Hello, I am using a VGG16 on keras to do some image classification on a dataset, What does it mean that the validation loss is smaller all the time than the train loss?. Is 80% a reasonable accuracy for this kind of problem?. Any tips on how to approach this kind of classification would very much be appreciated! Thanks in advance.
The problem you mentioned is known as generalization gap. There are some research papers regarding them you can search them in google scholar easily.
The point is you've to minimize the gap between your validation loss and training loss (which in turn might improve your accuracy).
There are several methods for minimizing generalization gap in deep learning models (such as VGG). The best is the Dropout[ Article Dropout: A Simple Way to Prevent Neural Networks from Overfitting
]. Please go through the paper to know how it generalizes various neural networks with a very less computational cost.This Dropout act as a layer in the Keras engine[ https://keras.io/api/layers/regularization_layers/dropout/ ] as Keras is designed to understand a neural network as a layered model (Either Sequential or Functional API). As Keras run tensorflow in the backend you can either impart[ https://www.tensorflow.org/api_docs/python/tf/keras/layers/Dropout ].
There are many such regularization methods such as Batch Normalization[ Article Batch Normalization: Accelerating Deep Network Training by R...
], Layer Normalization (for text data)[Article Layer Normalization
] and some generally used methods[ https://keras.io/api/layers/regularization_layers/].I hope this information might be useful.
It Is good to have a low validation loss. This shows how your model behaves in unseen problems. We are more interested in the validation curve than the training curve. Our goal is not to have a minimum training loss as that would overfit our data. The validation set is a set that is use to modify our parameters of the classification model. We always try to have a low validation loss.
1)Loss can be calculated in different ways, and that's why its absolute value can be dependent on sizes of your training dataset and validation dataset.
2)if your validation dataset is a part of your training dataset, then for some "lucky" cases your validation accuracy can be better than training accuracy, and has better loss. But it's not a good situation. Use separated from training data validation dataset, or use cross-validation technique.
3)80 % accuracy for most real world problems is a very good result.
- Badges
- Science topic
- Similar topics
- Computer Science
- Images