I am performing a pattern recognition classification of force outputs with electromyography inputs. The inputs of EMGs are the features extraction from filtered EMG data. While performing classification; the classification accuracy is almost 100% in a few cases. Under what circumstances; should I decide if the case is causing overfitting? Similarly; when is classification accuracy judged as underfitting?
I suggest you follow https://towardsdatascience.com/handling-overfitting-in-deep-learning-models-c760ee047c6e
https://elitedatascience.com/overfitting-in-machine-learning
https://www.kdnuggets.com/2019/12/5-techniques-prevent-overfitting-neural-networks.html
Overfitting occurs when the classifier becomes too specific while being trained by the training data. As a result, it can classify the samples on the training data very accurately (as it has already seen them) but performs poorly on the testing data (as it has never seen them). Broadly, overfitting can be judged by looking at the training accuracy and test accuracy scores; If the difference is huge then the model is overfitted. If the model can't classify the training samples as well, then it is underfitted.
You can shuffle the samples properly/multiple times before dividing the training and testing subsets and/or employ some pruning/dropout conditions to reduce the effects of overfitting. To solve underfitting you might need to employ a better classifier (that can create complex and more flexible version spaces, see the link below) and/or extract new/better/more features from the main data that represent the samples more elaborately.
You mentioned that you are working with EMG data but didn't mention the classifier or the feature extraction algorithm. Time-domain signals are very noisy, transforming them into frequency/wavelet domain is likely to improve the classification accuracy.
https://scikit-learn.org/stable/auto_examples/classification/plot_classifier_comparison.html
Niloy Sikder has explained very well the overfitting and underfitting of the trained model. As you are working with EMG data, the data preparation of biomedical signals is a bit tricky part to be used for pattern recognition. I suggest, before moving forward, look for data preparation. The problem might be with the data preparation part. I have explained all the steps from the data preprocessing to training your model in one of my papers. Please have a look:
Article Determination of Optimum Segmentation Schemes for Pattern Re...
If you need any clarity, do not hesitate to contact me.
check this paper, overfitting section
Article Review of deep learning: concepts, CNN architectures, challe...
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification