MUSIC Mikael Djurfeldt, PDC/KTH HBP Neuromorphic SP Outline - - PowerPoint PPT Presentation

▶
music
SMART_READER_LITE
LIVE PREVIEW

MUSIC Mikael Djurfeldt, PDC/KTH HBP Neuromorphic SP Outline - - PowerPoint PPT Presentation

Nov 20, 2023 486 likes •788 views

MUSIC Mikael Djurfeldt, PDC/KTH HBP Neuromorphic SP Outline Interfaces in computatjonal neuroscience sofuware What is MUSIC? Two problems solved by MUSIC: spatjal aliasing problem temporal aliasing problem How to use MUSIC

slide-1
SLIDE 1

MUSIC

Mikael Djurfeldt, PDC/KTH HBP Neuromorphic SP

slide-2
SLIDE 2

Outline

  • Interfaces in computatjonal neuroscience sofuware
  • What is MUSIC?
  • Two problems solved by MUSIC:
  • spatjal aliasing problem
  • temporal aliasing problem
  • How to use MUSIC from C++, Python and PyNN
  • Use cases
  • Where to get sofuware and documentatjon
slide-3
SLIDE 3

Interfaces in computational neuroscience

  • Simulatjon environments in computatjonal neuroscience,

such as NEURON, NEST or Brian, each provide many tools needed by the user to carry out high-quality simulatjon studies.

  • Models described difgerently,

environments have specifjc features => hard to move models

  • Diffjcult to build larger simulatjons which re-use existjng models
slide-4
SLIDE 4

Interfaces in computational neuroscience

  • As systems grow and encompass more subsystems, they rapidly

become unwieldy to develop

  • In general, sofuware in computatjonal neuroscience tends to have

a monolithic structure

  • Sofuware interfaces (APIs) allow for use of difgerent implementatjons
  • f sofuware components
  • MUSIC is an API, and implementatjon in the form of a C++ library,

supportjng fmowing of data between tools during simulatjon (INCF initjatjve, originally developed by Ö. Ekeberg and M. Djurfeldt)

slide-5
SLIDE 5

MUSIC co-simulations

A co-simulatjon with multjple parallel applicatjons (A, B, C) exchanging runtjme data (such as neuronal events)

Shipping data around between applicatjons during simulatjon useful e.g. for:

  • Building larger models by combining models as components
  • Modeling multjple scales and/or combining difgerent

formalisms simultaneously

  • Pre/postprocessing and visualizatjon
  • Interfacing to external hardware
slide-6
SLIDE 6

Network simulation

slide-7
SLIDE 7

Using MUSIC to expose data

slide-8
SLIDE 8

Co-simulation

slide-9
SLIDE 9

Loop

slide-10
SLIDE 10

Spatial aliasing problem

slide-11
SLIDE 11

Spatial aliasing

slide-12
SLIDE 12

Scheduling of communication

Handling of tjme in MUSIC

  • An applicatjon calls MUSIC tick() at points regularly spaced in

simulated tjme

  • This is where data may be sent and/or received
  • Difgerent applicatjons are allowed to call tick()at difgerent rates
  • MUSIC may allow applicatjons to run out-of-sync (each with its own
  • fgset between simulatjon tjme and wallclock tjme)
  • MUSIC allows complex topology of port connectjvity

Scheduling problem

  • How to deliver data in tjme while avoiding deadlocks
  • How to interpolate contjnuous data given difgerent tjck rates
slide-13
SLIDE 13

Interfaces to MUSIC

slide-14
SLIDE 14

C++ app: eventsource

slide-15
SLIDE 15

C++ app: eventsource

slide-16
SLIDE 16

C++ app: eventlogger

slide-17
SLIDE 17

MUSIC configuration file

slide-18
SLIDE 18

Interfaces to MUSIC

slide-19
SLIDE 19

Python app: eventsource

slide-20
SLIDE 20

Interfaces to MUSIC

slide-21
SLIDE 21
slide-22
SLIDE 22

Usage scenarios

slide-23
SLIDE 23

Integrated simulation of the whole-body musculo- skeletal-nervous system for clarifjcation of motor dysfunctions due to neurological diseases

Jun Igarashi¹, Jan Moren¹, Osamu Shouno³, Kazuya Shimizu², Naoto Yamamura², Junichiro Yoshimoto¹ Shu Takagi² & Kenji Doya¹ 1: Okiniwa Insitutiute of Science and Technology (OIST)

2: Tokyo University 3: Honda research Institute Japan

slide-24
SLIDE 24
  • Full size, neuron count of rat BG and motor
  • cortex. (~3.2 million neurons)
  • Conductance-based IaF or Izhikevich-type

neuron models, static or STDP synapses.

  • PyNEST and SLI models, connected with

MUSIC

  • dDIMS: Volumetric FEM-based fluid-

mechanical muscle model, full skeletal physics model

slide-25
SLIDE 25

MUSIC organization

Sample patches/columns on each surface:

  • L5BCS → Striatum: 50×50 patches
  • L5BPT → Spinal cord: 20x20 patches
  • Gpi/EP → Thalamus TC/HT: 1 neuron per

channel (3100 total)

  • spinal cord → muscle: 1 motor neuron per

channel (~750/biceps, 1500/triceps)

slide-26
SLIDE 26
  • Up: GPe, STN and GPi neurons
  • scillating at about 14.7Hz.
  • Left: power spectrum of GPi,

Thalamic CT neurons and L5B PT neurons.

preliminary results

slide-27
SLIDE 27

Other highlighted use cases

  • Bluebrain Monsteer

Library for interactjve visualizatjon

  • MUSIC-ROS toolchain

Philipp Weidel Thursday 10:10

slide-28
SLIDE 28

Where to get MUSIC

  • Github INCF/MUSIC
  • MUSIC manual in the distributjon
  • Djurfeldt et al. (2010) “Run-tjme interoperability between

neuronal network simulators based on the MUSIC framework” Neuroinform.

slide-29
SLIDE 29

Thanks

  • Ekaterina Brocke, SciLife lab, KI – communicatjon algorithms
  • Alexander Peyser, Simlab neurosci, FZJ – Python interface
  • Andrew Davison, Jochen Eppler and Eilif Muller – PyNN

interface

  • Rajalekshmi Deepu, Simlab neurosci, FZJ – Travis integratjon
  • Jan Morén, OIST – MUSIC applicatjon example
  • INCF
  • HBP
  • Simlab neuroscience
  • INM6, FZ Juelich