Sensitivity analysis for deep learning: ranking hyper-parameter influence

We propose a novel approach to ranking Deep Learning (DL) hyper-parameters through the application of Sensitivity Analysis (SA). DL hyper-parameters play an important role in model accuracy however, choosing optimal values for each parameter can be time and resource-intensive. A better understanding of the importance of parameters in relation to data and model architecture would benefit DL research. SA provides a quantitative measure by which hyper-parameters can be ranked in terms of their contribution to model accuracy allowing comparisons to be made across datasets and architecture. The results showed the importance of optimal architecture with activation function being highly influential. The influence of learning rate decay was ranked highest, with model performance being sensitive to this parameter regardless of architecture or dataset. The influence of a model's initial learning rate was proven to be low, contrary to the literature. The results also showed that the importance of a parameter is closely linked to model architecture. Shallower models showed susceptibility to hyper-parameters affecting the stochasticity of the learning process whereas deeper models showed sensitivity to hyper-parameters affecting the convergence speed. Additionally, the complexity of the dataset can affect the margin of separation between the sensitivity measures of the most and the least influential parameters, making the most influential hyper-parameter an ideal candidate for tuning compared to the other parameters