Combining Deep Learning with Traditional Machine Learning to Improve Classification Accuracy on Small Datasets
MetadataShow full metadata
Feature extraction and selection are essential phases in building machine learning classification models, and they have a great impact on the accuracy and the performance of the model. However, these phases are expensive, and there is no guarantee that manually extracted features will generalize well in different data modalities. Deep learning models integrate the phases of feature extraction, selection, and classification into a single optimization process. However, they are very computationally expensive compared to traditional machine learning algorithms, and they require large training datasets to achieve good classification performance.
This work explores ways of combining the advantages of deep learning and traditional machine learning models by building a hybrid classification scheme. The first few layers of a convolutional neural network are utilized for feature extraction and selection. Subsequently, the extracted features are fed to a traditional supervised learning algorithm to perform classification. We evaluate our method on sensor data coming from human physiological biosignal measurements and motion tracking data coming from accelerometers. Our experimental results show that our hybrid approach outperforms deep learning and traditional machine learning algorithms when those are used in isolation on small dataset.