AI & Radioastronomy

Requirements

1. Anaconda3
2. Python 3.6.x
3. jupyter notebook
4. Tensorflow 1.14
5. Keras 2.2.4
6. numpy 1.16.2
7. matplot 3.1.0

import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt

How to use

1. activate virtualenvironment with package management system such as pip or anaconda.
2. Run the jupyter notebook.
3. Run Exo.ipynb.
4. You can modify second block your own dataset about Exoplanet.
5. Run!

Results

We got 99.12% accuracy on the test dataset

Things we've tried to improve

1. Add an initializer to initialize every weight to a random number before training instead of 0 as default value.

tf.keras.initializers.RandomUniform(minval=-500, maxval=500, seed=None)

2. Playing with the optimizer by lowering the learning rate.

optimizer=(tf.keras.optimizers.Adagrad(lr=0.02, epsilon=None, decay=0.0)

3. Make more epochs with larger batches.

model.fit(data_train, label_train, validation_split=0.1, batch_size=100, epochs=5)

4. Make a deeper network with more layers. and changing the activation function.

tf.keras.layers.Dense(512, activation=tf.nn.relu),
tf.keras.layers.Dense(64, activation=tf.nn.sigmoid),
tf.keras.layers.Dense(2, activation=tf.nn.sigmoid)

5. Add dropout layers, it prevent for overfitting while allowing for better training.

tf.keras.layers.Dropout(0.5)

Things you could try to improve

1. Change the loss function.

loss='binary_crossentropy

2. Change/Tweak the optimizer.

optimizer=(tf.keras.optimizers.Adagrad(lr=0.02, epsilon=None, decay=0.0)

3. Trying other type of network.

tf.keras.layers.Dense(512, activation=tf.nn.relu),
tf.keras.layers.Dense(64, activation=tf.nn.sigmoid),
tf.keras.layers.Dense(2, activation=tf.nn.sigmoid)