2021.2

Big Data and Astroinformatics:

Exercises:

1. Create a new column from a FITS catalog and save the new catalog

2. Estimate the limiting magnitude of the provided catalog. Use the relation between mag_err and S/N

sextractor file_name.fits -c config.sex -PARAMETERS_NAME sex.param -CATALOG_NAME output.fits -CATALOG_TYPE FITS_1.0 -GAIN 8116 -PIXEL_SCALE 0.06 -SEEING_FWHM 0.1 -MAG_ZEROPOINT 25.6651 -PHOT_FLUXFRAC 0.682 -DEBLEND_MINCONT 0.0015 -DEBLEND_NTHRESH 32

You have to create your sex.param file. Please refer to Source Extraction documentation.

For the lecture on Tuesday, Sept-28, please install:

4. Derive your Hubble constant from several datasets hubble_fit_data

5. Demonstrate that an Ordinary Least Squares fit is equivalent to a Maximum Likelihood of a Gaussian variable with a linear model

6. Demonstrate that the Maximum Likelihood Method for a Bernoulli variable with a linear model is equivalent to minimizing the cross-entropy between predictions and measured variables.

7. Demonstrate that the minimum of cross-entropy is obtained when the two probabilities are equal.

Proposed projects for 2020.b

1- Not so few, not so Big. What is the data limit to train and converge gravitational arc identification models (Images)?

2- What is the detection limit for a gravitational lens (images)?

3 - Few shot learning for classification of gravitational lenses (images).

4 - Model for star/galaxy separation (tabular/images).

5 - Uncertainty for regression in gravitational lenses. (inverse modeling in images).

6 - Exoplanet detection https://www.kaggle.com/keplersmachines/kepler-labelled-time-series-data (sequence)

7 - Galaxy morphological classification with Deep Learning https://www.kaggle.com/c/galaxy-zoo-the-galaxy-challenge/data (images)

8 - Transient classification using photometry https://www.kaggle.com/c/PLAsTiCC-2018 (sequence)

9 - Credit card fraud detection https://www.kaggle.com/mlg-ulb/creditcardfraud (tabular)

10 - Pneumonia detection https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia (images)

11 - Cardiac arrest mortality prediction https://www.kaggle.com/andrewmvd/heart-failure-clinical-data/ (tabular)

12 - Wine quality prediction https://www.kaggle.com/uciml/red-wine-quality-cortez-et-al-2009/ (tabular)

13 - Heartbeat classification https://www.kaggle.com/kinguistics/heartbeat-sounds (audio)

14 - Sentiment analysis from audio https://www.kaggle.com/uwrfkaggler/ravdess-emotional-speech-audio (audio)

15 - CT segmentation for COVID-19 detection https://www.kaggle.com/andrewmvd/covid19-ct-scans (Images / Semantic segmentation)

Tasks

Task 1 - Diabetes diagnosis Diagnosis dataset

Task 2 - First dense network with MNIST

Task 3 - First CNN with MNIST

Task 4 - Classification in a mini-challenge:

Part 2: Evaluate your results by generating a ROC curve and calculating its area.

Part 3: Test the robustness by shuffling training and test sets multiple times and computing average and standard deviation of the ROC.

Part 4: Evaluate the overfitting of your network.

Part 5: Add another Conv+ReLu+Maxpool block and repeat step 3. Did it improve? Any insights?

Task 5 - First kaggle challenge: https://www.kaggle.com/c/titanic/overview/evaluation

Task 6 - Do it yourself:

Goal: Without Keras or Tensorflow, implement and train a network from scratch in Python.

Task 7 - Regression on kaggle: https://www.kaggle.com/c/house-prices-advanced-regression-techniques

1. Build a model with Random Forest

2. Build a model with LightGBM

3. Make an Ensemble

Task 8 - Apply your architecture to the Strong Lensing dataset and compare.

Task 9 - Time series with Deep Learning

1 - Follow the tutorial: aboveintelligent.com https://aboveintelligent.com/time-series-analysis-using-recurrent-neural-networks-lstm-33817fa4c47a

2 - Follow the tutorial: Kaggle https://www.kaggle.com/amirrezaeian/time-series-data-analysis-using-lstm-tutorial

3 - Implement an LSTM for the rainfall classification challenge https://www.kaggle.com/c/how-much-did-it-rain#description
Replicate the winning solution: https://github.com/danzelmo/how-much-did-it-rain
Compare your results using the official metric: https://www.kaggle.com/c/how-much-did-it-rain#evaluation

Task 10 - AE vs PCA

Show that PCA is equivalent to AE with linear functions under a specific loss function.

Introduction to Scientific Methodology

Introduction to Scientific Methodology from UFSM Methodology (Unit 1 and 2)Capes Methodology Guide (Unit 1 and 2)

P2:

Seminar topics, groups of up to 4 people - 20 min

Antiscience movements: Flat Earth
Antiscience movements: Anti-Vaccine Movement
Theory of Evolution vs Intelligent Design
From the scientific method and reproducible observations, discuss Heliocentrism.
Genetically modified foods under the lens of science: are they harmful?
Practices that claim scientific support: Homeopathy
Practices that claim scientific support: Quantum Healing
Global Warming

Additional material:

Flat Earth - Documentary
Mindwalk - Film
The Name of the Rose - Film
Why is philosophy important in teaching science in universities? - Folha de SP column
There is no such thing as exact science (and let’s agree they’re all human)
Why do I teach Plato to plumbers
Philosophy of science and science education: an analogy - Alberto Villani

2018.2

• Deep Learning
• Basic Electricity
• Scientific Methodology
• Tim 4
• Tim 2

Deep Learning

Task 1 (classification):
https://bitbucket.org/kognitalab/images_mini_challange/src/master/
Part 2: Evaluate your results by generating a ROC curve and calculating its area.
Part 3: Assess the robustness of your results by randomly redefining the training and test samples at least 10-20 times and computing the curve's mean and standard deviation.
Part 4: Evaluate the overfitting of your network.
Part 5: Add another Conv+ReLu+Maxpool block and repeat step 3.

Task 2 (regression):
Where is Waldo?
Goal: Given a set of images, build a network to identify the location of an object by defining 4 parameters: xmin, ymin, xmax, ymax.
The data is available at: /home/Dados/GeoSimula
Evaluate your results using a completeness vs centroid distance plot. Calculate the area of this figure and produce some example images with bounding boxes.
If you really want to find Waldo, go to: findingwally.pythonanywhere.com

Task 3 (Do it yourself!):
Goal: Without Keras or Tensorflow, implement and train a network with a few layers using only Python.
Use the example from: https://machinelearningmastery.com/implement-backpropagation-algorithm-scratch-python/
Challenge: Implement a convolutional layer and create a notebook using your example.
Part 2: Compare your results with an identical network built in Keras/Tensorflow.

Task 4:
Apply an architecture used in Task 01 (with proper adaptations) — like ResNet, ResNeXt, Inception — to the Strong Lensing Challenge dataset...

Task 5 - Time Series:
1 - Follow the tutorial at: aboveintelligent.com - LSTM Tutorial
2 - Follow the tutorial at: Kaggle - LSTM Time Series Tutorial
3 - Implement an LSTM network for the rainfall classification challenge: Kaggle - How much did it rain?
Build your own implementation of the winning solution: GitHub - danzelmo
and compare its performance using the evaluation metric: Kaggle - Evaluation

Task 6 - AE vs PCA:
Demonstrate that PCA is equivalent to an AutoEncoder with linear functions for a specific choice of loss function.

Recommended reading:
Introduction to Convolutional Neural Networks - Wu
A Beginner’s Guide To Understanding Convolutional Neural Networks
The Deep Learning Book

Suggested tools:
Tensorflow, Anaconda Python 2.X or 3.X, Keras, Jupyter

Additional reading:
Planes don’t flap their wings: does AI work like a brain?

Important dates:

Project definition deadline - 29/08
Work summary submission - 26/09
Seminars with preliminary results - 24/10 and 31/10
Updated code on GitHub - 14/12
NT submitted - 14/12

The final course project must be submitted (at least) to the Notas Técnicas journal: http://revistas.cbpf.br/index.php/nt/about/submissions#authorGuidelines

Introduction to Scientific Methodology (Production Engineering)

Supplementary material:
Why is philosophy important in science education at universities? - Folha de SP
There is no exact science (and let's agree all are human)
Why do I teach Plato to plumbers
Philosophy of science and Science Teaching: an analogy - Alberto Villani

Suggested debate topics:
Galileo and Giordano Bruno Intelligent design vs. Evolution Flat Earth vs. Round Earth Did man go to the Moon? Do cell phones cause cancer? Are GMOs harmful? Do vaccines cause more harm than good?

Rules:
Arguments based on inaccessible technology are not allowed... All claims must be sourced and demonstrable.

Debate dates:
10/04 - Flat Earth / Moon landing 10/11 - GMOs / Intelligent Design 10/18 - Vaccines / Cell phone cancer

Tim 2
list01
list02
list03

Tim 4
list01
list02 - Gravitation
Note: In some exercises, you may use online values such as:
https://pt.wikipedia.org/wiki/Terra
https://pt.wikipedia.org/wiki/Lua

Notice: No classes on 10/05 (Friday) and 10/26 (Friday).