reaction times even with slow elements due to parallel processing.

The University of Zurich and ETH Zurich have been developing neuromorphic systems for edge computing applications.

Neuromorphic chips can be used for practical applications like industrial monitoring and intelligent machine vision.

Startups are now utilizing neuromorphic technology to solve real-world problems with low-power, high-efficiency AI systems.

Transcripts

hello

this is giacomo individi from the

university of zurich and eth zurich at

the institute of neuro-informatics it's

going to be a pleasure for me to give

you a talk about brain inspired

strategies for low-power artificial

intelligence computing systems

so the term artificial intelligence

actually has become very very popular in

recent times

in fact artificial intelligence

algorithms and networks go back to the

late 80s

and although the first successes of

these networks were demonstrated in the

80s only recently

these algorithms and the computing

systems started to outperform

conventional

approaches for solving problems

in fact in

the field of machine vision uh from 2009

on

in 2011 in fact the first convolutional

neural network trained using back

propagation achieved impressive results

that made the whole field explode

and the reason for the success of this

approach

really even though as i said it was

started many many years ago only

recently we started to be able to follow

this

success because um and achieve really

impressive performance

because the technologies the hardware

technologies started to provide

enough computing power for these

networks to actually really perform well

in addition there's been now

the availability of large data sets that

can be used to train such networks which

also were not there in the 80s

and finally

several tricks and hacks and

improvements in the algorithms have been

proposed to actually

make these networks very robust and very

performant

however they do have some problems

most of these algorithms require a large

amount of resources in terms of memory

and energy to be trained

and in fact if we

do an estimate and we try to see how

much energy is required by all of the

computational devices that are in the

world

to implement such

neural networks it is estimated that by

2025 the ict industry will consume about

20 percent of the entire world's energy

this is clearly a problem which is not

sustainable

the other

reason for or one of the main reasons

for these networks to be extremely power

hungry is because they are

requiring large amounts of data and

memory resources and in particular

they're required to move data from the

memory to the computing and from the

computing back to the memory so

typically memory is used

in dram chips and these dram are at

least a thousand five hundred times more

costly than any compute operation mac

operations in these cnn accelerators

so it's it's really not the fact that

we're doing lots of computation it's

it's really the fact that we're moving

bits uh back and forth

that is is

burning all of this energy

the other problem is more fundamental

it's not only related to technology it's

related to the theory and these

algorithms these algorithms actually as

i said are actually are very very uh

powerful in terms of recognizing images

and solving but they are very narrow in

the sense that they are very specialized

to only a very specific domain

these networks are programmed to perform

a limited set of tasks and they operate

within a predetermined and predefined

range

they are not nearly as general purpose

as uh animal brains are so even though

we do we do call it artificial

intelligence it's really different from

natural intelligence the type of

intelligence that we see in in animals

and in humans

and the backbone of these artificial

intelligence algorithms is the back

propagation algorithm or if we're

looking at time series and sequences the

back propagation through time bpttt

algorithm

this this is uh really an algorithmic

limitation even though it can be used to

solve very powerful problems

trying to improve this bpdt

by making incremental changes is

probably not going to lead to

breakthroughs in understanding how to go

from artificial intelligence to natural

intelligence

and the way the brain works is actually

quite different from back propagation

through time

if you look at neuroscience and if you

study real neurons and real synapses

and

the sort of computational principles of

the brain you will realize that it's

there there's a big difference so

this problem has been recognized by many

communities many agencies there is for

example a recent paper by john shelf

that shows how to go beyond these

problems and try to improve performance

of computation

and if we look at this particular path

that basically tries to put together new

architectures new packaging systems with

new memory devices and new theories

one of the most promising approaches is

the one that is here listed as

neuromorphic

so what is this neuromorphic this is the

sort of the the bulk of this talk that i

am going to show you what we can do at

the university of zurich but also at the

startup that comes out of the university

of zurich instance with this type of

approach which is as i said taking the

best of the new materials and devices

new architectures and new theories and

trying to go really beyond what we have

today

so the term neuromorphic was actually

invented or coined many many years ago

by carver mead in the late 80s

and is now being used to describe

different things there's at least three

big communities that are using the term

neuromorphic

the original one that goes back to

carver mead was referring to the design

of cmos electronic circuits that were

used to emulate the brain

basically as a basic research attempt to

try to understand how the brain works by

building

circuits that are equivalent so trying

to really reproduce the physics

and and because of that these circuits

were using sub-threshold analog

transistors

for the neurodynamics and the

computation and asynchronous digital

logic for communicating spikes across

chips across cores it was really

fundamental research

the other big community that now started

to use the term neuromorphic is the

community building uh

practical devices for you know solving

practical problems

in that case these these this community

is is building chips that can implement

spiking neural network accelerators or

simulators not emulation but but really

now at this point it's it's more an

exploratory approach it's being used to

try to understand what can be done

with this approach of using digital

circuits to simulate spiking neural

networks

finally the last community or another

large community that is started to use

the term neuromorphic is the one that

has been developing emerging memory

technologies looking at nanoscale

devices to implement long-term

non-volatile memories

or if you like memoristive devices

so this community also started using the

turn neuromorphic because these devices

they can actually store

a change in the conductance which is

very similar to the way the real

synapses work when that they actually

change their conductance when they

change their synaptic weight

and

this allows them to build in memory

computing architectures that are also as

you will see very similar to the way

real biological neural networks work and

it can really create high density arrays

so we can actually by using analog

circuits the approach of simulating

digital um spike neural networks and by

using in-memory computing technologies

the hope is that we create a new field

which i'm calling here neuromorphic

intelligence that will lead to the

creation of

compact intelligent brain inspired

devices

and really to understand how to do these

brain inspired devices it's important to

look at the brain to go back to carver

meats approach and really do fundamental

research in studying biology and try to

really get the best out of all all of

these communities of the devices of the

sort of the computing principles using

simulations and and machine learning

approaches but also of neuroscience and

studying the brain

and so here i'd like to just to

highlight the main differences that are

there between

simulated artificial neural networks and

really

the biological neural networks those

that are in the brain

in simulated artificial neural networks

as you probably know there is a weighted

sum of inputs the inputs are all coming

in a point neuron which is basically

just doing the sum or the integral of

the inputs and multiplying all of them

by a weight so it's it's really

characterized by a big uh weight

multiplication or matrix multiplication

operation and then there is a

non-linearity either a spiking

non-linearity if it's a spike neural

network or a thresholding non-linearity

if it's an artificial neural network

in biology the neurons are also

integrating all of their synaptic inputs

with different weights so there is this

analogy of weighted inputs but it's all

happening through the physics of the

devices so the the physics is playing an

important role for computation

the synapses are not just doing a

multiplication they're actually

implementing some temporal operators

integrating applying non-linearities uh

dividing summing it's much more

complicated than just a weighted sum of

inputs

in addition the neuron actually has an

axon and it's sending its output through

the axon using also basically an all or

none event a spike

through time because the longer the axon

the longer it will take for the for the

spike to travel and reach the

destination

and depending on how thick the axon is

how much myelination there is it will be

slower or faster so also here the

temporal dimension is is really

important

in summary if we really want to see the

big difference is that artificial neural

networks the one that are being

simulated on computers and gpus are

actually algorithms that simulate some

basic properties of real neurons

whereas biological neural networks

really use time dynamics and the physics

of their computing elements to run the

algorithm actually in these networks the

structure of the architecture is the

algorithm there is no distinction

between the hardware and the software

everything is one and understanding how

to build

these types of hardware architectures

wet wear or hardware

using cmos using memristors maybe even

using alternative you know dna computing

or other approaches

will

hopefully and probably lead to much more

efficient and powerful computing systems

compared to the artificial neural

networks so if we want to understand how

to do this we really need to do a

radical paradigm shift in computing

standard computing architectures are

basically based on the phenomenon

system where you have a cpu on one side

and

memory uh on the other and as i said

transferring data back and forth from

the cpu to the memory and back is

actually what's burning all the power

doing the computation inside the cpu is

much much uh more energy efficient and

less costly than transferring the data

in brains what's happening is that

inside the neuron there are synapses

which store

the value of the weight so

memory and computation are co-localized

there is no transfer of data back and

forth everything is happening at the

synapse at the neuron and there is many

distributed synapses as many distributed

neurons so the memory and the

computation are intertwined together in

a distributed system

and this is really a big difference so

if we want to understand how to really

save power we have to look at how the

brain does it we have to use these brain

inspire strategies and the main three

points that i'd like you to remember is

that you we have to use basically

parallel arrays of processing elements

that have computation and memory

co-localized and this is radically

different from time multiplexing a

circuit here for example if we have one

cpu two cpus but even 64 cpus to

simulate

thousands of neurons we are time

multiplexing the integration of the

differential equations in these 64 cpus

here if we look at how to do it

following this brain inspired strategies

if we want to emulate a thousand neurons

we really have to have a thousand

different circuits that are laid out in

the in the layout of the of the chip of

the wafer and then run these

through their physics through the

physics of the circuits analog circuits

digital circuits but they have to be

many parallel circuits that operate in

parallel with the memory and the

computation co-localized that's really

the trick to to save power

the other is if we have analog circuits

we can use the physics of the circuits

to carry out the computation that really

instead of abstracting away some

differential equations and integrating

numerically the differential equation we

really use the physics of the device to

carry out the computation it's much more

efficient in terms of power latency time

and area

and finally the temporal domain is

really important the temporal dynamics

of the system have to be well matched to

the signals that we want to process so

if we want to have very low power

systems and for example we want to

process speech we have to have elements

in our computing substrate in our brain

like computer that have the same time

constants speech for example phonemes

have time constants of the order of 50

milliseconds so we have to slow down

silicon to have dynamics and time

constants of the order of 50

milliseconds so our our chips will be

firing and going at you know hertz or

maybe hundreds of hertz but definitely

not that megahertz or gigahertz like our

cpus or our gpus are doing

and and by having parallel arrays of

very slow elements we can still get very

fast computation even if we have slow

elements it doesn't mean that we don't

have a very fast reactive system

it's because they're in working in

parallel and so at some point there will

always be one or two of these elements

that are about to fire whenever the

input arrives and we can have

microsecond nanosecond reaction times

even though we have millisecond dynamics

and this is another key trick to

remember if we want to understand how to

do this radical paradigm shift

and at the university of zurich at eth

at the institute of neuroinformatics

we've been building these types of

systems

for many many years and we are uh now

also building these systems at our new

startup at since

the type of systems are shown here

basically we create arrays of neurons

with analog circuits

these circuits as i told you are slow

they have slow temporal non-linear

dynamics

as i told you they are massively

parallel we do massively parallel

operations all of the circuits work in

parallel

the fact that they are analog actually

brings this

feature that are that is basically

device mismatch all the circuits are

inhomogeneous they are not equal and

this actually can be used as an

advantage to carry out robust

computation it's counter-intuitive but i

will show you that it's actually an

advantage to have variability in your

devices and this actually is also very

nice for people doing memory still

devices that are typically very very

variable

the other features are that they are

adaptive all of these circuits that we

have have negative feedback loops they

have learning

adaptation plasticity so the learning

actually helps in creating robust

computation through the noisy and

variable elements

by construction there are many of these

in working in parallel even if some of

these stop working the system is fault

tolerant you don't have to throw away

the chip like you would with a standard

processor if one transistor breaks

probably performance will degrade

smoothly but at least the system will be

fault tolerant and because we use both

the best of both worlds analog circuits

for the dynamics and digital circuits

for the communication we can program the

routing tables and configure these

networks so we have flexibility in being

able to program these dynamical systems

like you would program a neural network

on a cpu on a computer

uh of course it's it's more complex we

still have to develop all the tools and

and simpsons and and other colleagues

around the world are still busy

developing the tools to program these

dynamical systems

it's not nearly as well developed as you

know having a java or a c or a python

piece of code but

there is very promising work going on

and now the question always comes why do

you do it if analog is noisy and

annoying in homogeneous why do you go

through the effort of building these

analog circuits so let me just try to

explain that there are several

advantages if you think of having large

networks in which you have many elements

working in parallel for example these

memorizative devices in a crossbar array

and you want to send data through them

these membership devices if you use the

physics they use analog variables so

if you just send these variables in an

asynchronous mode you don't need to use

a clock so you can avoid using digital

clock circuitry which is actually

extremely expensive in terms of area

requirements in large complex chips and

extremely power hungry so avoiding

clocks is is something really really

useful

if we don't use digital if we're staying

analog all the way from the input to the

output we don't need to convert we don't

need to convert from digital to analog

to to run the physics of these

memristors and we don't need to convert

back from analog to digital and these

adcs and these dacs are actually very

expensive in terms of size and power so

again if we don't use them we save in

size and power

if we use transistors to do for example

exponentials we don't need to have a

very complicated uh digital circuitry to

do that so again we can use a single

device that through the physics of the

device can do complex nonlinear

operation that saves area and power as

well

and finally if we have analog variables

like variable voltage heights variable

voltage pulses widths

and and

other types of variable currents we can

control the properties of the

of the devices that we use these

memorized devices we can make them uh

depending on how strong we we drive them

we can make them volatile or

non-volatile we can use their intrinsic

non-linearities

depending on how strongly we derive them

we can even make them switch with a

probability so we can use their

intrinsic stochasticity to do stochastic

gradient descent or to do probabilistic

graphical networks to do probabilistic

computation

and we can also use them in their

standard way of operation in their

non-volatile way of operation as

long-term memory elements so we don't

need to shift data back and forth from

peripheral memory back we can just store

the value of the sinuses directly in

these membrane devices

so if we use analog for our neurons and

synapses in cmos

we can then best best benefit the use of

future emerging memory technologies uh

reducing power consumption

and uh in a very recent in the last

escas conference which was just you know

a few weeks ago we did show with the pcm

trace algorithm or or the pcm series

experiments that we can exploit the

drift of pcm devices which are these

shown here in a picture from ibm

to implement eligibility traces which is

a very useful feature to have for

reinforcement learning

so if if we are interested in building

reinforcement learning algorithms for

example for for having behaving robots

that run with

brains that are implemented using these

chips we can actually take advantage of

the properties of these pcm devices

none that are typically thought as

non-idealities we can use them to our

advantage for computation now

analog circuits are noisy i told you

there are the they are variable in

homogeneous for example if you take one

of our chips you you stimulate the

neurons with the same current to all the

neurons to 256 different neurons and you

see how long it takes for the neuron to

fire

not only these neurons are slow but

they're also variable the time at which

they fire can greatly change depending

on which circuit you're using

and there is this noise which is usually

typically you have variance of 20

over the mean so the coefficient of

variation is about 20 percent

so the question is how can you do robust

computation using this noisy

computational substrate and the obvious

answer the easiest thing that people do

when they have noise is to average

so we can do that we can average over

space and we can average over time if we

use populations of neurons not single

neurons we can just take you know two

three four six eight neurons and look at

the average time that it took for them

to spike or if they're spiking uh

periodically we can look at the average

firing rate

and then if we integrate over long

periods of time we can average over time

so these these two strategies are going

to be useful for uh reducing the effect

of device mismatch if we do need to have

precise computation

and we are doing experiments in this in

these very days this is actually a very

recent experiment where we took these

neurons and we put two of them together

four of them together eight sixteen if

you look at the cluster size basically

that's the number of neurons that we are

using for the average over space

and then we are computing the um

firing rate over two milliseconds five

milliseconds 50 milliseconds 100 and so

on and then what we do is we calculate

the coefficient of variation basically

how much device mismatch there is the

larger the coefficient of variation the

more noise the smaller the coefficient

of variation the less noise so we can go

from a very large coefficient of

variation of say 12

actually 18 as i said 20

by integrating over long periods of time

or by integrating over large numbers of

neurons we can decrease this all the way

to 0.9

and you can take this coefficient of

variation and you can calculate the

equivalent number of bits if we were

using digital circuits how many would

bits would this correspond to so for by

just integrating over larger number of

neurons and over longer periods we can

have for example a sweet spot where we

have eight bit resolution just by using

16 neurons

and integrating for example over 50

milliseconds

this can be changed at runtime if we if

we want to have a very fast reaction

time and a course idea of what the

result is we can have only two neurons

and integrate only over two milliseconds

then there's many false myths when we we

use spike neural networks people tell us

oh but if you have to wait until you

integrate enough it's going to be slow

if you have to average over time it's

going to take area all of these are

actually false myths that can be

debunked by looking at neuroscience

neuroscience has been studying how the

brain works the brain is extremely fast

it's extremely low power we don't have

to wait long periods of time to make a

decision

so if you use populations of neurons to

average out it's been shown for example

experimentally with real learners

that populations of neurons have

reaction times that can be even 50 times

faster than single neurons

so by using populations we can really

speed up the computation

if we use the populations of neurons we

don't need them to be every neuron to be

very highly accurate they can be noisy

and they can be very low precision but

by

using populations and using sparse

coding we can have very accurate

representation of data and there is a

lot of work for example by sophie the

nerve that has been showing how to do

this by by training populations of

neurons to do that

and now it should be also known in

technology that if you have variability

it actually helps in transferring

information across multiple layers

and here what i'm showing here is data

from one of our chips where we use 16 of

these neurons per core and we basically

provide some in desired target as the

input we drive a motor with a pid

controller and we minimize the error

it's just to show you that by using

spikes we can actually have very fast

reaction times in robotic platforms

using these types of uh chips that you

saw in the previous slides

in fact we've been building many chips

for many years also at since the the the

colleagues are building chips and the

latest one that we have built at the

university as a academic prototype

is um

is called the dynamic neuromorphic

asynchronous processor it was built

using a very old

technology 180 nanometer as i said

because it's an academic exercise but

still this has a thousand neurons it has

four cores of 256 neurons each we can

actually do very interesting edge

computing applications just with a few

hundred neurons and then of course then

the idea is to use both the best of both

worlds to see where we can do

analog circuits to really have low power

or digital circuits to have you know to

verify principles and make you know

practical problems solve practical

problems quickly and then by combining

analog and digital we can also there get

the best of both worlds

so to to conclude actually i would just

like to show you examples of

applications that we built

using this

this is

a long list of applications but if you

have

the slides you can actually click on the

on the references and you can get the

paper the ones highlighted in red are

the ones that have been done by simpsons

uh by our colleagues from instance on

ecg anomaly detection and the detection

of vibration anomalies was actually done

by simpsons and by university of zurich

in parallel independently and just go to

the last slide where i basically tell

you that we are now at a point where we

can actually use all of the knowledge

from the university about brain inspire

strategies

to develop this neuromorphic

intelligence field transfer all of this

know-how into

technology and in the in with new

startups that can actually use this

know-how to solve practical problems

and try to find you know what is the

best market for this

as i said industrial monitoring for

example for vibrations is something but

something that can also be done using

both sensors and processors is is really

what the simpsons company has been

developing and it's it's been really

successful at for doing intelligent

machine vision for putting both uh

sensing and

artificial intelligence algorithm on the

same chip

and having a very low power in the order

of microwave uh tens or hundreds of

microwatt power dissipation for solving

practical problems that can be useful in

society and in fact even solving

consumer applications

so with this um sorry i went a bit over

time but i would just like to thank you

for your attention

you