This package facilitates application of DeepInsight (DI) and Visible Neural Network (VNN) algorithms from Alok Sharma1 and Michael Ku Yu2, respectively. The application is intended for supervised machine learning by convolutional neural network (CNN). DeepInsight converts non-image data into image-like data by dimensionality reduction algorithms. This package maps the data into a multi-dimensional array. Meanwhile, VNN determines a neural network architecture by hierarchical clustering algorithms, particularly for data-driven ontology. This package generate a CNN model based on the ontology using the DeepInsight array as the input. However, this package includes neither dimensionality reduction nor data-driven ontology inference. A comprehensive guide to orchestrate this package and other packages to develop the DI-VNN model is described in this package vignette. The inputs are instance-feature value table, outcome vector, feature similarity table, feature three-dimensional mapping table, and ontology source-target-similarity-relation table. The outputs are tidy (expression) set, training array, and Keras CNN model.
Read vignette for simple example in R Read vignette for simple example in python[1] Sharma A, Vans E, Shigemizu D, Boroevich KA, Tsunoda T, DeepInsight: a methodology to transform a non-image data to an image for convolution neural network architecture, Scientific Reports, 9:11399, pp. 1-7, 2019. Paper.pdf Supplement .pdf DeepInsight Matlab Code: Tested on Ubuntu 18.10.
[2] Ma, J., Yu, M., Fong, S. et al. Using deep learning to model the hierarchical structure and function of a cell. Nat Methods 15, 290–298 (2018). Paper Author GitHub DCell code