High-Resolution Simulation of
Biological Neural Networks
Simulations of Cortical Slow Waves and
Transition toward Awake States and of the
interplay between sleep and memory
consolidation and memory homeostasis in the
HBP (- until 2023)
and
Prototyping of Brain-Inspired Multi-area
Associative Recurrent Deep-Learning
and
Architecture Benchmark for ExaNest project and
follow-up EuroExa
project (2018-2020)
Pier Stanislao Paolucci
for the APE parallel/distributed computing lab - INFN Roma
Application used by WaveScalES team (INFN, IDIBAPS, UNIMI, IBEC, ISS)
Benchmarks of EXANEST project (and of follow-up project)
Computational Needs / Year
• High-Resolution Simulation of
Biological Neural Networks /
Associative Recurrent Deep-Learning
• 2018, 2019, 2020
• 4.0, 6.0, 8.0 Million core hours / year
• (ref. 2016 Intel-cores @2.4 GHz)
• Typical conf for production runs:
• 1024 - 8192 cores
• 0.6 – 4.8 TB memory
3-4 November 2016
NEST User Workshop 2016
2
INFN Status on
High. Res. Sim. Of Neural Nets
Simulated synaptic events per second
Simulation speed
(simulated synaptic events per second)
depending on number of processors
1.1G Syn,
0.7M Neu
@3.1Hz,
24x24 grid
1.00E+10
1.00E+09
4.4G Syn,
2.8M Neu
@3.1Hz,
48x48 grid
1.00E+08
17.6G Syn,
12M Neu
@3.1Hz,
96x96 grid
1.00E+07
Ideal
1.00E+06
1
8
64
Number of hardware cores
512
On proprietary DPSNN (Distributed Plastic
Spiking Net Simulator) to be ported to
3-4 November 2016
NEST User Workshop 2016
Efficient simulation of tens of
billions of synapses,
projected by grids of
columns of point spiking
neurons, distributed on
thousands of hardware cores
and software processes.
Study of hardware and
software technologies for
neural simulations
The APE
parallel
computing lab.
of INFN
HBP research
infrastructure
3
Backup material
WaveScalES in the HBP
Towards a multi-scale perturbational atlas
of the cerebral cortex. Linking the macro
(TMS/EEG) to the microscale (cortical
slices) through simultaneous recording of
hd-EEG and Stereo-EEG responses to
intracortical stimulation
Photostimulation and
photoinhibition of
slow-wave activity by
light-regulated ligands
of neuronal receptors
Multiscale theory/model of cortical
dynamics of slow-waves. Matching
theory and simulations with
experiments
Parallel simulations
Maria Victoria
Sanchez-Vives
3-4 November 2016
Marcello
Massimini
Pau
Gorostiza
Maurizio
Mattia
NEST User Workshop 2016
Pier Stanislao
Paolucci
5
Experimental components
Perturb and measure at several
spatio-temporal scales using
multiple methodologies
Novel photostimulation and
photoinibition tools based on
small molecules
3-4 November 2016
NEST User Workshop 2016
6
Matching Experiments and Theory
through High-resolution Simulations
Experiments
Meso-Macroscale
Micro-Mesoscale
Theory and simulations
3-4 November 2016
NEST User Workshop 2016
7
High-resolution Fast Cortical Simulations
Parallelized on Thousands of Processors
WaveScalES models of cortical slow-waves and awakening: single and multiple
area, tens of billions of synapses, projected by grids of layered columns, point
spiking neurons, distributed on thousands of hardware cores, STDP and LTP/LTD
Cooperation between SP3 and SP7
(exchange of requisites, specific
developments, reciprocal support)
to port WaveScalES models from DPSNN
SP3 – WP2 DPSNN sim. engine (INFN/ISS):
• fast, scalable, no user-interface, model
specific optimizations
• hardware/software co-design tool and
application benchmark
The APE parallel comp. lab.
of Italian INFN (Istituto
Nazionale di Fisica Nucleare
3-4 November 2016
research
community
SP7 – NEST Platform:
• configurable, user friendly, large
user community, multiple
execution platforms, multiple
models, scalable
In cooperation
with
NEST User Workshop 2016
8
Collaboration with NEST
Experiments
on mouse
on human
NEST
Requirements
Management
Perturbative
Models
Plasticity
Models
Theory
DPSNN
Multi-scale
software model
of cortical SW
and transition to
consciousness
states
NEST simulator
Multiarea
Singlearea nonlaminar
Singlearea
laminar
3-4 November 2016
Neuromodulated
Plastic
NEST User Workshop 2016
Stimulation-Recording
Connectivity Tools
NEST Support for
Modellers
9
Three simulation models for cortical
SW and transition to AW states
• Single-area non-laminar model
• Two-dimensional grid of cortical modules of point-like neurons (LIF model
variant with spike frequency adaptation / Izhikevich model) interconnected with
probability laws depending on distance and source-target population kinds
• Single-area laminar model
• Two-dimensional grid of cortical modules with laminar structure, accounting for
the different properties of the cortical layers (neural dynamics, connections,
synaptic weights, etc.)
• Neuromodulated Plasticity model
• Model incorporating neuromodulated STDP, LTP/LTD, to investigate the role of
deep sleep in memory consolidation and synaptic homeostasis
• Multi-area model
• Single cortical areas connected using low-complexity (effective/resting states)
connectivity matrixes
3-4 November 2016
NEST User Workshop 2016
10
Conclusion
• WaveScalES: models of deep-sleep and transition to
awakeness
•
•
•
•
Single-area non-laminar
Single-area laminar
Multi-area
Neuromodulated STDP, LTP/LTD
• Tight collaboration with NEST and NESTML teams for
requirements fulfilment
• New models for neurons, probability kernels functions,
stimulus, observables, etc.
• WaveScalES models initially prototyped using the
proprietary DPSNN engine, eventually ported to NEST
and offered to HBP community
3-4 November 2016
NEST User Workshop 2016
11
WaveScalES: Long Term Impact
Societal long-term impact
Better management of:
• sleep disorders
• anaesthesia
• memory disorders
• degenerative disorders
• epilepsy
Technological impact areas
• photostimulation tools
• HBP data-bases of measures
• brain simulation software and
hardware interconnect on large
scale parallel computers
Scientific objective
- Better understanding of the cortical organization starting from a
detailed comprehension of its lowest complexity state and of the
initial phases of transition to higher states of consciousness
- Brain inspired Software and Hardware innovations of simulation
systems
3-4 November 2016
NEST User Workshop 2016
12