A data scientist’s guide to direct imaging of exoplanets

Invited talk at the RADA big data workshop (https://as595.github.io/RADABigData/) in Medellín, Colombia. The workshop is part of the Radio Astronomy for Development in the Americas (RADA) Big Data program, which is jointly funded by the UK Newton Fund and the Ministerio de Tecnologías de la Información y las Comunicaciones de Colombia. Feb 12, 2019.

Transcript

1. A data scientist’s guide to direct imaging of exoplanets Carlos Alberto Gomez Gonzalez RADA big data workshop (Medellín). Feb 12, 2019
2. whoami Colombia Russia Spain Belgium France (PhD astrophysics and ML/CV)
3. whoami Research data scientist for Earth, Space and environmental sciences
4. whoami I have a very particular set of skills :)
5. A data scientist’s guide to direct imaging of exoplanets
6. A data scientist’s guide to direct imaging of exoplanets
7. A data scientist’s guide to direct imaging of exoplanets
8. A data scientist’s guide to direct imaging of exoplanets Domain expertise
9. A data scientist’s guide to direct imaging of exoplanets Programming
10. A data scientist’s guide to direct imaging of exoplanets Stats, ML
11. A data scientist’s guide to direct imaging of exoplanets Visualization, communication skills
12. Credit: Martin Vargic
13. Solar System A data scientist’s guide to direct imaging of exoplanets
14. Exo from extrasolar (outside of the Solar System) A data scientist’s guide to direct imaging of exoplanets
15. Exo from extrasolar (outside of the Solar System)
16. A data scientist’s guide to direct imaging of exoplanets
18. Credit: NASA, http://planetquest.jpl.nasa.gov High-contrast imaging Lighthouse -> Star Firefly -> exoplanet
19. Credit: NASA, http://planetquest.jpl.nasa.gov High-contrast imaging Lighthouse -> Star Firefly -> exoplanet Stars never shut off !!!
20. http://exoplanetarchive.ipac.caltech.edu, 25 Jan 2018
21. http://exoplanetarchive.ipac.caltech.edu, 25 Jan 2018 Task: try to spot this color and estimate the total contribution of direct imaging detection… about 1% !!!
22. Debris disk Beta Pictoris (occulted)
23. Yep, this is the exoplanet (Beta Pictoris b)
24. HR8799 bcde (Marois et al. 2008-2010) Several epochs (final images)
25. HR8799, Marois et al. 2010 20 AU 0.5” Konopacky et al. 2013 Bowler 2016 Why direct imaging? Milli et al. 2016
26. Fake planet (low contrast) Signal and noise
27. spatially rescaled images t0 t0 with an integral field spectrograph Signal and noise
28. Multi-dimensional arrays ( ) exoplanet t, λ, x, y
29. Image processing pipeline
30. Image processing pipeline Getting rid of the noise…
31. Image processing pipeline Look ma, I’ve found an exoplanet!
32. Raw astronomical images Sequence of calibrated images Basic calibration and “cosmetics” • Dark/bias subtraction • Flat fielding • Sky/thermal background subtraction • Bad pixel correction Image recentering Bad frames removal Image combination PSF modeling • Median • Pairwise, ANDROMEDA • LOCI • PCA/KLIP, NMF • LLSG Model PSF subtraction Detection on residual frame or detection map Characterization of “detected” companions R1 ,!1 ,F1 R2 ,!2 ,F2 R3 ,!3 ,F3 R4 ,!4 ,F4 Raw Clean
33. Algorithms Ten years of research: • Median frame subtraction • Pairwise subtraction • Least squares image combination • PCA (forward modeling), NMF • Low-rank plus sparse decompositions • Matched filtering • Maximum likelihood estimation
35. Researcher creates an open-source library
38. • https://github.com/vortex-exoplanet/VIP • http://vip.readthedocs.io/ • CI, test suite, jupyter tutorials, paper on A&A
40. VIP: algorithmic zoo Comparing the resulting images (using different algorithms)
41. Task: remove the noise and reveal the exoplanet signal VIP: algorithmic zoo
42. Designing APIs for science
43. Designing APIs for science
44. Building blocks… D — M = R Noise reduction algorithm R - residuals containing the exoplanet signal
45. Observer (classifier) D — M = R or Noise reduction algorithm R - residuals containing the exoplanet signal Building blocks…
46. Model PSF subtraction drawbacks Planet signal is subtracted along with the speckles
47. Image sequence Final residual image Detection
48. Image sequence Final residual image ? ? ? ? ? ? ? Detection
49. Image sequence Final residual image ? ? ? ? ? ? ? Speckles (?) Real planet Synthetic planets Detection
50. Unsupervised Supervised PC 1 PC 2 Dimensionality reduction Clustering Reinforcement Density estimation Regression Classification ML
51. f = arg min fθ,θ∈Θ n i=1 L(yi, fθ (xi )) + g(θ) f : X → Y, Training data chihuahua muffin (xi, yi )i=1,...,n Supervised learning
52. { Model architecture f = arg min fθ,θ∈Θ n i=1 L(yi, fθ (xi )) + g(θ) f : X → Y, (xi, yi )i=1,...,n Supervised learning
53. Loss function and regularization f = arg min fθ,θ∈Θ n i=1 L(yi, fθ (xi )) + g(θ) f : X → Y, (xi, yi )i=1,...,n Supervised learning
54. { Optimization f = arg min fθ,θ∈Θ n i=1 L(yi, fθ (xi )) + g(θ) f : X → Y, (xi, yi )i=1,...,n Supervised learning
55. Input X 1st Layer (data transformation)
56. Input X 1st Layer (data transformation) Perceptron (Rosenblatt 1958)
57. ▸ Activation function Input X 1st Layer (data transformation) 2nd Layer (data transformation) Nth Layer (data transformation) …
58. ▸ Max pooling ▸ Dropout ▸ BatchNorm Input X 1st Layer (data transformation) 2nd Layer (data transformation) Nth Layer (data transformation) …
59. Input X Nth Layer (data transformation) … 1st Layer (data transformation) 2nd Layer (data transformation) Predictions Ŷ Labels Y Loss function weights weights weights weight update Optimizer loss score
60. Reframing the problem: from unsupervised to supervised learning N … • Sequences of images without labels • Not enough archival data (observed stars) • We can generate semi- synthetic data by injecting a planet (PSF) template! • We grab patches: signal/noise PSF template
61. Labeled data generation Neural network design and training Prediction (detection) Gomez Gonzalez et al. 2018
62. Gomez Gonzalez et al. 2018
64. • Reproducible results • Hyper-parameter and network architecture tuning • Comparison of labeling strategies DataLabeler Model Predictor • Flux vs S/N sampling • Fluxes/contrast estimation • Training data generation • Data augmentation • Data persistence (load/save with HDF5) • Network creation (Keras and Tensorflow) • Model training • Model persistence (load/ save with HDF5) • Target samples generation • Predictions (based on trained model) • Probability map inspection • Results to HDF5 SODINN library
65. Corresponding labels: y ∈ {c−, c+} … MLAR samples C+ C-
66. Discriminative model: Neural Network Goal - to make correct predictions on new samples: f : X → Y ˆ y = p(c+| MLAR sample) SGD with a binary cross-entropy loss: L = − n (yn ln(ˆ yn ) + (1 − yn ) ln(1 − ˆ yn )) Learning a mapping function
67. 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 3d Convolutional layer kernel=(3x3x3), filters=40 3d Convolutional layer kernel=(2x2x2), filters=80 Dense layer units=128 Output dense layer units=1 3d Max pooling size=(2x2x2) 3d Max pooling size=(2x2x2) ReLU activation + dropout Sigmoid activation X and y to train/test/validation sets 2d Convolutional layer kernel=(3x3), filters=40 Dense layer units=128 Output dense layer units=1 ReLU activation + dropout Sigmoid activation X and y to train/test/validation sets 2d Convolutional layer kernel=(3x3), filters=40 2d Convolutional layer kernel=(3x3), filters=40 2d Convolutional layer kernel=(2x2), filters=80 2d Convolutional layer kernel=(2x2), filters=80 2d Convolutional layer kernel=(2x2), filters=80 … … B-LSTM / B-GRU layer 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 2d Max pooling size=(2x2) 3d Convolutional layer kernel=(3x3x3), filters=40 3d Convolutional layer kernel=(2x2x2), filters=80 Dense layer units=128 Output dense layer units=1 3d Max pooling size=(2x2x2) 3d Max pooling size=(2x2x2) ReLU activation + dropout Sigmoid activation X and y to train/test/validation sets 2d Convolutional layer kernel=(3x3), filters=40 Dense layer units=128 Output dense layer units=1 ReLU activation + dropout Sigmoid activation X and y to train/test/validation sets 2d Convolutional layer kernel=(3x3), filters=40 2d Convolutional layer kernel=(3x3), filters=40 2d Convolutional layer kernel=(2x2), filters=80 2d Convolutional layer kernel=(2x2), filters=80 2d Convolutional layer kernel=(2x2), filters=80 … … B-LSTM / B-GRU layer
68. Let’s inject some fake planets
69. VS Let’s inject some fake planets
70. PCA least squares Find the exoplanet ;)
71. VS VS PCA least squares SODINN SODINN
72. Performance assessment Receiver operating characteristic curves
73. https://carlgogo.github.io/exoimaging_challenge/ https://github.com/carlgogo/exoimaging_challenge_extras Do you have an idea for a new algorithm? We are about to launch a data challenge!
74. • Data from the most representative instruments • Metrics: • Phases: • Will run on Codalab https://carlgogo.github.io/exoimaging_challenge/ https://github.com/carlgogo/exoimaging_challenge_extras
75. Connections with other fields https://arxiv.org/pdf/1506.04214.pdf
76. Connections with other fields http://juliandewit.github.io/kaggle-ndsb2017/
77. Connections with other fields https://arxiv.org/pdf/1609.09143.pdf ReCTnet architecture for object detection in multi-slice medical images
78. Connections with other fields https://arxiv.org/pdf/1811.02471.pdf two-component convolutional long short-term memory network (LSTM) Convolutional LSTMs for Cloud-Robust Segmentation of Remote Sensing Imagery
79. ¡Gracias! carlgogo.github.io/ github.com/carlgogo/ speakerdeck.com/carlgogo/ linkedin.com/in/carlgogo/