Convolutional Neural Network Binarized And ECG Signal Analysis Application

As an important component of a broader family of machine learning methods,deep learning is generally considered as a representation learning approach based on artificial neural networks.Recent revolutionary advancement of deep learning technology has prompted neural network models to be used in a large number of realistic scenarios.However,deep neural networks are typically associated with large parameter size and immense computing complexity which limit its applications in wider scenarios.Hence research on neural network model compress and acceleration is one of the technology hotspots today.Arrhythmia,also called cardiac arrhythmia,is a group of conditions caused by abnormalities of cardiac rhythmic.Common arrhythmia symptoms include atrial fibrillation,ventricular fibrillation,tachycardia and so on.Certain arrhythmia conditions can develop into serious conditions if not treated in a timely manner.Currently,the main diagnostic method for arrhythmia is the electrocardiogram(ECG)test,but currently ECG analyses are mainly performed by physicians manually,which would be costly and often not accessible for real-time applications.Therefore,it is desirable to develop automatic ECG analysis algorithms.In this study,we propose to develop an algorithm based on Deep Convolutional Neural Network(CNN)for arrhythmia detection.We first build a full-precision deep convolutional network model.Our model achieved high performance on “The Physio Net Computing in Cardiology Challenge 2017” dataset.Full-precision models often can deliver good performance;however,they are typically associated with large computing complexity that is not feasible for embedding devices.To achieve a more efficient and more economical model,we propose to compress the network model by binarization.While a binarized model brings us the benefit of significantly lower computation and storage complexity,it also can lead to a significant reduction in model performance.Therefore,we propose to use the knowledge distillation to train the binarized network model with the full-precision network model as the teacher model.In this way,the knowledge embedded in the full-precision model is transferred to the binarized model.This leads to improved trainability and the final performance of the binarized model.In our experiment,the na?ve binarized model obtained an F1 score of 0.85,which is significantly lower than the score of 0.88 attained by the full-precision model.With the knowledge distillation training approach,the F1 score of the binarized model is improved to 0.87.This shows that the 1-dimensional convolutional network compressed by the binarization method can achieve performance close to that of a fullprecision network through proper training strategies.The relevant code in this paper has been open sourced in https://github.com/yangfansun/bnn-ecg.