in this  we'll discuss what it  means when a model is said to be  overfitting we'll also cover some  techniques we can use to try to reduce  overfitting when it happens  we briefly mentioned the concept of  overfitting in a previous lecture where we  discussed the purpose of a validation  set let's build more on this concept now  overfitting occurs when our model  becomes really good at being able to  classify or predict on data that was  included in the training set but is not  so good at classifying data that it  wasn't trained on so essentially the  model has over fit the data in the  training set so how do we know if our  models overfitting we can tell that the  models overfitting based on metrics that  are given for our training and  validation data during the training  process we previously saw that when we  specify a validation set during training  we get metrics for the validation  accuracy and loss as well as the  training accuracy and loss if the  validation metrics are considerably  worse than the training metrics then  that's indication that our model is  overfitting we can also get an idea that  our model is overfitting if during  training the models metrics were good  but when we use the model to predict on  test data it's not accurately  classifying the data in the test set the  concept of overfitting boils down to the  fact that the model is unable to  generalize well meaning it's learned the  features of the training set extremely  well but if we give the model any data  that slightly deviates from the exact  data used during training it's unable to  generalize and accurately predict the  output overfitting is an incredibly  common issue 

so how can we reduce it the  easiest thing we can do as long as we  have access to it is to add more data  the more data we have to train our model  on the more it will be able to learn  from the training set also with more  data we're hoping to be adding more  diversity to the training set as well  for example if we train a model to  classify whether an image is an image of  a dog or cat and the model has only seen  images of larger dogs like labs golden  retrievers and boxers that in practice  if it sees a Pomeranian it might not do  so well at recognizing that a Pomeranian  is a dog if we add more data to the  training set to encompass more breeds  then our training data will become more  die  and the model will be less likely to  overfit another technique we can deploy  to reduce overfitting  is to use data augmentation this is the  process of creating additional augmented  data by reasonably modifying the data in  our training set for image data for  example we can do these modifications by  cropping rotating flipping or zooming  we'll cover more on the concept of data  augmentation in a later video the general idea of data  augmentation allows us to add more data  to our training set that's similar to  the data that we already have but it's  just reasonably modified to some degree  so that it's not the exact same for  example if most of our dog images were  dogs facing to the left then it would be  a reasonable modification to add  augmented flipped images so that our  training set would also have dogs that  face to the right now something else  that we can do to reduce overfitting is  to reduce the complexity of our model we  could reduce a complexity by making  simple changes like removing some layers  from the model or reducing the number of  neurons in the layers this may help our  model generalize better to data it  hasn't seen before now the last step  I'll mention for reducing overfitting is  something called dropout the general  idea behind dropout is that if you add  it to a model it will randomly ignore  some subset of nodes in a given layer  during training ie it drops out the  nodes from the layer hence the name  dropout  this will prevent these dropped out  nodes from participating in producing a  prediction on the data this technique  may also help our model to generalize  better to data it hasn't seen before  we'll cover the full concept of drop out  as a regularization technique in another  video and there we'll understand why  this makes sense so hopefully now we  understand the concept of overfitting  and why it happens and also how we can  reduce it if we see it happening in one  of our models 

Dropouts