I'm implementing a neural network for a supervised classification task in MATLAB.
I have a training set and a test set to evaluate the results.
The problem is that every time I train the network for the same training set I get very different results (sometimes I get a 95% classification accuracy and sometimes like 60%) for the same test set.
Now I know this is because I get different initial weights and I know that I can use 'seed' to set the same initial weights but the question is what does this say about my data and what is the right way to look at this? How do I define the accuracy I'm getting using my designed ANN? Is there a protocol for this (like running the ANN 50 times and get an average accuracy or something)?
Thanks
Make sure your test set is large enough compared to the training set (e.g. 10% of the overall data) and check it regarding diversity. If your test set only covers very specific cases, this could be a reason. Also make sure you always use the same test set. Alternatively you should google the term cross-validation.
Furthermore, observing good training set accuracy while observing bad test set accuracy is a sign for overfitting. Try to apply regularization like a simple L2 weight decay (simply multiply your weight matrices with e.g. 0.999 after each weight update). Depending on your data, Dropout or L1 regularization could also help (especially if you have a lot of redundancies in your input data). Also try to choose a smaller network topology (fewer layers and/or fewer neurons per layer).
To speed up training, you could also try alternative learning algorithms like RPROP+, RPROP- or RMSProp instead of plain backpropagation.
Looks like your ANN is not converging to the optimal set of weights. Without further details of the ANN model, I cannot pinpoint the problem, but I would try increasing the number of iterations.
Related
Consider the training process of deep FF neural network using mini-batch gradient descent. As far as I understand, at each epoch of the training we have different random set of mini-batches. Then iterating over all mini batches and computing the gradients of the NN parameters we will get random gradients at each iteration and, therefore, random directions for the model parameters to minimize the cost function. Let's imagine we fixed the hyperparameters of the training algorithm and started the training process again and again, then we would end up with models, which completely differs from each other, because in those trainings the changes of model parameters were different.
1) Is it always the case when we use such random based training algorithms?
2) If it is so, where is the guaranty that training the NN one more time with the best hyperparameters found during the previous trainings and validations will yield us the best model again?
3) Is it possible to find such hyperparameters, which will always yield the best models?
Neural Network are solving a optimization problem, As long as it is computing a gradient in right direction but can be random, it doesn't hurt its objective to generalize over data. It can stuck in some local optima. But there are many good methods like Adam, RMSProp, momentum based etc, by which it can accomplish its objective.
Another reason, when you say mini-batch, there is at least some sample by which it can generalize over those sample, there can be fluctuation in the error rate, and but at least it can give us a local solution.
Even, at each random sampling, these mini-batch have different-2 sample, which helps in generalize well over the complete distribution.
For hyperparameter selection, you need to do tuning and validate result on unseen data, there is no straight forward method to choose these.
I've seen in multiple places that you should disable dropout during validation and testing stages and only keep it during the training phase. Is there a reason why that should happen? I haven't been able to find a good reason for that and was just wondering.
One reason I'm asking is because I trained a model with dropout, and the results turned out well - about 80% accuracy. Then, I went on to validate the model but forgot to set the prob to 1 and the model's accuracy went down to about 70%. Is it supposed to be that drastic? And is it as simple as setting the prob to 1 in each dropout layer?
Thanks in advance!
Dropout is a random process of disabling neurons in a layer with chance p. This will make certain neurons feel they are 'wrong' in each iteration - basically, you are making neurons feel 'wrong' about their output so that they rely less on the outputs of the nodes in the previous layer. This is a method of regularization and reduces overfitting.
However, there are two main reasons you should not use dropout to test data:
Dropout makes neurons output 'wrong' values on purpose
Because you disable neurons randomly, your network will have different outputs every (sequences of) activation. This undermines consistency.
However, you might want to read some more on what validation/testing exactly is:
Training set: a set of examples used for learning: to fit the parameters of the classifier In the MLP case, we would use the training set to find the “optimal” weights with the back-prop rule
Validation set: a set of examples used to tune the parameters of a classifier In the MLP case, we would use the validation set to find the “optimal” number of hidden units or determine a stopping point for the back-propagation algorithm
Test set: a set of examples used only to assess the performance of a fully-trained classifier In the MLP case, we would use the test to estimate the error rate after we have chosen the final model (MLP size and actual weights) After assessing the final model on the test set, YOU MUST NOT tune the model any further!
Why separate test and validation sets? The error rate estimate of the final model on validation data will be biased (smaller than the true error rate) since the validation set is used to select the final model After assessing the final model on the test set, YOU MUST NOT tune the model any further!
source : Introduction to Pattern Analysis,Ricardo Gutierrez-OsunaTexas A&M University, Texas A&M University (answer)
So even for validation, how would you determine which nodes you remove if the nodes have a random probability of being disactivated?
Dropout is a method of making bagging practical for ensembles of very many large neural networks.
Along the same line we may remember that using the following false explanation:
For the new data, we can predict their classes by taking the average of the results from all the learners:
Since N is a constant we can just ignore it and the result remains the same, so we should disable dropout during validation and testing.
The true reason is much more complex. It is because of the weight scaling inference rule:
We can approximate p_{ensemble} by evaluating p(y|x) in one model: the model with all units, but with the weights going out of unit i multiplied by the probability of including unit i. The motivation for this modification is to capture the right expected value of the output from that unit. There is not yet any theoretical argument for the accuracy of this approximate inference rule in deep nonlinear networks, but empirically it performs very well.
When we train the model using dropout(for example for one layer) we zero out some outputs of some neurons and scale the others up by 1/keep_prob to keep the expectation of the layer almost the same as before. In the prediction process, we can use dropout but we can only get different predictions each time because we drop the values out randomly, then we need to run the prediction many times to get the expected output. Such a process is time-consuming so we can remove the dropout and the expectation of the layer remains the same.
Reference:
Difference between Bagging and Boosting?
7.12 of Deep Learning
Simplest reason can be, during prediction(test, validation or after production deployment) you want to use the capability of each and every learned neurons and really don't like to skip some of them randomly.
Thats the only reason we set probability as 1 during testing.
There is a Bayesian technique called Monte Carlo dropout in which the dropout would be not disabled during testing. The model will run several times with the same dropout rate(or in one go as a batch), and the mean(line 6 depicted below) and variance(line 7 depicted below) of the results will be calculated to determine the uncertainty.
Here is Uber's application to quantify uncertainty:
Short answer:
Dropouts to bring down over fitting in the training data. They are used as a regularization parameters. So if you have high variance (i.e. look at the difference between training set and validation set accuracy for this) then use drop out on training data, as it won't be good enough to apply dropout on test and validation data as you haven't been sure about the neurons which are going to shut off hence laying off the importance of random neurons which can be important.
I have a quite simple ANN using Tensorflow and AdamOptimizer for a regression problem and I am now at the point to tune all the hyperparameters.
For now, I saw many different hyperparameters that I have to tune :
Learning rate : initial learning rate, learning rate decay
The AdamOptimizer needs 4 arguments (learning-rate, beta1, beta2, epsilon) so we need to tune them - at least epsilon
batch-size
nb of iterations
Lambda L2-regularization parameter
Number of neurons, number of layers
what kind of activation function for the hidden layers, for the output layer
dropout parameter
I have 2 questions :
1) Do you see any other hyperparameter I might have forgotten ?
2) For now, my tuning is quite "manual" and I am not sure I am not doing everything in a proper way.
Is there a special order to tune the parameters ? E.g learning rate first, then batch size, then ...
I am not sure that all these parameters are independent - in fact, I am quite sure that some of them are not. Which ones are clearly independent and which ones are clearly not independent ? Should we then tune them together ?
Is there any paper or article which talks about properly tuning all the parameters in a special order ?
EDIT :
Here are the graphs I got for different initial learning rates, batch sizes and regularization parameters. The purple curve is completely weird for me... Because the cost decreases like way slowly that the others, but it got stuck at a lower accuracy rate. Is it possible that the model is stuck in a local minimum ?
Accuracy
Cost
For the learning rate, I used the decay :
LR(t) = LRI/sqrt(epoch)
Thanks for your help !
Paul
My general order is:
Batch size, as it will largely affect the training time of future experiments.
Architecture of the network:
Number of neurons in the network
Number of layers
Rest (dropout, L2 reg, etc.)
Dependencies:
I'd assume that the optimal values of
learning rate and batch size
learning rate and number of neurons
number of neurons and number of layers
strongly depend on each other. I am not an expert on that field though.
As for your hyperparameters:
For the Adam optimizer: "Recommended values in the paper are eps = 1e-8, beta1 = 0.9, beta2 = 0.999." (source)
For the learning rate with Adam and RMSProp, I found values around 0.001 to be optimal for most problems.
As an alternative to Adam, you can also use RMSProp, which reduces the memory footprint by up to 33%. See this answer for more details.
You could also tune the initial weight values (see All you need is a good init). Although, the Xavier initializer seems to be a good way to prevent having to tune the weight inits.
I don't tune the number of iterations / epochs as a hyperparameter. I train the net until its validation error converges. However, I give each run a time budget.
Get Tensorboard running. Plot the error there. You'll need to create subdirectories in the path where TB looks for the data to plot. I do that subdir creation in the script. So I change a parameter in the script, give the trial a name there, run it, and plot all the trials in the same chart. You'll very soon get a feel for the most effective settings for your graph and data.
For parameters that are less important you can probably just pick a reasonable value and stick with it.
Like you said, the optimal values of these parameters all depend on each other. The easiest thing to do is to define a reasonable range of values for each hyperparameter. Then randomly sample a parameter from each range and train a model with that setting. Repeat this a bunch of times and then pick the best model. If you are lucky you will be able to analyze which hyperparameter settings worked best and make some conclusions from that.
I don't know any tool specific for tensorflow, but the best strategy is to first start with the basic hyperparameters such as learning rate of 0.01, 0.001, weight_decay of 0.005, 0.0005. And then tune them. Doing it manually will take a lot of time, if you are using caffe, following is the best option that will take the hyperparameters from a set of input values and will give you the best set.
https://github.com/kuz/caffe-with-spearmint
for more information, you can follow this tutorial as well:
http://fastml.com/optimizing-hyperparams-with-hyperopt/
For number of layers, What I suggest you to do is first make smaller network and increase the data, and after you have sufficient data, increase the model complexity.
Before you begin:
Set batch size to maximal (or maximal power of 2) that works on your hardware. Simply increase it until you get a CUDA error (or system RAM usage > 90%).
Set regularizes to low values.
The architecture and exact numbers of neurons and layers - use known architectures as inspirations and adjust them to your specific performance requirements: more layers and neurons -> possibly a stronger, but slower model.
Then, if you want to do it one by one, I would go like this:
Tune learning rate in a wide range.
Tune other parameters of the optimizer.
Tune regularizes (dropout, L2 etc).
Fine tune learning rate - it's the most important hyper-parameter.
I've made digit recognition (56x56 digits) using Neural Networks, but I'm getting 89.5% accuracy on test set and 100% on training set. I know that it's possible to get >95% on test set using this training set. Is there any way to improve my training so I can get better predictions? Changing iterations from 300 to 1000 gave me +0.12% accuracy. I'm also file size limited so increasing number of nodes can be impossible, but if that's the case maybe I could cut some pixels/nodes from the input layer.
To train I'm using:
input layer: 3136 nodes
hidden layer: 220 nodes
labels: 36
regularized cost function with lambda=0.1
fmincg to calculate weights (1000 iterations)
As mentioned in the comments, the easiest and most promising way is to switch to a Convolutional Neural Network. But with you current model you can:
Add more layers with less neurons each, which increases learning capacity and should increase accuracy by a bit. Problem is that you might start overfitting. Use regularization to counter this.
Use batch Normalization (BN). While you are already using regularization, BN accelerates training and also does regularization, and is a NN specific algorithm that might work better.
Make an ensemble. Train several NNs on the same dataset, but with a different initialization. This will produce slightly different classifiers and you can combine their output to get a small increase in accuracy.
Cross-entropy loss. You don't mention what loss function you are using, if its not Cross-entropy, then you should start using it. All the high accuracy classifiers use cross-entropy loss.
Switch to backpropagation and Stochastic Gradient Descent. I do not know the effect of using a different optimization algorithm, but backpropagation might outperform the optimization algorithm you are currently using, and you could combine this with other optimizers such as Adagrad or ADAM.
Other small changes that might increase accuracy are changing the activation functions (like ReLU), shuffle training samples after every epoch, and do data augmentation.
I'm relatively new to Matlab ANN Toolbox. I am training the NN with pattern recognition and target matrix of 3x8670 containing 1s and 0s, using one hidden layer, 40 neurons and the rest with default settings. When I get the simulated output for new set of inputs, then the values are around 0 and 1. I then arrange them in descending order and choose a fixed number(which is known to me) out of 8670 observations to be 1 and rest to be zero.
Every time I run the program, the first row of the simulated output always has close to 100% accuracy and the following rows dont exhibit the same kind of accuracy.
Is there a logical explanation in general? I understand that answering this query conclusively might require the understanding of program and problem, but its made of of several functions to clearly explain. Can I make some changes in the training to get consistence output?
If you have any suggestions please share it with me.
Thanks,
Nishant
Your problem statement is not clear for me. For example, what you mean by: "I then arrange them in descending order and choose a fixed number ..."
As I understand, you did not get appropriate output from your NN as compared to the real target. I mean, your output from NN is difference than target. If so, there are different possibilities which should be considered:
How do you divide training/test/validation sets for training phase? The most division should be assigned to training (around 75%) and rest for test/validation.
How is your training data set? Can it support most scenarios as you expected? If your trained data set is not somewhat similar to your test data sets (e.g., you have some new records/samples in the test data set which had not (near) appear in the training phase, it explains as 'outlier' and NN cannot work efficiently with these types of samples, so you need clustering approach not NN classification approach), your results from NN is out-of-range and NN cannot provide ideal accuracy as you need. NN is good for those data set training, where there is no very difference between training and test data sets. Otherwise, NN is not appropriate.
Sometimes you have an appropriate training data set, but the problem is training itself. In this condition, you need other types of NN, because feed-forward NNs such as MLP cannot work with compacted and not well-separated regions of data very well. You need strong function approximation such as RBF and SVM.