Please enable JavaScript.
Coggle requires JavaScript to display documents.
Anais Moller- Bayesian Neural Network light-curve classification for time…
Anais Moller- Bayesian Neural Network light-curve classification for time-domain astronomy
Hubble diagram and cosmic expansion
astrophysical objects
Measure distances (distance modulus)
Measure velocities (redshift)
model agreement with the data
Supernovae
Stellar explosions (transient events)
types:
1a (thermonuclear)
very luminous
Homogeneous spectral and photometric properties
11, 1b, 1c (Core collapse)
Find SNe 1a to study the cosmic expansion
Current surveys:
5 year survey, started 2013
4 primary probes: galaxy clusters, weak lensing, large scale structure, type 1a SNe
<2000 well measured SNe 1a
How to find SNe 1a
Difference imaging
Don't want bogus detections
ML classifier Random forest
skymapper images augmented for training
limitations: training sets! feature extraction
Selecting SNe 1a, & get host galaxy spectra for precise redshifts
zDES
Limited by spectrocopic efficiency
Also limited by Telescope time, telescope type, (limiting magnitude), Humans
Only get spectra of < 20% of ood quality SNe 1a
To improve limitations:
Improve selection function (non-human system?)
change how classification of SN is done
Photometric classification can help
Future Surveys
LSST
10 year survey starting 2022
1000 images/night = 15 TB/night
10,000 alerts/30 seconds = 1GB/30s
4000 well measured SNe 1a
photometric classification challenges
Volume
Nonuniform light-curves
Many transient types
many unknowns
Science time constraints
Typing supernovae with photometry
Early classification brokers Spectroscopic follow-up around maximum
Complete light-curve classification, larger and more reliable samples, probing new parameter space
Most approaches: fit-parameters, sequential cuts, machine learning an other cuts, science sample
New approach just machine learning to science sample
SuperNNova
Inputs
Flux in different band passes
Flux measurement errors
Time-step between measurements
optimal:other features (e.g. host redshift)
Deep learning for cosmology
Recurrent neural networks (RNNs)
Bayesian RNN
Convolutional NN
Early classification
For brokers and follow up for promising candidates
Reproducable selection functions
Improving photometric classification training samples
87.59
Complete
96.97
For larger and more reliable samples and probing parameter space
~2% contamination
Cosmology limitation: Modelling core-collapse contamination
Bayesian NNs
Bayesian: distribution of weights
Posterior is intractable for deep neural networks
maths
Approximating the variational distribution
MC dropout
Bayes by Backprop
Posterior that provides epistemic uncertainties
Epistemic uncertainties: express out ignorance about the model that generated the data
ML limitations
Training sets:
Not representitive
incomplete (we don't know/can't simulate)
Can we use output from ML classifiers for cosmology or any statistical analyses
Representativity
Combine simulations (Simplistic and representitive)
non-representative models give larger uncertainties
BNNs can give high probability but large uncertainty
ML probabilities for statistical analysis
Selecting a SN 1a sample: cutting on "classification probabilities" for selection
Can use a weight in the analysis using these classification probabilities