Binary-Classification--Equivariant-Neural-Network

Models for Binary Classification using Equivariant Neural Network

PROJECT

Equivariant Neural Networks for Dark Matter Morphology with Strong Gravitational Lensing.

GOAL

Use an Equivariant Neural Network to build a robust and efficient model for binary classification on the lensing dataset.

EVALUATION METRICS

ROC curve (Receiver Operating Characteristic curve) and AUC score (Area Under the ROC Curve)

APPROACH

• I have used General E(2)-Equivariant Steerable CNNs to implement the solution for the task.

• Test_Cyclic.ipynb builds a model equivariant to 8 rotations. We indicate the group of N discrete rotations as C_N which means the Cyclic Group of order N. We implement C₈.

• Test_Dihedral.ipynb builds a model based on Dihedral group D_2N which is group of N discrete planar rotations and reflections. D_N has an order of 2N. We implement D₄ in this notebook which expand to 4 rotations and 4 reflections.

• Model Weights for Dihedral Group Equivariance : https://drive.google.com/file/d/1pgGjBwAUHTiXejiWA3hpHuHRB1BkIwEf/view?usp=sharing

• Colab Notebook for Dihedral : https://colab.research.google.com/drive/1q6E-qFMwR7hxYfBbnV9YUZqR6VfAQDJ-?usp=sharing

• Model Weights for Cyclic Group Equivariance : https://drive.google.com/file/d/1AKAGe-cecKObsnBrK76LTBHNqFwz1bOa/view?usp=sharing

• Colab Notebook for Cyclic : https://colab.research.google.com/drive/1tHji-woLdPQsZFJ0ZD4NjD--GjtgFqwP?usp=sharing

ACCURACIES

Model	Architecture	Accuracy in % (on testing data)
C8	C8 Steerable CNN	99.6
D4	D4 Steerable CNN	99.4

AUC SCORE

Model	AUC Score
C8	1.0
D4	1.0