Podcast - Fullstack - Containers - Apps

March 20th 2013 at Melbourne, Australia

Show notes:

Computaiton and Simulation Sciences Conference
340ppl, 3 Plenary Speakers, 80 conference talks, 20 posters
12 Flagships, hundreds+ themes, ten-thousand+ projects

Capability hub:
  * Strategic team leader: Tim Gureyev (imaging and visuations), Sam Yang (Data Constrained material modelling)

Project leaders:
Tomasz Bednar - cloud based image aanlysis
Aaaron thorton - Aaron Thornton

Canidate material -> screening tools (HPC) -> industry capability material

Upcoming future tech:
  * Milxview (brain analysis MRI?) tool in the Cloud
  * Global virtual labs as a vision for the future, researchers conglomerate around data+tools.
  * GPU clusters (Kepler GPUs).
  * CSIRO Cloud (50Petabyte)
  * Visualising biological data (from Felice Frankel on Fri, MIT)
  * Accelerated Computing Service: optimising HPC to solve computation problems that need speeding up through developers.
  * Science Data Service: accelerated science data project to help get data into DAP.
  * HPC centres (iVec, NCI, VLSCI) - petaflop computing facilities coming online in 2013.
  * Windows Cloud, Desktop and Mobile

CSIRO eResearch Workflow: Research planning -> collaboration -> data collection -> data management -> computation -> discovery -> publication.

"I wish I would have had the DAP when I started my research, if we would have had it I would now be able to find that data which is sitting somewhere on old floppy disks, tapes and decripid laptop hard drives.  Use our eResearch services, you won't regret it."

=======================

Plenary talk - probabalistic and parallel computing
Lawrence Murray, CSIRO, Mathematic, informatics and Stats
Informatics from Edinburgh, MA from ANU
Bayesian inference, monte carlo, hpc, environment applications.

Weather prediction
  * uncertainty in predictions
  * quantitiy of uncertainty, "likely rain" or cyclone prediction
  * pairs of hurricaines, fujiwara affect
  * simple model that follows the tsunami as an object, how the two objects relate to one another and their individual trajectory, they were not able to predict that the hurricaine would turn back on itself.
  * observation uncertainty (where are the hurricaines exactly, e.g. marking the exact middle of the hurricaine?)
  * lots of possible datasets to integrate, e.g. sea surface temperature effects prediciton
  * probability distribution (bayesian): likelihood(prior)=joint
  * bayes rule = applied to product rule of data.
  * essentially setting up the formula so that it can be applied to the variables in the dataset, e.g. velocity, temperature, trajectory, etc.

Soil Carbon
  * carbon decay in the soil
  * decay rate of carbon + prior distributions of carbon + use (crop rotation, prior use, etc)
  * get resulting probability distribution = range of uncertainty

State-space model
  * stateChange(joint = prior(liklihood))
  * joint= parameter x initial x transition x observation
  * a state space model is a structure that works well to do bayesian inference on HPC (sequential monte carlo)
  * LibBi model is software code for modelling state-space model.
  * LibBi will then work on various h/w architecture, GPU, CPU, etc.
  * no longer can we just write a single process it and then parrallise it. 
  * data parrallism: same data items, I can then do muliple operations on that data and multiply, e.g. with GPU x32 operations
  * task parrallism on GPU can be 48-way
    * cores are misleading on GPU vs CPU
  * Clock rate (GhZ) & cache rate (Kb)
  * "Threads per warp"
  * parrallelism was a physical requirement because of power/heat problems
  * bragg cluster (cpu, gpu + shared memory setup)
  * networked systems are slower than in-memory systems
 
Sequential Monte Carlo applied to harware constraints
  * model+data+inference=>uncertainty?
  * given a fixed set of parameters, propogate where they will be in the first state versus second state (observation), resampling based on weight of particle => move to third state (weight & resample)...
    * Xt x Yt => uncertainty (estimate of parameter)
  * this is a good model for parrallising on SuperComputer in a shared memory environment, e.g. GPU
  * e.g. particle chain monte carlo
  * sample from a distribution queue to guess new parameter to walk through state model.
  * metrpolis-hasting rule
  * how does this run on a distrib memory system?
    * 'burn in' period can be reduced by guassian random walk.

Marine biogeochemistry in North Pacific
  * How much nitrogeon they hold (mixing, growth, grazing, respiration, mortality, reminerisation, respiration, sinking)
  * all species lumped into single species (phytoplankton, zooplankton) to do general modelling.
  * 15xParameters / 15xState variables
  * Cooler+lessLight = less phytoplankton
  * credibility of model is subject to spatial and instrument errors

SMC(squared)
  * burn in problem with Monte Carlo
  * introduces parameters along with state.
  * each state: propegate -> weight -> resample -> diversify (random walk)
  * this is very parrallisable, million-way parrallel
  * can't resample per state until each process is finished.
    * the number of steps are dependent on the number of parameter
    * load balancing heavy parameter tasks

Summary:
  * from prior to postering - this is a computation problem
  * algorithm needs to match the structure of the h/w
  * in many disciplines quantifiying uncertainty is the best signifier for moving research forward

Q&A:
  * Future is billion-way parrallelism
    * maximising the capability of that device, mobile or HPC
    * research needs to move on to take advantage of this computation power.
  * How does this work with Big Data?
    * "Fast data" (regardless of how 'big') will work with this model.
    * hierarchical data works with this.
    * Big data will require simple models.
    * complex models with big data is not easy.
  * Trend towards 'uncertainty quanitification' methods is misplaced?
    * William Stark from Berkely criticised this method as subject to noise processes, esp in earth sciences... black swans?
    * big forward physics often applies noise processes which corrodes the outcomes.

===================================================================
Keynote:
Cloud Computing and Big Data Analytics in Science
Dennis Gannon, Director Cloud Research Strategy
Microsoft Research Connectionis

The 4th Paradigm of Science
  * Every researcher is now a 'data science' researcher
  * Need data analytics tools
  * Long tail of data
    * short tail = physics
    * medium tail = genomics
    * long tail = economics, social science, anthropologists, etc.
      * less than a petabyte per year, becoming more quantitative
  * in US every funded research project has to have a 'data management plan'

Genesis of Big Data
  * early days of indexing the Web, e.g. multi-node, MapReduce
  * next steps are semantic modelling, e.g. concept clustering, relevance, etc.
  * Hadoop from Yahoo => aws, cloudera, datastax, dell, emc2, hartonworks, ibm, mapr, microsoft

Microsoft Data Centre Scale
  * Quincy, Des moines, San Antonio, Chicago, Boydton, Dublin, Amsterdam, Japan, Hong Kong, Singapore, 
  * Power consumption is the primary concern in designing Cloud
  * 90 Projects worldwide in collaboration with funding agencies
    * 30 Projects have done interesting things, the learning curve to change over is not insignificant, you have to want to do *cloud* computing, e.g. fault tolerances, inter-process communication, etc.
  * Bio-informaticians are starting to see advantages of Cloud 9Azure projects)
    * Protein folding (Univesity of Washington), Neuroimaging analysis (France INRIA), Comparative Genomics (U of North Carolina), Drug Discovery (Newcastle Uni, UK), Systems BIology (U of Trento)
  * bio-informatics didn't come from an HPC background to embracing the new technology of Cloud.
  * image analysis is a good job for the Cloud (astronomy, medical, physics, geosciences, crop disease and seismic analysis from satelites, etc). 
  * scientific document analysis is good job for the Cloud: recognising inconsistent scientific results, plagiarism, knowing what was wrong in history based on new findings, auto-ontology learning for greater ease of cross-discplinary understanding, predicting future evolution of science, etc.

Microsoft Cloud & Projects
  * Python as first class language in Azure: NumPy, SciPy, ScikitLearn, biopython, etc.
  * deploy data collection / volumes on cloud, e.g. genomics libraries, medical image libraries, geophysics, astronmy, etc.
  * Apps deployed on Cloud: WorldWide Telescope, Layerscape, Chronozoom, Fetch Climate.

Challenges of long tail of science
  * Scientists don't want to learn the cloud, they use standard packages tools for their work, e.g. spreadsheets, existing scripts that have been utilised, etc.
  * Researchers who use spreadsheets: DataUp by CDL
  * How do we finanace the data from the long tail?
  * Dataverse Network Project
  * How do we find a way to save this data with limited resources?
    * subscription model?
  * value in data?
    * well curated = great
    * w/ high quality services = greater
    * w/ experts = greatest

Q&A:
  * sharring data is at odds with the current academic culture, e.g. putting data up on the gene omnibus was a lot of effort and the researcher got nothing out of it.
  * RDA Alliance making progress on identifiers.
  * Research gate?

======================================================

= Visualisations Track =
Tuesday from 2pm to 3.30pm

== Biomedical Imaging at Microsoft Research ==
Dr Simon Mercer, Director of Health and Wellbeing

about Microsoft Research Connections
  * 850 Researchers, worldwide, labs
  * theoretical computer science turned into products.
  * the future of products
  * applied CS: natural user interface, earth/energy/environment, educationand scholarly communication, health and wellbeing

Image analysis
  * they do projects, e.g. Inner Eye: google "microsoft medicalimaginganalysis"
  * being able to anonymise medical images and make them more searchable.
  * tools to work on medical imaging
  * they want to do this in the open as it is more valuable to address the entire academic community.
    * this enables miscrosoft research to pull in better talent and move through to productisation

AIDA (Anonymized Image Data Archive): ethical and legal framework for anonymous image donations by patients receiving radiotherapy in NHS hospitals
    * hospitals are not concerned with opening up data, it is the hospitals who are concerned with patient privacy.
    * embagro meidcal images with the STFC 
  * AIDA has received ethical approval.

GeoS segmentation tool: 
  * able to swipe accross object in segementation to identify object
  * decision forest, same as in kinect
  * available for download soon
  * works on MRI, CT 3D and 3D scans, etc.

CodaLab
  * platform for data sharring to execute data algorithm competitions.
  * developed at Stanford w Percy Liang with OuterCurve Foundation
  * Python, Apache, GitHub

CrowdSynth: GalaxyZoo like project.
  * how to seperate machine capabilities and human capabilities to work together.

Google search keywords: medimaging microsoft
========================================================
Coputational Biomedical Imaging Research at Monash University
by Gary Egan

Neuro imaging: hard not to show images which are so compelling to the human brain to look at - afterall what brain doesn't want to look at itself?

Monash Facilities
  * Scanners: MRI Siemens Skyra 3T, CT small animal, Bioscan FLECT (3D florescence) scanner, Bioscan PET CT located at Baker IDI
  * Partners: Syntrotron, Southern Health, Alfred Hospital, etc.
  * Florey, Monash, Swinburne, Melbourne
  * Mental Health Research Institution
  * national imaging facility and capability: ansto, sync, ammrf, national imaging facility, VPTN, VBIC
  * overseas repo washington st louis = XNAT
  * DaRIS medical imaging repository
  * All part of MASSIVE capability.

Huntingtons disease is able to be detected 18 months in advance of the symptoms with MRI scanns of 

Europe has 1.2billion euro project to look at neuroscience and computing. = The Human Brain Project (HBP)
  * 2015 meeting to discuss invovlement with HPB

=======================================================
NeCTAR Project RT035: Cloud based image analysis and processing toolbox
  * Moving s/w to the Cloud: HCA vision, x-tract, MILXView

HCA-Vision
  * microscopy imaging analysis: detect nuclei, detect cells, etc.

X-Tract: x-ray image analysis
  * sinogram creation, etc

MILXView
  * 3D medical imaging analysis and visualisation platform
  * cortical surface extraction, etc.

Built workflows for access the various image recognition features from the above s/w
  * GLUE = Galaxy workflow tool
  * yahoo pipes like workflow to set up image analysis
  * selected from existing workflow tools: workspace, galaxy, kepler, yabi, yawl, taverna, andruil, etc.
  * Utilising hadoop, pig, hive, oozie, etc.

Users: neuroscience research community, cancer research, other

cloudimaging.blogspot.com.au

========================================================
X-ray imaging program at SSRF (Shanghai Synchrotron Radiation Facility), Shanghai Institute of Applied Physics, CAS.
by Xiao Ti-Qiao, Prof

Intro to Facility use:
  * 7 beamlines
  * X-ray imaging and tomography (no flux, steady beam)
  * Users from every provence: 131 university, 94 institute, 18 hospital, 24 company.  49% university use, 35% institute use, 9% company use.
  * 732 journal papers published, 157 in top journals, CNS papers 17
  * 6 new beamlines by end of this year = total 14
  * in five years up to 40 beamlines opened up.
 
Examples:
  * First live grasshopper imaged
  * Dentritical crystal growing process
  * Ion transportation in lithium cell via xray
  * high flux density results in better singal to noise ration = better iamges.
  * 12x orders highers than conventional x-ray
  * x-ray is most popular with 618 users in 2012
  * 47% biomedicine, 30% materials, 9% environment

Methodology for X-Ray imaging
  * x-ray tomography
  * x-ray florescence tomography
  * dynamic imaging for material science
  * Ring artefacts 
  * microtompography (drugs, paleontology)
  * nanoCT

https://ct.ssrf...
========================================================
Image segmentation methods for intracranial aneurysm Haemodynamic research, Australian School of Advanced Medicine, Macquarie University
...by Yi Qian (Itsu Sen)

Ways of detecting have improved:
  * 3D angiography
  * flat panel head scanner
  * flow analysis of aneurism
  * 


=======================================================
Tri-tracking: free software for large scale particle tracking
by Pascal Vallotton

Object tracking 
  * using graph object tracking (nodes, single direction graphs)
  * they like veolcity constancy to guess how object is moving
  * large particile jumps
  * 

========================================================
Notes for audio report:

Computaiton and Simulation Sciences Conference
340ppl, 3 Plenary Speakers, 80 conference talks, 20 posters
12 Flagships, hundreds+ themes, ten-thousand+ projects

Keynotes: (long, medium-genomics, short-physics tails)
  * being able to do science with computers is changing (again)
    * HPC modelling of data to do 'uncertainty analysis', e.g. weather analysis predicting where hurricaines will go requires a lot of different predictions.
  * Cloud computing as a new paradigm for different subject areas
    * it is the 'medium tale' (genomics) that is picking up Cloud, instead of HPC (the short tail -physics- is still most using HPC), 
  * Big Data is different depending on if it is fast data, complex data, distrib data, etc. (not just map-reduce)
 
Visualisation - attended this session as visualisin data is the future if you want to get your research read - naturally leads to wanting the data as well.
  * Microsoft is delving into data visualisations as turning data into something people can see is becoming essential in science.
  * There is a lot of science datasets which are images, we need more tools interpretting and working with thes image datasets.
  * NeCTAR has a virtual lab that has extracted several image analysis tools for various disciplines that can be coupled together in a workflow to analyse images, e.g. detect edges, lay out slices of 3D image with certain parts highlighted, etc.
  * In brain science the images of the brain are beautiful and attract the human eye (what brain doesn't like looking at itself).  The trend for more impressive images to express the data is continuing (and getting the money).
  * China's syncrotron will go from 7 beamlines to 40 beamlines in the next five years (this in comparison to the 9 beamlines in Australia, which are struggling for funding). 
  * China's presentation of x-ray images coming out of the their Shangahi snyc is further demonstration of the power of images bridging the language gap.
  * Once complicated tracking of complex moving objects is becoming more simple as algroythims are able to track multiple object based on guess algorithm for which way the objects is going to move. Imagine a bacteria on a petry dish under a microscope and watching as the different parts of that bacteria grew to consume the entire dish, tracking how all those objects of the bacteria move is a powerful tool for understanding the patterns of nature.  Could be applied to particles, food as it goes throughout your stomach, etc.


dfflanders