A very big epoch size does not always increase accuracy. After one epoch in a neural network, all of the training data had been used to refine the models’ parameters. Epoch sizes may boost precision up to a certain limit, beyond which the model begins to overfit the data. Having a really low level will also result in an improper fit. Observing the enormous discrepancy between epoch 99 and epoch 100 reveals that the model is already overfitting. As a general rule, the optimal number of epochs is between 1 and 10 and should be achieved when the accuracy in deep learning stops improving. 100 seems excessive already.
Batch size does not affect your precision. This is simply used to modify the pace or efficiency of the GPU’s memory. If you have a large amount of memory, you may have a large batch size, making training quicker.
To make sure that your accuracy increase, you can:
- Expand your training dataset;
- Try utilizing Convolutional Networks as an alternative; or
- Try alternative algorithms.
In machine learning, there is a technique called Early Stop. In this method, the error rate on validation and training data is shown. The horizontal axis corresponds to the number of epochs, while the vertical represents the error rate. The training phase should conclude when the error rate of the test dataset is minimal.
In the age of deep learning, it is less common to have an early halt. One of the reasons for this is that deep-learning techniques need so much data that showing the aforementioned graph would be very undulating. If you train excessively on the training data, your model may be overfitting. To address this issue, other strategies are used. Adding noise to various model components, such as drop-out or batch normalization with regulated batch size, prevents these learning methods from overfitting even after a large number of epochs.
In general, an excessive number of epochs may lead your model to overfit its training data. It indicates that your model is memorizing the data rather than learning it.