FROM BRAIN RESEARCH TO FUTURE TECHNOLOGIES. Dirk Pleiter Post-H2020 Vision for HPC Workshop, Frankfurt

Size: px

Start display at page:

Download "FROM BRAIN RESEARCH TO FUTURE TECHNOLOGIES. Dirk Pleiter Post-H2020 Vision for HPC Workshop, Frankfurt"

Abigayle Long
5 years ago
Views:

1 FROM BRAIN RESEARCH TO FUTURE TECHNOLOGIES Dirk Pleiter Post-H2020 Vision for HPC Workshop, Frankfurt

Science Challenge and Benefits Whole brain cm scale Understanding the human brain Understand the organisation of the brain Understand brain functioning and dysfunction Social and economic impact

2 Science Challenge and Benefits Whole brain cm scale Understanding the human brain Understand the organisation of the brain Understand brain functioning and dysfunction Social and economic impact Neuron μm to mm scale Enable treatment of brain diseases Enable new approaches to artificial intelligence Complexity of the brain About 90 billion nerve cells, each with 10,000 connections Multi-scale challenge Need for interdisciplinary approaches Neuroscience, medicine, physics etc. Computer science and engineering, mathematics etc. 2 Synapse nm to μm scale

The Human Brain Project Flagship project funded by the EC Science-driven, seeded from Future and Emerging Technologies, extending beyond ICT Started in 2013 as

research/brain-inspired sciences Gather, organise and disseminate data describing the brain and its diseases Simulate the brain Build multi-scale scaffold

3 The Human Brain Project Flagship project funded by the EC Science-driven, seeded from Future and Emerging Technologies, extending beyond ICT Started in 2013 as a 10 years efforts More than 110 partners in 19 countries Objectives of the HBP Realise an European scientific Research Infrastructure for brain research/brain-inspired sciences Gather, organise and disseminate data describing the brain and its diseases Simulate the brain Build multi-scale scaffold theory and models for the brain Develop brain-inspired computing, data analytics and robotics Ensure that the HBP's work is undertaken responsibly and that it benefits society 3

4 A Case for Co-Design Need for of a cutting-edge ICT-based scientific research infrastructure Simulate the brain Gather, organise and disseminate data Brain research seed future and emerging technology Develop brain-inspired computing, data analytics and robotics 4

5 ENABLING BRAIN RESEARCH THROUGH NEW ARCHITECTURES AND TECHNOLOGIES

brain based on histological input Approach

section images (1μm resolution) Re-construct

Reconstruct nerve fibers in histological

6 High-Resolution Brain Models from Histological Data [K. Amunts et al., Science 2013] Research goal Accurate, highly detailed computer model of the human brain based on histological input Approach Create high-resolution 2-dimensional brain section images (1μm resolution) Re-construct 3-dimensional models from these images Reconstruct nerve fibers in histological brain sections Automatised analysis of high-resolution images [M. Axer et al.] 6

7 Data Analytics on Extreme-Scale Data Volumes Data challenge [K. Amunts, T. Dickscheid et al.] Data acquisition using 8-9 microscopes at ~15 MByte/s ~10 TByte/day Archival data increase 1-2 PByte per year Data analytics challenge Data access speed challengee Example of multi-gpu training requirements (using P100 GPUs) N B IO > GPU ϵ 0.3 GByte / s 7 Site local storage Permanent HPC storage Deep learning methods for automatised cell segmentation Massively parallel feature extraction (e.g. cell count) Image registration Δ t IO < ϵ Δ t comp Data Acquisition Fast HPC storage Need for improved data transport capabilities and support for high performance data analytics workflows HPC/HPDA system

8 Dynamical Numerical Brain Models Research goal Create biologically realistic models of the brain Realistic size of models Realistic level of detail Multi-scale challenge Multi-scale simulations based on co-simulation of different models Supercomputer requirements Maximize memory footprint Application today is memory capacity limited Optimize processing performance memory performance Keep ratio simulation versus simulated time small Allow for interactive steering of the applications Requires interactive visualisation as well as low-latency control loop 8 Need for scalable compute resources (efficiently exploitable for applications)

9 Brain Simulation Workflows HPC system Interactive viz Knowledge Graph Fast HPC storage Jupyter notebook Permanent HPC storage Need for interactive compute services integrated in HPC centres 9

10 Coupling HPC Systems and External Facilities Closed-loop brain and robot model based experiments Brain simulation coupled to robot simulation engines Need for bridges to external services 10

11 Integration and Creation of Knowledge New experiments Enable ICT platforms integrating different discovery steps Experiment Creation of data sets curated following agreed procedures Improve on tapping all possible data sources (including unstructured data sources Creation of knowledge spaces New models Data Experiment Model building and simulation Data integration and strategic data collection Creation of knowledge spaces Knowledge and Theory Other data sources Simulation Need for technologies tapping all relevant data sources Federation of services Need for integration of various compute and data services No centralisation be open and allow for extensions Allow for integration of very heterogeneous compute services or data sources Need for infrastructure federation 11

12 Medical Informatics: Coping with Sensitive Data Objective Enable access to distributed harmonised pan-omics patient data for clinical and fundamental brain research E.g., electronic health records (EHR), picture archiving and communication systems (PACS) Challenge Use of personal data Strategies Securing data repositories for personal data Introduce de-identification mechanisms and services Pseudonomisation Anonymisation Need for technical means supporting - Securing data storage and access - Secure provisioning of de-identification services 12

Service-Oriented Provisioning of HPC/HPDA Resources Provisioning of broad range of integrated infrastructure services Different types of compute services including

13 Service-Oriented Provisioning of HPC/HPDA Resources Provisioning of broad range of integrated infrastructure services Different types of compute services including Scalable compute services Interactive compute services VM services Different types of data stores Develop and deploy services that facilitate federation Federation of AAI, monitoring etc. Encourage communities to build community specific platforms Push for further separation of concerns Close engagement with communities Need for support of new HPC use models including - Flexible provisioning of hardware resources - Supporting federation of infrastructure - Support for new compute devices (NC, QC) 13

14 FROM BRAIN RESEARCH TO NEW ARCHITECTURES AND TECHNOLOGIES

15 Learning from the Brain: Extremely High Connectivity Large Rent s exponent Rent s rule: Scaling of connections as function of logic element C=k N p Typical value for micro-processors: p Smaller values of p outside of micro-procesor Estimated value for mammalian brains: p 0.8 [Landman and Russo, 1971;M.Y. Lanzerotti et al., 2005; D.S. Bassett et al., 2010 ] Brain as inspiration for improving on new approaches to improve data transport capabilities and space filling computers [P. Ruch et al., 2011] 15

16 Learning from the Brain: Extremely Small Power Density Smaller frequency and power density Brain based on massively-parallel distributed architecture Micro-processors based on von-neumann architecture are much sequential and centralised Analogue logics Analogue logics can be much faster for certain operations, e.g. time integration for neural cell membrane Brain as inspiration for neuromorphic processing architectures and massively-parallel digitalanalogue circuitries for compute [Merolla et al., Science, 2014] 16

17 HBP s Neuromorphic Systems Manchester Heidelberg [S. Furber et al.] [K. Meier et al.] 17

New Approaches to Artificial Intelligence Today: Deep learning based on algorithms and implementations that are different from those used by our brains Currently used ANN+algorithms quite successful

18 New Approaches to Artificial Intelligence Today: Deep learning based on algorithms and implementations that are different from those used by our brains Currently used ANN+algorithms quite successful but very compute intensive Principles of learning in the brain seem to be different [W. Maass, 2017, /j.cobeha ] Neural circuits are highly recurrent networks consisting of different types of neurons and synapses with diverse dynamic properties Networks of neurons in the brain provide stable computational function in spite of ongoing rewiring and network perturbations Research starts leading to new learning algorithms Algorithms particular suitable for neuromorphic hardware Brain as inspiration for new approaches to artificial intelligence and possibly new semiconductor innovations [Nature Editorial, 2/2018] 18

19 SUMMARY AND CONCLUSIONS

20 Summary and Conclusions Science drivers are key for defining future HPC system/infrastructure architecture and technology roadmaps Brain research is only one research area Others include, e.g., earth system modelling (including weather) Brain research as driver for architectures and technologies Exascale computing and beyond Creation of federated, ICT-based research infrastructures integrating various compute and storage resources connected to various data sources and coupled to external experiments Convergence of HPC and big data and cloud Brain research as enabler for new technologies Neuromorphic computing New approaches to artificial intelligence 20

Publishable Summary for the Periodic Report Ramp-Up Phase (M1-12)

Publishable Summary for the Periodic Report Ramp-Up Phase (M1-12) Overview. As described in greater detail below, the HBP achieved all its main objectives for the first reporting period, achieving a high