M. Bari
T-Mobile,
United States
Keywords: base model, convolutional neural network, resnet, deep learning, inception net, imagenet, mnist database
Summary:
For any learning problem, the first task for data scientists after performing exploratory data analysis (EDA) is to have a base model (BM). The base model is then used as a benchmark to obtain better performing models. This summary proposes to utilize the most commonly used state-of-the-art convolutional neural network (CNN) architectures and its variations to obtain reasonably good base models. This work reports the algorithms' performance metrics, and training and testing times. The training and test times measure the computational complexity of the architecture. The accuracy, as well as the computational complexity, enable the readers to choose the right architecture in an informed way for their use-cases. For benchmarking purpose, the problem of handwritten digits recognition in the MNIST database is used in this summary. In the full paper, the performance for a few other publicly available datasets, including ImageNet, will also be reported. The approach used for images can be extended to other data types as well.