intelligent imaging for cta · 2018. 11. 19. · •particle showers and propagation through pair...
TRANSCRIPT
Intelligent Imaging for CTA
* Supported by the EU FP7-PEOPLE-2012-ITN project nr. 317446, INFIERI, “Intelligent Fast Interconnected and Efficient Devices for Frontier Exploitation in Research and Industry“. ESR9
Overview
• CTA Imaging Techniques
• Gamma Hadron Separation
• Classic Approaches• Hillas parameters, pulse timing,
• Neural Network design• Convolutional verses multi-layer perceptron
• Next Steps
• Summary
Atmospheric Imaging
• As gamma rays enter the atmosphere, they can interact with air particles • particle showers and propagation through
pair production, bremsstrahlung, and multiple scattering
• Originate from variety of astrophysical sources- AGN jets, gamma ray bursts, pulsar wind nebulae…
H.E.S.S. array of 4 telescopes
Paranal, Chile location with 3 sizes:
o 4 large (LST) in array centre for energies in 10s of GeV
o 25 medium (MST) for 100 GeV – 10 TeVo 70 small (SST) for energies > 10 TeV
Triggering• The camera for the SST needs to
trigger at approximately 1 kHz
• Night Sky Background: enters camera at a rate of ~ 100 Mhz
Considerations:• Can’t record continuously: rolling buffer of 16 μs• Night Sky Background: should only trigger on true events• Other particles also create air showers
• CHEC: Camera with 2048 pixels
• 512 super pixels (4-pixel summation) used for triggering
Gamma-HadronSeparation
• Cosmic Rays (protons and heavier nuclei) are ~104 as a frequent as gamma rays huge source of noise!• About 87% of these
are protons; the rest are heavier nuclei
• More efficiently separating out the gamma will mean a huge increase in the sensitivity of the array
Shower Characteristics• Hadron progenitors undergo inelastic collisions with
atmosphere, developing substantial transverse momentum
• Amount of Cherenkov light scales approximately linearly with energy of progenitor
• Asymmetry of hadron-originating air showers cut out about 70% of events
• Showers originate at ~20km
• Penetrate to about 4km, covering an area of approximately
gamma
proton
Approaches: Hillas Parameters
• Uses approximately elliptical shape of gamma ray showers to reduce complicated image to 6 parameters
• Second order moments describe rate of change in a shape’s area
Mo
han
tyet. al. 1
99
8
Gamma ray shower images are more compact (shorter length and width), uniform, and have principal axes pointing towards the shower centre
M. Naurois, 2015
Centre of intensity
𝑥 =Σ𝑛𝑖𝑥𝑖
Σ𝑛𝑖𝑦 =
Σ𝑛𝑖𝑦𝑖
Σ𝑛𝑖
Image moments:
𝑀𝜇𝜈 =1
𝑁Σ𝑖 𝑥𝑖𝜇𝑦𝑖𝜇𝑛𝑖
Making Tailcuts• The images need to be ”cleaned” so that
accurate calculation of the Hillas parameters can be performed
• Done through nearest neighbour cuts• a pixel is saved either if it is above the RMS intensity
or if the sum of neighbouring pixels is above a cutoff, which is a multiple of the RMS intensity
However: run the risk of losing important information!
Uncalibrated commission data from Paris
Is this information useful?
Simulations of gamma air showers- cleaning and tailcuts applied
Hillas Separation • Hillas parameters are adept at classifying low energy images but fail in the realm above ~10TeV
• Discerning characteristics of length and width increase for protons• If first shower interaction is deeper in the atmosphere, more Cherenkov light is produced because EM cascade develops in
area of larger n- more common for high energy protons
• My initial investigation of Hillas parameters for CHEC returned comparable results to previous IACT telescopes.
• Length, width and distance are some of the most discriminating parameters: individual quality factors between 1.5-3
• Examined through simulated air shower data
• maximize quality factor
𝑞 =𝑁𝑎𝛾
𝑁𝑇𝛾
𝑁𝑎𝐶𝑅
𝑁𝑇𝐶𝑅
− 1 2
Approaches: Alternates
• 3D Hillas: more detailed analysis that uses information from the 3 dimensional elliptical shape of the photosphere• Uses Gaussian photosphere to predict light
collected in each pixel• Rejects 70% of hadrons through rotational
symmetry, then width for rest as hadron are wider than gamma
• Model Analysis: pixel-by-pixel comparison to image template generated through simulation• Maximum likeliehood fitting reconstructs shower
parameters such as energy and direction• Hadron-gamma separation conducted through
goodness of fit
Pulse Timing
Razdan et. al. 2002
Smaller light pulses can arrive a couple nsecbefore the main shower
• Another strategy that has proven to be successful is to use characteristics of the light pulse itself
• Muons produced more commonly in showers of hadronic origin penetrate further to the ground• Because they are travelling faster than the speed of
light, their sub-shower’s light arrives first• Method is more effective at higher energies, where
more muons are produced.
• This is shown to be more successful in conjugation with other image characteristics
Gamma Proton
Rise Time 1.82 ns 2.71 ns
Base Width 7.29 ns 8.5 ns
Cab
ot et al., 1
99
8
Neural Networks
• Intelligent machine learning algorithm that is able to learn input-output matching over a vast set of data• Uses adaptive weights: numerical parameters tuned
by decent towards minimum error (Cost function)
• Current studies use Multi-Layer Perceptrons(MLPs)• each neuron in layer i is connected to every neuron in
layer i+1
• Previous studies input Hillas parameters, time parameters, and UV content parameters • achieve quality factors of ~3 and gamma acceptance
of • H.E.S.S. II used MLPs to reconstruct direct and energy
of initial particle and discriminate between gamma/hadron
Summary of Methods
METHOD % GAMMA
RETAINED
% HADRON
RETAINED
QF
Hillas 30-50 1.5 ~4
3D Hillas 40 1 4
Pulse Timing- jitter 67 8 2.3
Pulse Timing- width 98 75 1.3
MLP 90 7 3.4
Still many hadrons retained!
Neural Networks: CNN
• Convolutional neural networks- multiple layers of small neuron collections• Process portions of input image called “receptive fields” resulting in local connectivity
Gives better representation of initial image
• Learned filters pick out specific features and create feature maps in the next layer• Ex: can learn borders, speckling, high and low intensity etc.
Conventional neural networks such as MLPs have a handful of inputs-
Ours has n x 2048 and will explore their evolution in time Initial investigations ongoing
Hyper-Parameters• CNN are defined by several hyper-
parameters: 1. Number of filters
Reduction to 𝑁2 − 𝑁 − 𝑛 2
2. Type and interconnection pattern between neuron layers
3. Learning process for updating interconnection weights• Want to avoid local minima
4. Activation function: converts neuron’s weighted input to its output activation
𝑓 𝑥 = 𝜅
𝑖
𝑤𝑖 𝑔𝑖 (𝑥)
5. Training ’batch size’ and epoch size
• These must be set and experimented upon by the user• Use separate validation set when
training to test hyper-paramters
Layers
Pooling
Fully Connected
ReLURectified Linear Units
Convolutional
Initial Investigation
This learned filter, approximately resembling a Gaussian intensity spread, enables the neural network to recognize stray peaks in intensity
• Using MatConvNet: a neural network package for MatLab, designed by a team working on computer vision at Oxford• Building blocks for CNN assembly
and testing• CPU to GPU translation
Next Steps
• Adapt air shower image simulations to compatibility with MatConvNet CNN• Develop ideal network structure
• Training over vast data sets and picking hyper-parameters: number of layers, learning rate, etc.
• Parallelize for use with multiple telescopes
• Incorporating timing data
• Long Term: Examine possible incorporation of CNN onto FPGAs for triggering• Multi-telescope decision making for triggering
Summary
• Hillas parameters are the most widely used method of separating gamma and hadron shower images, and are still effective• However, conservative cuts leave a lot of CR noise• Initial Hillas results for CTA are comparable to previous IACT telescopes
• Other methods utilize pulse timing information, 3-dimensional shower parameters, or template libraries
• MLPs make more intelligent decisions and can combine the effects of multiple parameters
• CNN show promise as they are able to efficiently process larger amounts of spatially organized information
Still lots of work to do!
Intelligent Imaging for CTA
Laurel KayeSupervisor: Tim Greenshaw
“The research leading to these results has received funding from the People programme(Marie Curie Actions) of the European Unions Seventh Framework Programme FP7/2007-2013/ under REA grant agreement n [317446] INFIERI “INtelligent Fast Interconnected and Efficient Devices for Frontier Exploitation in Research and Industry“.
Thanks!