Suppose I split my data into training set and validation set. I perform a 5-fold cross-validation on my training set to obtain the optimal hyper-parameters for my model, then I use the optimal hyper-parameters to train my model and apply the resulting model on my validation set. My question is, is it reasonable to combine the training and validation set, and use the hyper-parameters obtained from the training set to build a final model?
It is resonable if training data was relatively small and adding validation set makes your model significantly stronger. However, at the same time, adding new data makes your previously selected hyperparameters possibly suboptimal (it is really hard to show what kind of transformation of hyperparameters you should apply when you add new data to your training set). Thus you balance two things - gain in model quality from more data and possible loss due to hard to predict change in hyperparameters meaning. To some extent you can simulate this process to make sure it makes sense, if you have N points in training data and M in validation, you can try to split training further to chunks with the same proportion (thus one is now N * (N/(N+M) and other N * (M/(N+M))), train on first one and check whether optimal hyperparameters transfer (more or less) to the optimal one on the whole training set - if so, you can safely add validation as they should transfer as well. If they do not - the risk might be not worth the gain.
Related
As I learned about cross-validation algorithm, from most of the articles on the web, there are variety of cross-validation methods. Here I want to be clear about the k-fold cross-validation technique.
In the k-fold cross-validation algorithm, we can split the training set in to k not-overlapped folds.
As we split the training data in to k folds, we have to train the model in k iterations.
So, in each iteration, we train the model with (k-1) folds and validate it with the remained fold.
In each split we can calculate the desired metric(s) of our model.
At the end we can report the training error by taking the average of scores of all iterations.
But what is the final trained model?
Some points in those articles are not clear for me?
Should I initiate model's parameters in each iteration?
I ask this, because if I don’t initialize the parameter's it could save the pattern of data which I want to be unseen in the next iteration and so on…
Should I save the initial parameter of the split in which I gained the best score, as the best initial values of the parameters?
Should I retrain the model initiating it with the initial values of the parameters gained in my second question and then feed it with whole training dataset and gain the final trained model?
Alright so before answering your question I will go a bit back to explain the purpose of cross validation and model evaluation. You can read these slides or research more about statistical learning theory if you want to go deeper.
Train/test split
Suppose you have a model with defined hyperparameter (or none) and you train it on the training split. If you calculate the metrics over the test split, this will give you the risk of the model on new data. Then you know that this particular model will perform like that on unseen data.
So we have a learning process B, that takes a dataset S (here the training dataset) as well as hyperparameters h, and gives a fitted model m; then B(S, h)->m (training B on S with hp h gives a model m, with its parameters). Then we tested this model to evaluate the risk R on the test dataset.
k-fold Cross validation
When doing k-fold cross validation, you fit k models using the learning process B. Each model is fitted on a different training set, and the risk is computed on non overlapping samples.
Then, you calculate the mean risk among the folds. A common mistake is that it gives you the performance of the model, that's not true. This gives you the mean (or expected) performances of the learning process B (and hyperparams h). That means, if you train a new model using B (and hyperparams h), its expected performance will be around the calculated metrics (of course this is not always true).
For your questions
Yes you should train the model from scratch, if possible with the same initial parameters (if initialization is not random) to avoid any difference between folds. Using a warm start with the previous parameters can modify the learning process, and the fitting.
No, if initialization is random let it be, if it is fixed use the same initial parameters for all folds
For the two previous questions, if by initial parameters you meant hyperparameters, then you should keep the same for all folds, otherwise the calculated risk will be useless. If you want to try multiple hyperparameters, you have to repeat the cross validation multiple times, and then you can select the best ones based on the risk calculated.
Once you tuned your hyperparameters you can train the model on your whole training set. This will give you a model m. Before your cross validation you can keep a small test split to evaluate this final model on unseen data
Let's say I initially split my dataset into training (80%) and test (20%) sets, perform a 10-fold CV on my training set and obtain an average R² of 75%. After that I check the best model's accuracy on the test set and obtain an R² of 74%, which indicates that the model is fairly robust. Now, before deploying it to real applications, I tune it with the whole data. Someone asks me the model's approximate R²; if I say 74% or 75%, I will me ignoring the fact that the model was now tunned with more data (test set). Is it a resonable approach to perform a leave one out CV on the chosen model with the whole data, compare the predicted targets with the real ones, check the R² (let's say it's 80% now) and say that the real-world model will most likely have an R² of 80%? I see no problems with that, but I do not know if this approach is correct.
It is true that you should train again on the whole data and it might lead to performance improvements. However, in this context your whole data should not be train + test! It should be just the trainign dataset but without any cross-validation. So before you had %80 for training and you were doing 10 fold CV, meaning that you were training your model actually on %72 of your complete data(train+test) and keeping the %8 for the validation. Now you should train it on the whole %80 percent and report your final results again on the unseen test set.
If you do LOOCV on the train + test, you can not report your performance on validation samples because this is how the model is finetuned and you might as well overfit to validation data.
What is the difference between the Holdout dataset and the Validation dataset in Machine Learning context?
The Validation dataset is used during training to track the performance of your model on "unseen" data. I wrote the unseen in quotes because although the model doesn't directly see the data in validation set, you will optimize the hyper-parameters to decrease the loss on validation set (since increasing val loss will mean over-fitting). However, by doing so, you may over-fit the hyper-parameters to validation set (So that the loss will be low on that specific validation set, but will become worse on any other unseen set). That's why you usually keep another 3rd set, called test set (or held-out set), which will be your truly unseen data, and you will test the performance of your model on that test set only once, after training your final model.
For a class project, I designed a neural network to approximate sin(x), but ended up with a NN that just memorized my function over the data points I gave it. My NN took in x-values with a batch size of 200. Each x-value was multiplied by 200 different weights, mapping to 200 different neurons in my first layer. My first hidden layer contained 200 neurons, each one a linear combination of the x-values in the batch. My second hidden layer also contained 200 neurons, and my loss function was computed between the 200 neurons in my second layer and the 200 values of sin(x) that the input mapped to.
The problem is, my NN perfectly "approximated" sin(x) with 0 loss, but I know it wouldn't generalize to other data points.
What did I do wrong in designing this neural network, and how can I avoid memorization and instead design my NN's to "learn" about the patterns in my data?
It is same with any machine learning algorithm. You have a dataset based on which you try to learn "the" function f(x), which actually generated the data. In real life datasets, it is impossible to get the original function from the data, and therefore we approximate it using something g(x).
The main goal of any machine learning algorithm is to predict unseen data as best as possible using the function g(x).
Given a dataset D you can always train a model, which will perfectly classify all the datapoints (you can use a hashmap to get 0 error on the train set), but which is overfitting or memorization.
To avoid such things, you yourself have to make sure that the model does not memorise and learns the function. There are a few things which can be done. I am trying to write them down in an informal way (with links).
Train, Validation, Test
If you have large enough dataset, use Train, Validation, Test splits. Split the dataset in three parts. Typically 60%, 20% and 20% for Training, Validation and Test, respectively. (These numbers can vary based on need, also in case of imbalanced data, check how to get stratified partitions which preserve the class ratios in every split). Next, forget about the Test partition, keep it somewhere safe, don't touch it. Your model, will be trained using the Training partition. Once you have trained the model, evaluate the performance of the model using the Validation set. Then select another set of hyper-parameter configuration for your model (eg. number of hidden layer, learaning algorithm, other parameters etc.) and then train the model again, and evaluate based on Validation set. Keep on doing this for several such models. Then select the model, which got you the best validation score.
The role of validation set here is to check what the model has learned. If the model has overfit, then the validation scores will be very bad, and therefore in the above process you will discard those overfit models. But keep in mind, although you did not use the Validation set to train the model, directly, but the Validation set was used indirectly to select the model.
Once you have selected a final model based on Validation set. Now take out your Test set, as if you just got new dataset from real life, which no one has ever seen. The prediction of the model on this Test set will be an indication how well your model has "learned" as it is now trying to predict datapoints which it has never seen (directly or indirectly).
It is key to not go back and tune your model based on the Test score. This is because once you do this, the Test set will start contributing to your mode.
Crossvalidation and bootstrap sampling
On the other hand, if your dataset is small. You can use bootstrap sampling, or k-fold cross-validation. These ideas are similar. For example, for k-fold cross-validation, if k=5, then you split the dataset in 5 parts (also be carefull about stratified sampling). Let's name the parts a,b,c,d,e. Use the partitions [a,b,c,d] to train and get the prediction scores on [e] only. Next, use the partitions [a,b,c,e] and use the prediction scores on [d] only, and continue 5 times, where each time, you keep one partition alone and train the model with the other 4. After this, take an average of these scores. This is indicative of that your model might perform if it sees new data. It is also a good practice to do this multiple times and perform an average. For example, for smaller datasets, perform a 10 time 10-folds cross-validation, which will give a pretty stable score (depending on the dataset) which will be indicative of the prediction performance.
Bootstrap sampling is similar, but you need to sample the same number of datapoints (depends) with replacement from the dataset and use this sample to train. This set will have some datapoints repeated (as it was a sample with replacement). Then use the missing datapoins from the training dataset to evaluate the model. Perform this multiple times and average the performance.
Others
Other ways are to incorporate regularisation techniques in the classifier cost function itself. For example in Support Vector Machines, the cost function enforces conditions such that the decision boundary maintains a "margin" or a gap between two class regions. In neural networks one can also do similar things (although it is not same as in SVM).
In neural network you can use early stopping to stop the training. What this does, is train on the Train dataset, but at each epoch, it evaluates the performance on the Validation dataset. If the model starts to overfit from a specific epoch, then the error for Training dataset will keep on decreasing, but the error of the Validation dataset will start increasing, indicating that your model is overfitting. Based on this one can stop training.
A large dataset from real world tends not to overfit too much (citation needed). Also, if you have too many parameters in your model (to many hidden units and layers), and if the model is unnecessarily complex, it will tend to overfit. A model with lesser pameter will never overfit (though can underfit, if parameters are too low).
In the case of you sin function task, the neural net has to overfit, as it is ... the sin function. These tests can really help debug and experiment with your code.
Another important note, if you try to do a Train, Validation, Test, or k-fold crossvalidation on the data generated by the sin function dataset, then splitting it in the "usual" way will not work as in this case we are dealing with a time-series, and for those cases, one can use techniques mentioned here
First of all, I think it's a great project to approximate sin(x). It would be great if you could share the snippet or some additional details so that we could pin point the exact problem.
However, I think that the problem is that you are overfitting the data hence you are not able to generalize well to other data points.
Few tricks that might work,
Get more training points
Go for regularization
Add a test set so that you know whether you are overfitting or not.
Keep in mind that 0 loss or 100% accuracy is mostly not good on training set.
When developing a neural net one typically partitions training data into Train, Test, and Holdout datasets (many people call these Train, Validation, and Test respectively. Same things, different names). Many people advise selecting hyperparameters based on performance in the Test dataset. My question is: why? Why not maximize performance of hyperparameters in the Train dataset, and stop training the hyperparameters when we detect overfitting via a drop in performance in the Test dataset? Since Train is typically larger than Test, would this not produce better results compared to training hyperparameters on the Test dataset?
UPDATE July 6 2016
Terminology change, to match comment below. Datasets are now termed Train, Validation, and Test in this post. I do not use the Test dataset for training. I am using a GA to optimize hyperparameters. At each iteration of the outer GA training process, the GA chooses a new hyperparameter set, trains on the Train dataset, and evaluates on the Validation and Test datasets. The GA adjusts the hyperparameters to maximize accuracy in the Train dataset. Network training within an iteration stops when network overfitting is detected (in the Validation dataset), and the outer GA training process stops when overfitting of the hyperparameters is detected (again in Validation). The result is hyperparameters psuedo-optimized for the Train dataset. The question is: why do many sources (e.g. https://www.cs.toronto.edu/~hinton/absps/JMLRdropout.pdf, Section B.1) recommend optimizing the hyperparameters on the Validation set, rather than the Train set? Quoting from Srivasta, Hinton, et al (link above): "Hyperparameters were tuned on the validation set such that the best validation error was produced..."
The reason is that developing a model always involves tuning its configuration: for example, choosing the number of layers or the size of the layers (called the hyper-parameters of the model, to distinguish them from the parameters, which are the network’s weights). You do this tuning by using as a feedback signal the performance of the model on the validation data. In essence, this tuning is a form of learning: a search for a good configuration in some parameter space. As a result, tuning the configuration of the model based on its performance on the validation set can quickly result in overfitting to the validation set, even though your model is never directly trained on it.
Central to this phenomenon is the notion of information leaks. Every time you tune a hyperparameter of your model based on the model’s performance on the validation set, some information about the validation data leaks into the model. If you do this only once, for one parameter, then very few bits of information will leak, and your validation set will remain reliable to evaluate the model. But if you repeat this many times—running one experiment, evaluating on the validation set, and modifying your model as a result—then you’ll leak an increasingly significant amount of information about the validation set into the model.
At the end of the day, you’ll end up with a model that performs artificially well on the validation data, because that’s what you optimized it for. You care about performance on completely new data, not the validation data, so you need to use a completely different, never-before-seen dataset to evaluate the model: the test dataset. Your model shouldn’t have had access to any information about the test set, even indirectly. If anything about the model has been tuned based on test set performance, then your measure of generalization will be flawed.
There are two things you are missing here. First, minor, is that test set is never used to do any training. This is a purpose of validation (test is just to asses your final, testing performance). The major missunderstanding is what it means "to use validation set to fit hyperparameters". This means exactly what you describe - to train a model with a given hyperparameters on the training set, and use validation to simply check if you are overfitting (you use it to estimate generalization) , but you do not really "train" on them, you simply check your scores on this subset (which, as you noticed - is way smaller).
You cannot "stop training hyperparamters" because this is not a continuous process, usually hyperparameters are just "possible sets of values", and you have to simply test lots of them, there is no valid way of defining a direct trainingn procedure between actual metric you are interested in (like accuracy) and hyperparameters (like size of the hidden layer in NN or even C parameter in SVM), as the functional link between these two is not differentiable, is highly non convex and in general "ugly" to optimize. If you can define a nice optimization procedure in terms of a hyperparameter than it is usually not called a hyperparameter but a parameter, the crucial distinction in this naming convention is what makes it hard to optimize directly - we call hyperparameter a parameter, than cannot be directly optimized against thus you need a "meta method" (like simply testing on validation set) to select it.
However, you can define a "nice" meta optimization protocol for hyperparameters, but this will still use validation set as an estimator, for example Bayesian optimization of hyperparameters does exactly this - it tries to fit a function saying how well is you model behaving in the space of hyperparameters, but in order to have any "training data" for this meta-method, you need validation set to estimate it for any given set of hyperparameters (input to your meta method)
simple answer: we do
In the case of a simple feedforward neural network you do have to select e.g. layer and unit count per layer, regularization (and non-continuous parameters like topology if not feedforward and loss function) in the beginning and you would optimize on those.
So, in summary you optimize:
ordinary parameters only during training but not during validation
hyperparameters during training and during validation
It is very important not to touch the many ordinary parameters (weights and biases) during validation. That's because there are thousands of degrees of freedom in them which means they can learn the data you train them on. But then the model doesn't generalize to new data as well (even when that new data originated from the same distribution). You usually only have very few degrees of freedom in the hyperparameters which usually control the rigidity of the model (regularization).
This holds true for other machine learning algorithms like decision trees, forests, etc as well.