Improving Deep Learning Models in Drug Discovery (Merck)

In this project, we aim to develop improvements to the current Deep Learning methods for drug discovery. These improvements could affect any of the following algorithmic components: * Activation functions: Sigmoids, Rectified Linear Units, Exponential Linear Units, Leaky Rectified Units, etc. * Architecture: Standard, Residual Networks, Highway Networks, deep or broad architectures, connectivity of the architecture, etc. * Regularization techniques: Dropout, weight decay, etc. * Learning techniques: Online learning, gradient descent, stochastic gradient descent, etc. * Initialization strategies * Strategies countering the vanishing gradient problem * Representation of the chemical input data: 2D&3D chemical descriptors, ECFP, DFS, toxicophore descriptors, molecular graph convolutions, etc.