Main projects
New York Taxi Trip Duration Kaggle challenge
Taxi trip duration regression (Kaggle competition) using a fully connected neural network implemented on tensorflow and whose hyperparameters are optimized using hyperopt.
The fully connected neural network is trained using warm restart cyclic learning rate with snapshot ensembling
(see the papers
Architecture and training: Neural network is composed of 10 dense layers with batch normalization and optimized using momentum SGD (ADAM performs better if the learning rate is constant but doesn't plays well with cosine learning rate scheduling and warm restart cycles). The output layer is a softmax layer trained for classification on buckets using cross entropy loss (Discretizing target instead of doing a regression improves performances).
1D GAN
Tensorflow implementation of 1D convolutional Generative Adversarial Network (improved WGAN variant, see the
paper
Learn error by explicit generalization (WIP)
Learn to generalize explicitly by learning error vector on testset batch (Tensorflow, WIP).