Neuromorphic Intelligence: Brain-inspired Strategies for AI Computing Systems

00:27:40
https://www.youtube.com/watch?v=6lLHVcBkwgI

Summary

TLDRI denne præsentation diskuterer Giacomo Indyvidi fra universitetet i Zürich banebrydende innovationer indenfor hjerneinspireret teknologi til kunstig intelligens. Traditionelle kunstige netværk, kendt siden 1980'erne, er blevet begrænset af deres enorme krav til energi og hukommelse. Ifølge Indyvidi kan vi ved at drage nytte af neuromorfik teknologi, som kopierer biologiske hjerners energibesparende mekanismer, opnå en betydelig forbedring. Neuromorfik tilgange integrerer både hardware og beregning tættere sammen, hvilket potentielt kan skabe intelligente enheder med lavt energiforbrug, egnet til en bred vifte af industrielle og forbrugsmæssige applikationer. Dette kan hjælpe med at reducere kunstig intelligensens energifodaftryk betydeligt. Desuden kan videre forskning inden for neuromorfik og analog computing åbne nye døre for robuste og fleksible AI-løsninger der fungerer effektivt under realistiske forhold.

Takeaways

  • 🧠 Neuromorfik teknologi efterligner hjernen for at spare energi i kunstig intelligens.
  • ⚙️ Hjerneinspirerede strategier bruger parallelle kredsløb til at samlokalisere hukommelse og beregning.
  • 🔋 Kunstige neurale netværk kræver meget energi grundet deres datakrævning.
  • 💽 Analog teknologi kan reducere omkostningen ved dataoverførsel i AI-systemer.
  • 📉 Problemet med energiintensiv AI er ikke i beregningen, men i dataflytningen.
  • 🌐 Neuromorfik forskning omfatter design på tværs af hardware og algoritmiske strukturer.
  • 🔄 Variabilitet i kredsløb kan bruges til robust beregning ved hjælp af gennemsnitsteknikker.
  • 🌍 Potentialet for energibesparelse i AI kan reducere den globale teknologis energiforbrug.
  • 🎓 Forskning i Zürich viser praktiske anvendelser af hjernensinspireret computing.
  • 🚀 Neuromorfisk intelligens kan føre til mere effektive og kraftfulde systemer.

Timeline

  • 00:00:00 - 00:05:00

    Foredraget handler om hjerneinspirerede strategier for lav-energi kunstig intelligens. AI er blevet populært, men har sine rødder i 1980'erne. Succesen skyldes forbedret hardware og større datasæt, men AI er energikrævende, og fremtidens energiforbrug fra AI-systemer kan blive uholdbart.

  • 00:05:00 - 00:10:00

    Neuromorphic computing er en løsning, der kombinerer nye materialer og teorier. Dette felt, skabt af Carver Mead, bruger CMOS kredsløb til at imitere hjernens funktioner. Det sigter mod at forbedre forståelsen og effektiviteten af computing gennem biologiinspirerede metoder.

  • 00:10:00 - 00:15:00

    Hovedforskellene mellem neurale netværk i biologiske hjerner og kunstige neurale netværk er måden de bearbejder information på. Biologiske systemer integrerer tid og fysik på en måde, der sammensmelter hardware og software, hvilket kunstige systemer ikke gør.

  • 00:15:00 - 00:20:00

    Strategier til at udvikle energieffektive systemer inkluderer brug af parallelle, analoge kredsløb, der anvender fysikkens love. Målet er at reducere energiforbruget ved at efterligne hjernens måde at integrere hukommelse og beregning på.

  • 00:20:00 - 00:27:40

    Projektet på universitetet i Zürich udvikler neurale netværk baseret på analoge kredsløb. Dette inkluderer robust beregning på trods af støj og fejl i hardware. Målet er at kombinere analoge og digitale systemer for at løse praktiske problemer effektivt.

Show more

Mind Map

Mind Map

Frequently Asked Question

  • Er kunstige neurale netværk lige så alsidige som menneskelige hjerner?

    Nej, de er stadig meget specialiserede sammenlignet med dyre- og menneskehjerner.

  • Hvordan kan hjerneinspirerede strategier spare energi i AI?

    Effektiviteten forbedres ved at designe parallelle kredsløb, der samtidig udfører beregning og hukommelsesopgaver ligesom hjerner.

  • Hvordan bidrager analoge kredsløb til energibesparelse?

    Ved at integrere analog beregning for non-lineariteter og reducere behovet for adskilte digitale konverteringer.

  • Hvilke udfordringer har kunstige neurale netværk?

    De begrænses af energiforbrug, data- og hukommelsesbehov, samt specialisering til bestemte opgaver.

  • Hvordan håndteres variabiliteten i analoge enheder til beregning?

    Mengem varianter af devices bruges til robust beregning og udvinding af mønstre, ligesom i hjerneaktivitet.

  • Hvad er formålet med neuromorfik forskning?

    Neuromorfik udforskninger er ofte fokuseret på at efterligne hjernens funktion for at forbedre hardwarearkitekturer.

View more video summaries

Get instant access to free YouTube video summaries powered by AI!
Subtitles
en
Auto Scroll:
  • 00:00:04
    hello
  • 00:00:05
    this is giacomo individi from the
  • 00:00:07
    university of zurich and eth zurich at
  • 00:00:09
    the institute of neuro-informatics it's
  • 00:00:11
    going to be a pleasure for me to give
  • 00:00:12
    you a talk about brain inspired
  • 00:00:14
    strategies for low-power artificial
  • 00:00:16
    intelligence computing systems
  • 00:00:19
    so the term artificial intelligence
  • 00:00:21
    actually has become very very popular in
  • 00:00:24
    recent times
  • 00:00:25
    in fact artificial intelligence
  • 00:00:27
    algorithms and networks go back to the
  • 00:00:29
    late 80s
  • 00:00:30
    and although the first successes of
  • 00:00:32
    these networks were demonstrated in the
  • 00:00:34
    80s only recently
  • 00:00:36
    these algorithms and the computing
  • 00:00:38
    systems started to outperform
  • 00:00:40
    conventional
  • 00:00:41
    approaches for solving problems
  • 00:00:44
    in fact in
  • 00:00:46
    the field of machine vision uh from 2009
  • 00:00:49
    on
  • 00:00:50
    in 2011 in fact the first convolutional
  • 00:00:53
    neural network trained using back
  • 00:00:55
    propagation achieved impressive results
  • 00:00:58
    that made the whole field explode
  • 00:01:00
    and the reason for the success of this
  • 00:01:03
    approach
  • 00:01:04
    really even though as i said it was
  • 00:01:06
    started many many years ago only
  • 00:01:09
    recently we started to be able to follow
  • 00:01:11
    this
  • 00:01:12
    success because um and achieve really
  • 00:01:15
    impressive performance
  • 00:01:16
    because the technologies the hardware
  • 00:01:18
    technologies started to provide
  • 00:01:21
    enough computing power for these
  • 00:01:22
    networks to actually really perform well
  • 00:01:25
    in addition there's been now
  • 00:01:28
    the availability of large data sets that
  • 00:01:30
    can be used to train such networks which
  • 00:01:32
    also were not there in the 80s
  • 00:01:34
    and finally
  • 00:01:36
    several tricks and hacks and
  • 00:01:38
    improvements in the algorithms have been
  • 00:01:39
    proposed to actually
  • 00:01:41
    make these networks very robust and very
  • 00:01:44
    performant
  • 00:01:46
    however they do have some problems
  • 00:01:49
    most of these algorithms require a large
  • 00:01:52
    amount of resources in terms of memory
  • 00:01:55
    and energy to be trained
  • 00:01:58
    and in fact if we
  • 00:02:00
    do an estimate and we try to see how
  • 00:02:02
    much energy is required by all of the
  • 00:02:04
    computational devices that are in the
  • 00:02:06
    world
  • 00:02:06
    to implement such
  • 00:02:09
    neural networks it is estimated that by
  • 00:02:12
    2025 the ict industry will consume about
  • 00:02:14
    20 percent of the entire world's energy
  • 00:02:17
    this is clearly a problem which is not
  • 00:02:19
    sustainable
  • 00:02:21
    the other
  • 00:02:22
    reason for or one of the main reasons
  • 00:02:24
    for these networks to be extremely power
  • 00:02:27
    hungry is because they are
  • 00:02:29
    requiring large amounts of data and
  • 00:02:31
    memory resources and in particular
  • 00:02:33
    they're required to move data from the
  • 00:02:36
    memory to the computing and from the
  • 00:02:38
    computing back to the memory so
  • 00:02:40
    typically memory is used
  • 00:02:42
    in dram chips and these dram are at
  • 00:02:45
    least a thousand five hundred times more
  • 00:02:46
    costly than any compute operation mac
  • 00:02:50
    operations in these cnn accelerators
  • 00:02:53
    so it's it's really not the fact that
  • 00:02:55
    we're doing lots of computation it's
  • 00:02:57
    it's really the fact that we're moving
  • 00:02:58
    bits uh back and forth
  • 00:03:01
    that is is
  • 00:03:02
    burning all of this energy
  • 00:03:05
    the other problem is more fundamental
  • 00:03:07
    it's not only related to technology it's
  • 00:03:09
    related to the theory and these
  • 00:03:10
    algorithms these algorithms actually as
  • 00:03:13
    i said are actually are very very uh
  • 00:03:15
    powerful in terms of recognizing images
  • 00:03:18
    and solving but they are very narrow in
  • 00:03:21
    the sense that they are very specialized
  • 00:03:23
    to only a very specific domain
  • 00:03:26
    these networks are programmed to perform
  • 00:03:28
    a limited set of tasks and they operate
  • 00:03:30
    within a predetermined and predefined
  • 00:03:32
    range
  • 00:03:33
    they are not nearly as general purpose
  • 00:03:35
    as uh animal brains are so even though
  • 00:03:39
    we do we do call it artificial
  • 00:03:41
    intelligence it's really different from
  • 00:03:43
    natural intelligence the type of
  • 00:03:45
    intelligence that we see in in animals
  • 00:03:47
    and in humans
  • 00:03:49
    and the backbone of these artificial
  • 00:03:51
    intelligence algorithms is the back
  • 00:03:53
    propagation algorithm or if we're
  • 00:03:56
    looking at time series and sequences the
  • 00:03:58
    back propagation through time bpttt
  • 00:04:01
    algorithm
  • 00:04:03
    this this is uh really an algorithmic
  • 00:04:06
    limitation even though it can be used to
  • 00:04:08
    solve very powerful problems
  • 00:04:11
    trying to improve this bpdt
  • 00:04:14
    by making incremental changes is
  • 00:04:17
    probably not going to lead to
  • 00:04:19
    breakthroughs in understanding how to go
  • 00:04:21
    from artificial intelligence to natural
  • 00:04:24
    intelligence
  • 00:04:25
    and the way the brain works is actually
  • 00:04:27
    quite different from back propagation
  • 00:04:29
    through time
  • 00:04:30
    if you look at neuroscience and if you
  • 00:04:32
    study real neurons and real synapses
  • 00:04:35
    and
  • 00:04:36
    the sort of computational principles of
  • 00:04:38
    the brain you will realize that it's
  • 00:04:40
    there there's a big difference so
  • 00:04:43
    this problem has been recognized by many
  • 00:04:45
    communities many agencies there is for
  • 00:04:48
    example a recent paper by john shelf
  • 00:04:50
    that shows how to go beyond these
  • 00:04:52
    problems and try to improve performance
  • 00:04:55
    of computation
  • 00:04:56
    and if we look at this particular path
  • 00:04:58
    that basically tries to put together new
  • 00:05:00
    architectures new packaging systems with
  • 00:05:02
    new memory devices and new theories
  • 00:05:05
    one of the most promising approaches is
  • 00:05:07
    the one that is here listed as
  • 00:05:08
    neuromorphic
  • 00:05:10
    so what is this neuromorphic this is the
  • 00:05:13
    sort of the the bulk of this talk that i
  • 00:05:16
    am going to show you what we can do at
  • 00:05:18
    the university of zurich but also at the
  • 00:05:21
    startup that comes out of the university
  • 00:05:22
    of zurich instance with this type of
  • 00:05:25
    approach which is as i said taking the
  • 00:05:27
    best of the new materials and devices
  • 00:05:30
    new architectures and new theories and
  • 00:05:32
    trying to go really beyond what we have
  • 00:05:35
    today
  • 00:05:37
    so the term neuromorphic was actually
  • 00:05:40
    invented or coined many many years ago
  • 00:05:42
    by carver mead in the late 80s
  • 00:05:44
    and is now being used to describe
  • 00:05:47
    different things there's at least three
  • 00:05:49
    big communities that are using the term
  • 00:05:51
    neuromorphic
  • 00:05:52
    the original one that goes back to
  • 00:05:54
    carver mead was referring to the design
  • 00:05:57
    of cmos electronic circuits that were
  • 00:06:00
    used to emulate the brain
  • 00:06:03
    basically as a basic research attempt to
  • 00:06:05
    try to understand how the brain works by
  • 00:06:07
    building
  • 00:06:08
    circuits that are equivalent so trying
  • 00:06:10
    to really reproduce the physics
  • 00:06:12
    and and because of that these circuits
  • 00:06:14
    were using sub-threshold analog
  • 00:06:16
    transistors
  • 00:06:18
    for the neurodynamics and the
  • 00:06:20
    computation and asynchronous digital
  • 00:06:22
    logic for communicating spikes across
  • 00:06:25
    chips across cores it was really
  • 00:06:27
    fundamental research
  • 00:06:29
    the other big community that now started
  • 00:06:31
    to use the term neuromorphic is the
  • 00:06:33
    community building uh
  • 00:06:35
    practical devices for you know solving
  • 00:06:37
    practical problems
  • 00:06:39
    in that case these these this community
  • 00:06:41
    is is building chips that can implement
  • 00:06:44
    spiking neural network accelerators or
  • 00:06:46
    simulators not emulation but but really
  • 00:06:49
    now at this point it's it's more an
  • 00:06:51
    exploratory approach it's being used to
  • 00:06:53
    try to understand what can be done
  • 00:06:55
    with this approach of using digital
  • 00:06:58
    circuits to simulate spiking neural
  • 00:07:00
    networks
  • 00:07:01
    finally the last community or another
  • 00:07:04
    large community that is started to use
  • 00:07:05
    the term neuromorphic is the one that
  • 00:07:07
    has been developing emerging memory
  • 00:07:09
    technologies looking at nanoscale
  • 00:07:11
    devices to implement long-term
  • 00:07:14
    non-volatile memories
  • 00:07:16
    or if you like memoristive devices
  • 00:07:19
    so this community also started using the
  • 00:07:21
    turn neuromorphic because these devices
  • 00:07:23
    they can actually store
  • 00:07:25
    a change in the conductance which is
  • 00:07:27
    very similar to the way the real
  • 00:07:28
    synapses work when that they actually
  • 00:07:31
    change their conductance when they
  • 00:07:32
    change their synaptic weight
  • 00:07:34
    and
  • 00:07:35
    this allows them to build in memory
  • 00:07:37
    computing architectures that are also as
  • 00:07:40
    you will see very similar to the way
  • 00:07:42
    real biological neural networks work and
  • 00:07:45
    it can really create high density arrays
  • 00:07:47
    so we can actually by using analog
  • 00:07:50
    circuits the approach of simulating
  • 00:07:52
    digital um spike neural networks and by
  • 00:07:55
    using in-memory computing technologies
  • 00:07:57
    the hope is that we create a new field
  • 00:08:00
    which i'm calling here neuromorphic
  • 00:08:01
    intelligence that will lead to the
  • 00:08:04
    creation of
  • 00:08:05
    compact intelligent brain inspired
  • 00:08:08
    devices
  • 00:08:09
    and really to understand how to do these
  • 00:08:11
    brain inspired devices it's important to
  • 00:08:14
    look at the brain to go back to carver
  • 00:08:16
    meats approach and really do fundamental
  • 00:08:18
    research in studying biology and try to
  • 00:08:21
    really get the best out of all all of
  • 00:08:23
    these communities of the devices of the
  • 00:08:25
    sort of the computing principles using
  • 00:08:28
    simulations and and machine learning
  • 00:08:30
    approaches but also of neuroscience and
  • 00:08:33
    studying the brain
  • 00:08:34
    and so here i'd like to just to
  • 00:08:36
    highlight the main differences that are
  • 00:08:37
    there between
  • 00:08:39
    simulated artificial neural networks and
  • 00:08:41
    really
  • 00:08:42
    the biological neural networks those
  • 00:08:44
    that are in the brain
  • 00:08:45
    in simulated artificial neural networks
  • 00:08:48
    as you probably know there is a weighted
  • 00:08:50
    sum of inputs the inputs are all coming
  • 00:08:52
    in a point neuron which is basically
  • 00:08:54
    just doing the sum or the integral of
  • 00:08:56
    the inputs and multiplying all of them
  • 00:08:58
    by a weight so it's it's really
  • 00:09:00
    characterized by a big uh weight
  • 00:09:02
    multiplication or matrix multiplication
  • 00:09:04
    operation and then there is a
  • 00:09:06
    non-linearity either a spiking
  • 00:09:08
    non-linearity if it's a spike neural
  • 00:09:09
    network or a thresholding non-linearity
  • 00:09:12
    if it's an artificial neural network
  • 00:09:14
    in biology the neurons are also
  • 00:09:17
    integrating all of their synaptic inputs
  • 00:09:20
    with different weights so there is this
  • 00:09:22
    analogy of weighted inputs but it's all
  • 00:09:24
    happening through the physics of the
  • 00:09:25
    devices so the the physics is playing an
  • 00:09:28
    important role for computation
  • 00:09:31
    the synapses are not just doing a
  • 00:09:32
    multiplication they're actually
  • 00:09:34
    implementing some temporal operators
  • 00:09:37
    integrating applying non-linearities uh
  • 00:09:40
    dividing summing it's much more
  • 00:09:42
    complicated than just a weighted sum of
  • 00:09:44
    inputs
  • 00:09:45
    in addition the neuron actually has an
  • 00:09:47
    axon and it's sending its output through
  • 00:09:49
    the axon using also basically an all or
  • 00:09:52
    none event a spike
  • 00:09:54
    through time because the longer the axon
  • 00:09:57
    the longer it will take for the for the
  • 00:09:58
    spike to travel and reach the
  • 00:10:00
    destination
  • 00:10:02
    and depending on how thick the axon is
  • 00:10:04
    how much myelination there is it will be
  • 00:10:06
    slower or faster so also here the
  • 00:10:08
    temporal dimension is is really
  • 00:10:10
    important
  • 00:10:11
    in summary if we really want to see the
  • 00:10:13
    big difference is that artificial neural
  • 00:10:14
    networks the one that are being
  • 00:10:16
    simulated on computers and gpus are
  • 00:10:18
    actually algorithms that simulate some
  • 00:10:21
    basic properties of real neurons
  • 00:10:24
    whereas biological neural networks
  • 00:10:25
    really use time dynamics and the physics
  • 00:10:28
    of their computing elements to run the
  • 00:10:31
    algorithm actually in these networks the
  • 00:10:34
    structure of the architecture is the
  • 00:10:36
    algorithm there is no distinction
  • 00:10:38
    between the hardware and the software
  • 00:10:40
    everything is one and understanding how
  • 00:10:43
    to build
  • 00:10:44
    these types of hardware architectures
  • 00:10:47
    wet wear or hardware
  • 00:10:49
    using cmos using memristors maybe even
  • 00:10:51
    using alternative you know dna computing
  • 00:10:54
    or other approaches
  • 00:10:55
    will
  • 00:10:56
    hopefully and probably lead to much more
  • 00:10:58
    efficient and powerful computing systems
  • 00:11:01
    compared to the artificial neural
  • 00:11:03
    networks so if we want to understand how
  • 00:11:05
    to do this we really need to do a
  • 00:11:07
    radical paradigm shift in computing
  • 00:11:10
    standard computing architectures are
  • 00:11:12
    basically based on the phenomenon
  • 00:11:15
    system where you have a cpu on one side
  • 00:11:18
    and
  • 00:11:18
    memory uh on the other and as i said
  • 00:11:22
    transferring data back and forth from
  • 00:11:24
    the cpu to the memory and back is
  • 00:11:27
    actually what's burning all the power
  • 00:11:29
    doing the computation inside the cpu is
  • 00:11:31
    much much uh more energy efficient and
  • 00:11:34
    less costly than transferring the data
  • 00:11:37
    in brains what's happening is that
  • 00:11:39
    inside the neuron there are synapses
  • 00:11:41
    which store
  • 00:11:43
    the value of the weight so
  • 00:11:45
    memory and computation are co-localized
  • 00:11:48
    there is no transfer of data back and
  • 00:11:50
    forth everything is happening at the
  • 00:11:52
    synapse at the neuron and there is many
  • 00:11:54
    distributed synapses as many distributed
  • 00:11:56
    neurons so the memory and the
  • 00:11:58
    computation are intertwined together in
  • 00:12:00
    a distributed system
  • 00:12:02
    and this is really a big difference so
  • 00:12:04
    if we want to understand how to really
  • 00:12:06
    save power we have to look at how the
  • 00:12:08
    brain does it we have to use these brain
  • 00:12:10
    inspire strategies and the main three
  • 00:12:13
    points that i'd like you to remember is
  • 00:12:15
    that you we have to use basically
  • 00:12:17
    parallel arrays of processing elements
  • 00:12:19
    that have computation and memory
  • 00:12:21
    co-localized and this is radically
  • 00:12:23
    different from time multiplexing a
  • 00:12:26
    circuit here for example if we have one
  • 00:12:27
    cpu two cpus but even 64 cpus to
  • 00:12:31
    simulate
  • 00:12:32
    thousands of neurons we are time
  • 00:12:34
    multiplexing the integration of the
  • 00:12:36
    differential equations in these 64 cpus
  • 00:12:40
    here if we look at how to do it
  • 00:12:42
    following this brain inspired strategies
  • 00:12:44
    if we want to emulate a thousand neurons
  • 00:12:46
    we really have to have a thousand
  • 00:12:48
    different circuits that are laid out in
  • 00:12:50
    the in the layout of the of the chip of
  • 00:12:52
    the wafer and then run these
  • 00:12:55
    through their physics through the
  • 00:12:56
    physics of the circuits analog circuits
  • 00:12:59
    digital circuits but they have to be
  • 00:13:01
    many parallel circuits that operate in
  • 00:13:03
    parallel with the memory and the
  • 00:13:05
    computation co-localized that's really
  • 00:13:07
    the trick to to save power
  • 00:13:09
    the other is if we have analog circuits
  • 00:13:11
    we can use the physics of the circuits
  • 00:13:13
    to carry out the computation that really
  • 00:13:14
    instead of abstracting away some
  • 00:13:16
    differential equations and integrating
  • 00:13:18
    numerically the differential equation we
  • 00:13:20
    really use the physics of the device to
  • 00:13:22
    carry out the computation it's much more
  • 00:13:24
    efficient in terms of power latency time
  • 00:13:27
    and area
  • 00:13:28
    and finally the temporal domain is
  • 00:13:31
    really important the temporal dynamics
  • 00:13:32
    of the system have to be well matched to
  • 00:13:34
    the signals that we want to process so
  • 00:13:37
    if we want to have very low power
  • 00:13:38
    systems and for example we want to
  • 00:13:40
    process speech we have to have elements
  • 00:13:42
    in our computing substrate in our brain
  • 00:13:44
    like computer that have the same time
  • 00:13:47
    constants speech for example phonemes
  • 00:13:49
    have time constants of the order of 50
  • 00:13:51
    milliseconds so we have to slow down
  • 00:13:53
    silicon to have dynamics and time
  • 00:13:55
    constants of the order of 50
  • 00:13:57
    milliseconds so our our chips will be
  • 00:14:00
    firing and going at you know hertz or
  • 00:14:03
    maybe hundreds of hertz but definitely
  • 00:14:05
    not that megahertz or gigahertz like our
  • 00:14:07
    cpus or our gpus are doing
  • 00:14:10
    and and by having parallel arrays of
  • 00:14:13
    very slow elements we can still get very
  • 00:14:15
    fast computation even if we have slow
  • 00:14:18
    elements it doesn't mean that we don't
  • 00:14:19
    have a very fast reactive system
  • 00:14:21
    it's because they're in working in
  • 00:14:23
    parallel and so at some point there will
  • 00:14:25
    always be one or two of these elements
  • 00:14:26
    that are about to fire whenever the
  • 00:14:28
    input arrives and we can have
  • 00:14:31
    microsecond nanosecond reaction times
  • 00:14:33
    even though we have millisecond dynamics
  • 00:14:36
    and this is another key trick to
  • 00:14:38
    remember if we want to understand how to
  • 00:14:40
    do this radical paradigm shift
  • 00:14:43
    and at the university of zurich at eth
  • 00:14:45
    at the institute of neuroinformatics
  • 00:14:47
    we've been building these types of
  • 00:14:48
    systems
  • 00:14:49
    for many many years and we are uh now
  • 00:14:53
    also building these systems at our new
  • 00:14:55
    startup at since
  • 00:14:57
    the type of systems are shown here
  • 00:14:59
    basically we create arrays of neurons
  • 00:15:02
    with analog circuits
  • 00:15:04
    these circuits as i told you are slow
  • 00:15:06
    they have slow temporal non-linear
  • 00:15:08
    dynamics
  • 00:15:09
    as i told you they are massively
  • 00:15:11
    parallel we do massively parallel
  • 00:15:13
    operations all of the circuits work in
  • 00:15:15
    parallel
  • 00:15:16
    the fact that they are analog actually
  • 00:15:18
    brings this
  • 00:15:20
    feature that are that is basically
  • 00:15:23
    device mismatch all the circuits are
  • 00:15:25
    inhomogeneous they are not equal and
  • 00:15:28
    this actually can be used as an
  • 00:15:29
    advantage to carry out robust
  • 00:15:31
    computation it's counter-intuitive but i
  • 00:15:33
    will show you that it's actually an
  • 00:15:35
    advantage to have variability in your
  • 00:15:38
    devices and this actually is also very
  • 00:15:40
    nice for people doing memory still
  • 00:15:42
    devices that are typically very very
  • 00:15:44
    variable
  • 00:15:46
    the other features are that they are
  • 00:15:48
    adaptive all of these circuits that we
  • 00:15:50
    have have negative feedback loops they
  • 00:15:51
    have learning
  • 00:15:52
    adaptation plasticity so the learning
  • 00:15:55
    actually helps in creating robust
  • 00:15:58
    computation through the noisy and
  • 00:16:00
    variable elements
  • 00:16:02
    by construction there are many of these
  • 00:16:04
    in working in parallel even if some of
  • 00:16:05
    these stop working the system is fault
  • 00:16:08
    tolerant you don't have to throw away
  • 00:16:09
    the chip like you would with a standard
  • 00:16:11
    processor if one transistor breaks
  • 00:16:14
    probably performance will degrade
  • 00:16:16
    smoothly but at least the system will be
  • 00:16:17
    fault tolerant and because we use both
  • 00:16:20
    the best of both worlds analog circuits
  • 00:16:22
    for the dynamics and digital circuits
  • 00:16:24
    for the communication we can program the
  • 00:16:26
    routing tables and configure these
  • 00:16:28
    networks so we have flexibility in being
  • 00:16:31
    able to program these dynamical systems
  • 00:16:34
    like you would program a neural network
  • 00:16:35
    on a cpu on a computer
  • 00:16:38
    uh of course it's it's more complex we
  • 00:16:40
    still have to develop all the tools and
  • 00:16:42
    and simpsons and and other colleagues
  • 00:16:44
    around the world are still busy
  • 00:16:46
    developing the tools to program these
  • 00:16:47
    dynamical systems
  • 00:16:49
    it's not nearly as well developed as you
  • 00:16:52
    know having a java or a c or a python
  • 00:16:55
    piece of code but
  • 00:16:57
    there is very promising work going on
  • 00:17:00
    and now the question always comes why do
  • 00:17:02
    you do it if analog is noisy and
  • 00:17:04
    annoying in homogeneous why do you go
  • 00:17:07
    through the effort of building these
  • 00:17:08
    analog circuits so let me just try to
  • 00:17:11
    explain that there are several
  • 00:17:12
    advantages if you think of having large
  • 00:17:15
    networks in which you have many elements
  • 00:17:17
    working in parallel for example these
  • 00:17:19
    memorizative devices in a crossbar array
  • 00:17:22
    and you want to send data through them
  • 00:17:24
    these membership devices if you use the
  • 00:17:26
    physics they use analog variables so
  • 00:17:31
    if you just send these variables in an
  • 00:17:33
    asynchronous mode you don't need to use
  • 00:17:35
    a clock so you can avoid using digital
  • 00:17:38
    clock circuitry which is actually
  • 00:17:40
    extremely expensive in terms of area
  • 00:17:42
    requirements in large complex chips and
  • 00:17:45
    extremely power hungry so avoiding
  • 00:17:47
    clocks is is something really really
  • 00:17:49
    useful
  • 00:17:51
    if we don't use digital if we're staying
  • 00:17:52
    analog all the way from the input to the
  • 00:17:54
    output we don't need to convert we don't
  • 00:17:56
    need to convert from digital to analog
  • 00:17:58
    to to run the physics of these
  • 00:18:00
    memristors and we don't need to convert
  • 00:18:02
    back from analog to digital and these
  • 00:18:04
    adcs and these dacs are actually very
  • 00:18:06
    expensive in terms of size and power so
  • 00:18:09
    again if we don't use them we save in
  • 00:18:11
    size and power
  • 00:18:13
    if we use transistors to do for example
  • 00:18:15
    exponentials we don't need to have a
  • 00:18:17
    very complicated uh digital circuitry to
  • 00:18:20
    do that so again we can use a single
  • 00:18:22
    device that through the physics of the
  • 00:18:24
    device can do complex nonlinear
  • 00:18:26
    operation that saves area and power as
  • 00:18:28
    well
  • 00:18:29
    and finally if we have analog variables
  • 00:18:31
    like variable voltage heights variable
  • 00:18:34
    voltage pulses widths
  • 00:18:37
    and and
  • 00:18:38
    other types of variable currents we can
  • 00:18:40
    control the properties of the
  • 00:18:42
    of the devices that we use these
  • 00:18:44
    memorized devices we can make them uh
  • 00:18:48
    depending on how strong we we drive them
  • 00:18:50
    we can make them volatile or
  • 00:18:51
    non-volatile we can use their intrinsic
  • 00:18:54
    non-linearities
  • 00:18:56
    depending on how strongly we derive them
  • 00:18:58
    we can even make them switch with a
  • 00:18:59
    probability so we can use their
  • 00:19:01
    intrinsic stochasticity to do stochastic
  • 00:19:04
    gradient descent or to do probabilistic
  • 00:19:06
    graphical networks to do probabilistic
  • 00:19:08
    computation
  • 00:19:10
    and we can also use them in their
  • 00:19:12
    standard way of operation in their
  • 00:19:15
    non-volatile way of operation as
  • 00:19:17
    long-term memory elements so we don't
  • 00:19:19
    need to shift data back and forth from
  • 00:19:22
    peripheral memory back we can just store
  • 00:19:24
    the value of the sinuses directly in
  • 00:19:27
    these membrane devices
  • 00:19:29
    so if we use analog for our neurons and
  • 00:19:31
    synapses in cmos
  • 00:19:33
    we can then best best benefit the use of
  • 00:19:36
    future emerging memory technologies uh
  • 00:19:39
    reducing power consumption
  • 00:19:41
    and uh in a very recent in the last
  • 00:19:44
    escas conference which was just you know
  • 00:19:45
    a few weeks ago we did show with the pcm
  • 00:19:49
    trace algorithm or or the pcm series
  • 00:19:51
    experiments that we can exploit the
  • 00:19:54
    drift of pcm devices which are these
  • 00:19:56
    shown here in a picture from ibm
  • 00:19:59
    to implement eligibility traces which is
  • 00:20:02
    a very useful feature to have for
  • 00:20:04
    reinforcement learning
  • 00:20:06
    so if if we are interested in building
  • 00:20:08
    reinforcement learning algorithms for
  • 00:20:10
    example for for having behaving robots
  • 00:20:12
    that run with
  • 00:20:14
    brains that are implemented using these
  • 00:20:16
    chips we can actually take advantage of
  • 00:20:18
    the properties of these pcm devices
  • 00:20:21
    none that are typically thought as
  • 00:20:23
    non-idealities we can use them to our
  • 00:20:25
    advantage for computation now
  • 00:20:28
    analog circuits are noisy i told you
  • 00:20:30
    there are the they are variable in
  • 00:20:31
    homogeneous for example if you take one
  • 00:20:33
    of our chips you you stimulate the
  • 00:20:36
    neurons with the same current to all the
  • 00:20:38
    neurons to 256 different neurons and you
  • 00:20:41
    see how long it takes for the neuron to
  • 00:20:42
    fire
  • 00:20:43
    not only these neurons are slow but
  • 00:20:45
    they're also variable the time at which
  • 00:20:47
    they fire can greatly change depending
  • 00:20:49
    on which circuit you're using
  • 00:20:51
    and there is this noise which is usually
  • 00:20:53
    typically you have variance of 20
  • 00:20:55
    over the mean so the coefficient of
  • 00:20:57
    variation is about 20 percent
  • 00:21:00
    so the question is how can you do robust
  • 00:21:01
    computation using this noisy
  • 00:21:04
    computational substrate and the obvious
  • 00:21:06
    answer the easiest thing that people do
  • 00:21:08
    when they have noise is to average
  • 00:21:11
    so we can do that we can average over
  • 00:21:13
    space and we can average over time if we
  • 00:21:16
    use populations of neurons not single
  • 00:21:18
    neurons we can just take you know two
  • 00:21:20
    three four six eight neurons and look at
  • 00:21:22
    the average time that it took for them
  • 00:21:23
    to spike or if they're spiking uh
  • 00:21:26
    periodically we can look at the average
  • 00:21:28
    firing rate
  • 00:21:29
    and then if we integrate over long
  • 00:21:31
    periods of time we can average over time
  • 00:21:33
    so these these two strategies are going
  • 00:21:35
    to be useful for uh reducing the effect
  • 00:21:38
    of device mismatch if we do need to have
  • 00:21:40
    precise computation
  • 00:21:42
    and we are doing experiments in this in
  • 00:21:44
    these very days this is actually a very
  • 00:21:46
    recent experiment where we took these
  • 00:21:48
    neurons and we put two of them together
  • 00:21:51
    four of them together eight sixteen if
  • 00:21:53
    you look at the cluster size basically
  • 00:21:54
    that's the number of neurons that we are
  • 00:21:56
    using for the average over space
  • 00:21:59
    and then we are computing the um
  • 00:22:02
    firing rate over two milliseconds five
  • 00:22:03
    milliseconds 50 milliseconds 100 and so
  • 00:22:06
    on and then what we do is we calculate
  • 00:22:09
    the coefficient of variation basically
  • 00:22:11
    how much device mismatch there is the
  • 00:22:13
    larger the coefficient of variation the
  • 00:22:15
    more noise the smaller the coefficient
  • 00:22:17
    of variation the less noise so we can go
  • 00:22:19
    from a very large coefficient of
  • 00:22:21
    variation of say 12
  • 00:22:23
    actually 18 as i said 20
  • 00:22:26
    by integrating over long periods of time
  • 00:22:29
    or by integrating over large numbers of
  • 00:22:31
    neurons we can decrease this all the way
  • 00:22:32
    to 0.9
  • 00:22:34
    and you can take this coefficient of
  • 00:22:35
    variation and you can calculate the
  • 00:22:37
    equivalent number of bits if we were
  • 00:22:39
    using digital circuits how many would
  • 00:22:42
    bits would this correspond to so for by
  • 00:22:45
    just integrating over larger number of
  • 00:22:47
    neurons and over longer periods we can
  • 00:22:49
    have for example a sweet spot where we
  • 00:22:51
    have eight bit resolution just by using
  • 00:22:54
    16 neurons
  • 00:22:58
    and integrating for example over 50
  • 00:22:59
    milliseconds
  • 00:23:01
    this can be changed at runtime if we if
  • 00:23:03
    we want to have a very fast reaction
  • 00:23:05
    time and a course idea of what the
  • 00:23:07
    result is we can have only two neurons
  • 00:23:10
    and integrate only over two milliseconds
  • 00:23:13
    then there's many false myths when we we
  • 00:23:15
    use spike neural networks people tell us
  • 00:23:17
    oh but if you have to wait until you
  • 00:23:19
    integrate enough it's going to be slow
  • 00:23:22
    if you have to average over time it's
  • 00:23:23
    going to take area all of these are
  • 00:23:26
    actually false myths that can be
  • 00:23:27
    debunked by looking at neuroscience
  • 00:23:29
    neuroscience has been studying how the
  • 00:23:31
    brain works the brain is extremely fast
  • 00:23:34
    it's extremely low power we don't have
  • 00:23:36
    to wait long periods of time to make a
  • 00:23:37
    decision
  • 00:23:39
    so if you use populations of neurons to
  • 00:23:41
    average out it's been shown for example
  • 00:23:43
    experimentally with real learners
  • 00:23:45
    that populations of neurons have
  • 00:23:47
    reaction times that can be even 50 times
  • 00:23:49
    faster than single neurons
  • 00:23:51
    so by using populations we can really
  • 00:23:53
    speed up the computation
  • 00:23:55
    if we use the populations of neurons we
  • 00:23:57
    don't need them to be every neuron to be
  • 00:23:59
    very highly accurate they can be noisy
  • 00:24:01
    and they can be very low precision but
  • 00:24:04
    by
  • 00:24:05
    using populations and using sparse
  • 00:24:08
    coding we can have very accurate
  • 00:24:09
    representation of data and there is a
  • 00:24:12
    lot of work for example by sophie the
  • 00:24:14
    nerve that has been showing how to do
  • 00:24:15
    this by by training populations of
  • 00:24:18
    neurons to do that
  • 00:24:19
    and now it should be also known in
  • 00:24:21
    technology that if you have variability
  • 00:24:24
    it actually helps in transferring
  • 00:24:26
    information across multiple layers
  • 00:24:28
    and here what i'm showing here is data
  • 00:24:30
    from one of our chips where we use 16 of
  • 00:24:32
    these neurons per core and we basically
  • 00:24:35
    provide some in desired target as the
  • 00:24:38
    input we drive a motor with a pid
  • 00:24:41
    controller and we minimize the error
  • 00:24:44
    it's just to show you that by using
  • 00:24:45
    spikes we can actually have very fast
  • 00:24:48
    reaction times in robotic platforms
  • 00:24:50
    using these types of uh chips that you
  • 00:24:52
    saw in the previous slides
  • 00:24:55
    in fact we've been building many chips
  • 00:24:56
    for many years also at since the the the
  • 00:24:59
    colleagues are building chips and the
  • 00:25:00
    latest one that we have built at the
  • 00:25:02
    university as a academic prototype
  • 00:25:06
    is um
  • 00:25:08
    is called the dynamic neuromorphic
  • 00:25:09
    asynchronous processor it was built
  • 00:25:11
    using a very old
  • 00:25:13
    technology 180 nanometer as i said
  • 00:25:15
    because it's an academic exercise but
  • 00:25:18
    still this has a thousand neurons it has
  • 00:25:21
    four cores of 256 neurons each we can
  • 00:25:24
    actually do very interesting edge
  • 00:25:25
    computing applications just with a few
  • 00:25:28
    hundred neurons and then of course then
  • 00:25:30
    the idea is to use both the best of both
  • 00:25:32
    worlds to see where we can do
  • 00:25:34
    analog circuits to really have low power
  • 00:25:37
    or digital circuits to have you know to
  • 00:25:39
    verify principles and make you know
  • 00:25:42
    practical problems solve practical
  • 00:25:44
    problems quickly and then by combining
  • 00:25:46
    analog and digital we can also there get
  • 00:25:49
    the best of both worlds
  • 00:25:50
    so to to conclude actually i would just
  • 00:25:53
    like to show you examples of
  • 00:25:54
    applications that we built
  • 00:25:57
    using this
  • 00:25:59
    this is
  • 00:26:00
    a long list of applications but if you
  • 00:26:02
    have
  • 00:26:03
    the slides you can actually click on the
  • 00:26:07
    on the references and you can get the
  • 00:26:08
    paper the ones highlighted in red are
  • 00:26:11
    the ones that have been done by simpsons
  • 00:26:13
    uh by our colleagues from instance on
  • 00:26:15
    ecg anomaly detection and the detection
  • 00:26:18
    of vibration anomalies was actually done
  • 00:26:19
    by simpsons and by university of zurich
  • 00:26:22
    in parallel independently and just go to
  • 00:26:24
    the last slide where i basically tell
  • 00:26:26
    you that we are now at a point where we
  • 00:26:29
    can actually use all of the knowledge
  • 00:26:31
    from the university about brain inspire
  • 00:26:34
    strategies
  • 00:26:35
    to develop this neuromorphic
  • 00:26:36
    intelligence field transfer all of this
  • 00:26:39
    know-how into
  • 00:26:40
    technology and in the in with new
  • 00:26:43
    startups that can actually use this
  • 00:26:45
    know-how to solve practical problems
  • 00:26:47
    and try to find you know what is the
  • 00:26:48
    best market for this
  • 00:26:50
    as i said industrial monitoring for
  • 00:26:52
    example for vibrations is something but
  • 00:26:54
    something that can also be done using
  • 00:26:56
    both sensors and processors is is really
  • 00:26:59
    what the simpsons company has been
  • 00:27:01
    developing and it's it's been really
  • 00:27:03
    successful at for doing intelligent
  • 00:27:05
    machine vision for putting both uh
  • 00:27:07
    sensing and
  • 00:27:09
    artificial intelligence algorithm on the
  • 00:27:11
    same chip
  • 00:27:12
    and having a very low power in the order
  • 00:27:14
    of microwave uh tens or hundreds of
  • 00:27:16
    microwatt power dissipation for solving
  • 00:27:19
    practical problems that can be useful in
  • 00:27:21
    society and in fact even solving
  • 00:27:23
    consumer applications
  • 00:27:25
    so with this um sorry i went a bit over
  • 00:27:27
    time but i would just like to thank you
  • 00:27:29
    for your attention
  • 00:27:39
    you
Tags
  • AI
  • hjernensinspireret
  • neuromorfik
  • lavenergi
  • computing
  • analoge kredsløb
  • energibesparelse
  • neurale netværk
  • hukommelse
  • specialisering