deep data-driven astronomyrvanderb/kes/sedaghat/dl_eso.pdfneural network... nima sedaghat, sep....

Post on 29-Jun-2020

5 Views

Category:

Documents

0 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Nima Sedaghat, Sep. 2019, ESO, Garching

Deep Data-driven Astronomy

Nima Sedaghat

Nima Sedaghat, Sep. 2019, ESO, Garching

Outline

● Basic Ideas behind Machine Learning● Classical Machine Learning vs. Deep Learning● Rapid Evolution of Deep Learning

→ Encoder-Decoder Architectures

● TransiNet

● Letting Spectra Speak for Themselves

Nima Sedaghat, Sep. 2019, ESO, Garching

Machine Learning

Nima Sedaghat, Sep. 2019, ESO, Garching

30 second intro to machine learning

Nima Sedaghat, Sep. 2019, ESO, Garching

30 second intro to machine learning

DOG? DOG?

DOG?

CAT?

CAT?CAT?

Guess

Nima Sedaghat, Sep. 2019, ESO, Garching

30 second intro to machine learning

DOG? DOG?

DOG?

CAT?

CAT?CAT?

Update

Nima Sedaghat, Sep. 2019, ESO, Garching

30 second intro to machine learning

DOG? DOG?

CAT?

CAT?

DOG?DOG?

EducatedGuess

Nima Sedaghat, Sep. 2019, ESO, Garching

30 second intro to machine learning

Update

DOG? DOG?

CAT?

CAT?

DOG?DOG?

Nima Sedaghat, Sep. 2019, ESO, Garching

30 second intro to machine learning

CAT? DOG?

CAT?

CAT?

DOG?CAT?

EducatedGuess

Nima Sedaghat, Sep. 2019, ESO, Garching

Real-world Example

Human or not?

www.istockphoto.com

Nima Sedaghat, Sep. 2019, ESO, Garching

“Feature”

Nima Sedaghat, Sep. 2019, ESO, Garching

if (there is a circle somewhere close to the top)

if (there are two parallel vertical lines below it)

if ...Probably a human!

else

if ...Probably a human without a neck!

else

not a human

Featrues → Final Answers(inference)

Nima Sedaghat, Sep. 2019, ESO, Garching

From If-else to Learning by Examples

Nima Sedaghat, Sep. 2019, ESO, Garching

From If-else to Learning by Examples

“Model”

Nima Sedaghat, Sep. 2019, ESO, Garching

Model

Human?Not Human?

Nima Sedaghat, Sep. 2019, ESO, Garching

Different Methods – Same Idea

Support Vector Machine

DecisionTree

NeuralNetwork

...

Nima Sedaghat, Sep. 2019, ESO, Garching

Classical Machine Learning vs.

Deep Learning

Nima Sedaghat, Sep. 2019, ESO, Garching

From Hand-Crafted Features to Representation Learning

Alex, D., 2017

Nima Sedaghat, Sep. 2019, ESO, Garching

Why didn’t we do it before?

We needed to go deeper: more parameters, more sophisticated features

It would take forever to tell a cat from a dog!

Nima Sedaghat, Sep. 2019, ESO, Garching

What made it possible?

Convolutional Layers

Nima Sedaghat, Sep. 2019, ESO, Garching

GPU: Graphical Processing Unit

https://www.wikihow.com/Connect-an-External-TV-Tuner-Card-to-a-Desktop

Nima Sedaghat, Sep. 2019, ESO, Garching

GPU: Graphical Processing Unit

Nima Sedaghat, Sep. 2019, ESO, Garching

Convolutional Layers

https://www.analyticsindiamag.com/convolutional-neural-network-image-classification-overview

Nima Sedaghat, Sep. 2019, ESO, Garching

Why Convolutional Layers help?

“Shared Weights”

Seongwon Hwang, “Convolutional Neural Network (CNN) presentation from theory to code in Theano”

Nima Sedaghat, Sep. 2019, ESO, Garching

Deep ConvNets

Nima Sedaghat, Sep. 2019, ESO, Garching

Hierarchical Representation Learning

Nima Sedaghat, Sep. 2019, ESO, Garching

On the Importance of Data

Nima Sedaghat, Sep. 2019, ESO, Garching

On the Importance of Data

over-fittingSmall Dataset

It’s a cat

Nima Sedaghat, Sep. 2019, ESO, Garching

The Rapid Evolution of Deep Learning

Nima Sedaghat, Sep. 2019, ESO, Garching

Classification

CAT? DOG?

CAT?

CAT?

DOG?CAT?

EducatedGuess

Nima Sedaghat, Sep. 2019, ESO, Garching

Classification

https://github.com/EdwardJKim/dl4astro/

Nima Sedaghat, Sep. 2019, ESO, Garching

Regression

Sedaghat, Brox, “Unsupervised Generation of a Viewpoint Annotated Car Dataset from Videos”

Nima Sedaghat, Sep. 2019, ESO, Garching

● Question: What examples can you think of in Astronomy?

Regression

Nima Sedaghat, Sep. 2019, ESO, Garching

Encoder-Decoder Networks

Nima Sedaghat, Sep. 2019, ESO, Garching

E.g. Auto-Encoder

Nima Sedaghat, Sep. 2019, ESO, Garching

Why an Auto-Encoder?

Code

Nima Sedaghat, Sep. 2019, ESO, Garching

E.g. De-noising

Question: How would you train this?

Nima Sedaghat, Sep. 2019, ESO, Garching

E.g. Segmentation

Nima Sedaghat, Sep. 2019, ESO, Garching

E.g. Optical Flow (Motion) Estimation

Nima Sedaghat, Sep. 2019, ESO, Garching

Application Example: TransiNet

Reference Image Science Image Ideal output

Nima Sedaghat, Ashish Mahabalhttps://doi.org/10.1093/mnras/sty613

Watch presentation @ Astroinformatics '18

Nima Sedaghat, Sep. 2019, ESO, Garching

Application Example

Letting Spectra Speak for themselves!

Nima Sedaghat, Sep. 2019, ESO, Garching

The Idea

Nima Sedaghat, Sep. 2019, ESO, Garching

First Results

Nima Sedaghat, Sep. 2019, ESO, Garching

Retrieval

Query Spectrum

Nima Sedaghat, Sep. 2019, ESO, Garching

Retrieval

Search Results:

Nima Sedaghat, Sep. 2019, ESO, Garching

Serendipitous Finding

● Does the network know something we don’t?

top related