Deep Learning Model Deployment in Multiple Cloud Providers: an Exploratory Study Using Low Computing Power Environments

The deployment of Machine Learning models at cloud have grown by tech companies. Hardware requirements are higher when these models involve Deep Learning (DL) techniques and the cloud providers' costs may be a barrier. We explore deploying DL models using for experiments the GECToR model, a DL solution for Grammatical Error Correction, across three of the major cloud platforms (AWS, Google Cloud, Azure). We evaluate real-time latency, hardware usage and cost at each cloud provider by 7 execution environments with 10 experiments reproduced. We found that while GPUs excel in performance, they had an average cost 300% higher than solutions without GPU. Our analysis also identifies that processor cache size is crucial for cost-effective CPU deployments, enabling over 50% of cost reduction compared to GPUs. This study demonstrates the feasibility and affordability of cloud-based DL inference solutions without GPUs, benefiting resource-constrained users like startups.

Optimal Categorical Attribute Transformation for Granularity Change in Relational Databases for Binary Decision Problems in Educational Data Mining

This paper presents an approach for transforming data granularity in hierarchical databases for binary decision problems by applying regression to categorical attributes at the lower grain levels. Attributes from a lower hierarchy entity in the relational database have their information content optimized through regression on the categories histogram trained on a small exclusive labelled sample, instead of the usual mode category of the distribution. The paper validates the approach on a binary decision task for assessing the quality of secondary schools focusing on how logistic regression transforms the students and teachers attributes into school attributes. Experiments were carried out on Brazilian schools public datasets via 10-fold cross-validation comparison of the ranking score produced also by logistic regression. The proposed approach achieved higher performance than the usual distribution mode transformation and equal to the expert weighing approach measured by the maximum Kolmogorov-Smirnov distance and the area under the ROC curve at 0.01 significance level.