DSIR-426: BLOG POST 5

‘Viva La Convolución’

Despite rather intimidating technical nomenclature, convolutional neural networks (CNN) have grown increasingly popular and are now commonly implemented by data scientists and programmers agnostic of industries.

CNN’s are very well-suited to handle complex data of high dimensionality, specializing in image or visual processing. Furthermore, CNN’s offer powerful functionality in identifying and filtering out lower-correlation features within a given data set, allowing for refined processing of higher-correlated features or variables of greater importance.

Initial assessment or a general prediction of processing power necessary to compute results is a top priority when passing data through any CNN model. Adding more pooling layers might improve model scores, yet will significantly increase time required for model fitting / evaluation. Assessing train / test scores for both accuracy & loss metrics is strongly recommended whilst interpreting CNN model results.

Though model selection is absolutely requisite upon the shape, size, components, and data types of the data under review, CNN’s are most appropriate for more granular or complex data of greater dimensionality.

CNN’s are best utilized with multi-dimensional data or images. Model results can often be enhanced through additional pooling layers, densification, dropout layers, or ensemble methods such as bootstrapping. http://deeplearning.stanford.edu/tutorial/supervised/ConvolutionalNeuralNetwork/

FINAL WORDS OF WISDOM: Tread with caution around CNN model processing / fit time!! It is easy to get greedy with hyper-parameter tuning and refinement, which can just as easily blow up a model in waiting time as it might marginally improve that same model’s results.