Archive for October 2016
With sophisticated Neural Networks, you are dealing with a quite complicated nonlinear function. When fitting a high degree polynomial to a few data points, the polynomial can go through all the points, but have such steep slopes that it is useless for predicting points between the training points, we get this same sort of behaviour in Neural Networks. In a way you are training the Neural Network to exactly memorize all the training data exactly rather than figuring out the trends and patterns that you can use to predict other values.
One solution is to perhaps gather more training data, however this may be impossible or quite expensive. It also might be that the training data is missing some representative samples. Here we’ll concentrate on what we can do with the algorithm rather than trying to improve the data.
Interpolation and Extrapolation
Here we refer to generalization as wanting to get answers to data that isn’t in the training data. We refer to overfitting as the case where the model works really well for the training data but doesn’t do nearly as well for anything else.
There are two distinct cases we want to worry about. One is interpolation, this is trying to estimate values where the inputs are surrounded by data in the training set. Extrapolation is the process of trying to predict what happens beyond the training data. Our stock market data is an example of extrapolation. Recognizing handwriting is an example of interpolation (assuming you have a good sample of training data)
Extrapolation tends to be a much harder problem than interpolation, but both a strongly affected by overfitting.
What we often do is divide our training data into three groups. The largest of these we call the training data and use for training. Another is the test data which we run after training to see how well the algorithm works on data that hasn’t been seen by training. To help with detecting overfitting we create a third group which we run after a certain number of steps during training. The following screenshot shows the results for the training and validation sets (this is for a Kaggle competition so the test set needs to be submitted to Kaggle to get the answer). Here smaller values are better. Notice that the training data gets better starting at 3209.5 and going down to 712.8 which indicates training is working. However the validation data starts at 3014.3 goes down to the 1160s and then starts increasing. This indicates we are overfitting the data.
The approach here is really simple, let’s just stop once the validation data starts increasing. So let’s just stop at this point and say we’re done. This is actually a pretty simple and effective way to prevent overfitting. As an added bonus this is a rare technique that leads to faster training.
Penalizing Large Weights
A sign of overfitting is that the slope of our function is high at the points in the training data. Since the slope is approximated by the appropriate weights in our matrix, we would want to keep the weights in our weight matrices low. The way we accomplish this is to add a penalty to the loss function based on the size of the weights.
loss = (tf.nn.l2_loss( tf.sub(logits, tf_train_labels)) + tf.nn.l2_loss(layer1_weights)*beta + tf.nn.l2_loss(layer2_weights)*beta + tf.nn.l2_loss(layer3_weights)*beta + tf.nn.l2_loss(layer4_weights)*beta)
Here we add the sum of the squares of the weights to our loss function. The factor beta is there to let us scale this value to be in the same magnitude as the main loss function. I’ve found that in some problems making the loss due to the weights about equal to the main loss works quite well. In another problem I found choosing beta so that the weights are 10% of the main loss worked quite well.
I have found that combining this with early stopping works quite well. The weight penalty lets us train longer before we start overfitting, which leads to a better overall result.
One property of the Neural Networks in our brain is that brain cells die, but our brain seems to mostly keep on working. In this sense the brain is far more resilient to damage than a computer. The idea behind dropout is to try to add rules to train the Neural Network to be resilient to Neurons being removed. This means the Neural Network can’t be completely reliant on any given Neuron since it could die (be removed from the model).
The way we accomplish this is we add a dropout activation function at some point:
if dropout: hidden = tf.nn.dropout(hidden, 0.5)
This activation function will remove 50% of the neurons at this layer and scale up its outputs by a matching amount. This is so the sum stays the same which means you can use the same weights whether dropout is present or not.
The reason for the if statement is that you only want to do dropout during training and not during validation, testing or production.
You would do this on each hidden layer. It’s rather surprising that the Neural Network still works as well as it does with this much dropout.
I find dropout doesn’t always help, but when it does you can combine it with penalizing the weights and then you can train longer before you need to stop during overfitting. This can sometimes help a network find finer details without overfitting.
When you do dropout, you do have to train for a longer time, so if this is too time prohibitive you might not want to use it.
I think it’s a good sign that Neural Networks can exhibit the same resilience to damage that the brain shows. Perhaps a bit of biological evidence that we are on the correct track.
These are a few techniques you can use to avoid overfitting your model. I generally use all three so I can train a bit longer without overfitting. If you can get more good training data that can also help quite a bit. Using a simpler model (with fewer hidden nodes) can also help with overfitting, but perhaps not provide as good a functional approximation as the more complicated model. As with all things in computer science you are always trading off complexity, overfitting and performance.
We’ve been playing with TensorFlow for a while now and we have a working model for predicting the stock market. I’m not too sure if we’re beating the stocking picking cat yet, but at least we have a good model where we can experiment and learn about Neural Networks. In this article we’re going to look at the optimization methods available in TensorFlow. There are quite a few of these built into the standard toolkit and since TensorFlow is open source you could create your own optimizer. This article follows on from our previous article on optimization and training.
Weaknesses in Gradient Descent
Gradient Descent has worked for us pretty well so far. Basically it calculates the gradients of the loss function (the partial derivatives of loss by each weight) and moves the weights in the direction of lowering the loss function. However finding the minimums of a complicated nonlinear function is a non-trivial exercise and compound this with the fact that a lot of the data we are feeding in during training is very noisy. In our case the stock market historical data is probably quite contradictory and is probably presenting a good challenge to the training algorithm. Here are some weaknesses these other algorithms attempt to address:
- Learning rate. We have one fixed learning rate (how far we move in the direction of the sign of the gradient). We added an optimization to reduce this learning rate as we proceed, but we use the same learning rate for everything at each step. But some parts of our weight matrix may be changing quickly and other parts remaining close to constant. So perhaps use a different learning rate for each weight/bias and vary it by how fast it’s moving and whether it’s moving consistently in the same direction.
- Getting stuck in local minimums or wandering around plateaus. Are we getting stuck in a local minimum which is much worse than the global minimum we would like to find? How can we power past global minimums and continue to the real goal?
The optimizers included with TensorFlow are all variations on Gradient Descent. There are many other optimizers that people use like simulated annealing, conjugate gradient and ant colony optimization but these tend to either not work well with multi-layer Neural Networks or don’t parallelize well to run on GPUs or a distributed network or are far too computationally intensive for large matrices. We added to the code all the optimizers and you just uncomment the one that you want to use.
# optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(loss, global_step=global_step) # optimizer = tf.train.AdadeltaOptimizer(starter_learning_rate).minimize(loss) # optimizer = tf.train.AdagradOptimizer(starter_learning_rate).minimize(loss) # promising # optimizer = tf.train.AdamOptimizer(starter_learning_rate).minimize(loss) # promising # optimizer = tf.train.MomentumOptimizer(starter_learning_rate, 0.001).minimize(loss) # diverges # optimizer = tf.train.FtrlOptimizer(starter_learning_rate).minimize(loss) # promising optimizer = tf.train.RMSPropOptimizer(starter_learning_rate).minimize(loss) # promising
Perhaps it would be less hacky to make this a parameter to the program, but we’ll leave that till we need it.
Let’s quickly summarize what each optimizer tries to accomplish:
- MomentumOptimizer: If gradient descent is navigating down a valley with steep sides, it tends to madly oscillate from one valley wall to the other without making much progress down the valley. This is because the largest gradients point up and down the valley walls whereas the gradient along the floor of the valley is quite small. Momentum Optimization attempts to remedy this by keeping track of the prior gradients and if they keep changing direction then damp them, and if the gradients stay in the same direction then reward them. This way the valley wall gradients get reduced and the valley floor gradient enhanced. Unfortunately this particular optimizer diverges for the stock market data.
- AdagradOptimizer: Adagrad is optimized to finding needles in haystacks and for dealing with large sparse matrices. It keeps track of the previous changes and will amplify the changes for weights that change infrequently and suppress the changes for weights that change frequently. This algorithm seemed promising for the stock market data.
- AdadeltaOptimizer: Adadelta is an extension of Adagrad that only remembers a fixed size window of previous changes. This tends to make the algorithm less aggressive than pure Adagrad. Adadelta seemed to not work as well as Adagrad for the stock market data.
- AdamOptimizer: Adaptive Moment Estimation (Adam) keeps separate learning rates for each weight as well as an exponentially decaying average of previous gradients. This combines elements of Momentum and Adagrad together and is fairly memory efficient since it doesn’t keep a history of anything (just the rolling averages). It is reputed to work well for both sparse matrices and noisy data. Adam seems promising for the stock market data.
- FtrlOptimizer: Ftrl-Proximal was developed for ad-click prediction where they had billions of dimensions and hence huge matrices of weights that were very sparse. The main feature here is to keep near zero weights at zero, so calculations can be skipped and optimized. This algorithm was promising on our stock market data.
- RMSPropOptimizer: RMSprop is similar to Adam it just uses different moving averages but has the same goals.
Neural networks can be quite different and the best algorithm for the job may depend a lot on the data you are trying to train the network with. Each of these optimizers has several tunable parameters. Besides initial learning rate, I’ve left all the others at the default. We could write a meta-trainer that tries to find an optimal solution for which optimizer to use and with which values of its tunable parameters. You would want a quite powerful distributed set of computers to run this on.
Optimization is a tricky subject with Neural Networks, a lot depends on the quality and quantity of your data. It also depends on the size of your model and the contents of the weight matrices. A lot of these optimizers are tuned for rather specific problems like image recognition or ad click-through prediction; however, if you have a unique problem them largely you are left to trial and error (whether automated or manual) to determine the best solution.
Note that a lot of practitioners stick with basic gradient descent since they know it quite well, rather than relying on the newer algorithms. Often massaging your data or altering the random starting point can be a better area to focus on.
We’ve spent some time developing a Neural Network model for predicting the stock market. TensorFlow has produced a fairly black box implementation that is trained by historical data and then can output predictions for tomorrow’s prices.
But what confidence do we have that this model is really doing what we want? Last time we discussed some of the meta-parameters that configure the model. How do we know these are vaguely correct? How do we know if the weights we are training are converging? If we want to step through the model, how do we do that?
TensorFlow comes with a tool called TensorBoard which you can use to get some insight into what is happening. You can’t easily just print variables since they are all internal to the TensorFlow engine and only have values when required as a session is running. There is also the problem with how to visualize the variables. The weights matrix is very large and is constantly changing as you train it, you certainly don’t want to print this out repeatedly, let alone try to read through it.
To use TensorBoard you instrument your program. You tell it what you want to track and assign useful names to those items. This data is then written to log files as your model runs. You then run the TensorBoard program to process these log files and view the results in your Web Browser.
Something Went Wrong
Due to household logistics I moved my TensorFlow work over to my MacBook Air from running in an Ubuntu VM image on our Windows 10 laptop. Installing Python 3, TensorFlow and the various other libraries I’m using was quite simple and straight forward. Just install Python from Python.org and then use pip3 to install any other libraries. That all worked fine. But when I started running the program from last time, I was getting NaN results quite often. I wondered if TensorFlow wasn’t working right on my Mac? Anyway I went to debug the program and that led me to TensorBoard. As it turns out there was quite a bad bug in the program presented last time due to un-initialized variables.
You tend to get complacent programming in Python about un-initialized variables (and array subscript range errors) because usually Python will raise and exception if you try to use a variable that hasn’t been initialized. The problem is NumPy which is a library written in C for efficiency. When you create a NumPy array, it is returned to Python, telling Python its good to go. But since its managed by C code you don’t get the usual Python error checking. So when I changed the program to add the volumes to the price changes, I had a bug that left some of the data arrays uninitialized. I suspect on the Windows 10 laptop that these were initialized to zero, but that all depends on which exact C runtime is being used. On the Mac these values were just random memory and that immediately led to program errors.
Adding the TensorBoard initialization showed the problem was originating with the data and then it was fairly straight forward to zero in on the problem and fix it.
As a result, for this article, I’m just going to overwrite the Python file from last time with a newer one (tfstocksdiff2.py) which is posted here. This version includes TensorBoard instrumentation and a couple of other improvements that I’ll talk about next time.
First we’ll start with some of the things that TensorBoard shows you. If you read an overview of TensorFlow it’s a bit confusing about what are Tensors and what flows. If you’ve looked at the program so far, it shows quite a few algebraic matrix equations, but where are the Tensors? What TensorFlow does is break these equations down into nodes where each node is a function execution and the data flows along the edges. This is a fairly common way to evaluate algebraic expressions and not unique to TensorFlow. TensorFlow then supports executing these on GPUs and in distributed environments as well as providing all the node types you need to create Neural Networks. TensorBoard gives you a way to visualize these graphs. The names of the nodes are from the program instrumentation.
When the program was instrumented it grouped things together. Here is an expansion of the trainingmodel box where you can see the operations that make up our model.
This gives us some confidence that we have constructed our TensorFlow graph correctly, but doesn’t show any data.
We can track various statistics of all our TensorFlow variables over time. This graph is showing a track of the means of the various weight and bias matrixes.
TensorBoard also lets us look at the distribution of the matrix values over time.
TensorBoard also lets us look at histograms of the data and how those histograms evolve over time.
You can see how the layer 1 weights start as their seeded normal distribution of random numbers and then progress to their new values as training progresses. If you look at all these graphs you can see that the values are still progressing when training stops. This is because TensorBoard instrumentation really slows down processing, so I shortened the training steps while using TensorBoard. I could let it run much longer over night to ensure that I am providing sufficient training for all the values to settle down.
Rather than include all the code here, check out the Google Drive for the Python source file. But quickly we added a function to get all the statistics on a variable:
def variable_summaries(var, name): """Attach a lot of summaries to a Tensor.""" with tf.name_scope('summaries'): mean = tf.reduce_mean(var) tf.scalar_summary('mean/' + name, mean) with tf.name_scope('stddev'): stddev = tf.sqrt(tf.reduce_mean(tf.square(var - mean))) tf.scalar_summary('stddev/' + name, stddev) tf.scalar_summary('max/' + name, tf.reduce_max(var)) tf.scalar_summary('min/' + name, tf.reduce_min(var)) tf.histogram_summary(name, var)
We define names in the various section and indicate the data we want to collect:
with tf.name_scope('Layer1'): with tf.name_scope('weights'): layer1_weights = tf.Variable(tf.truncated_normal( [NHistData * num_stocks * 2, num_hidden], stddev=0.1)) variable_summaries(layer1_weights, 'Layer1' + '/weights') with tf.name_scope('biases'): layer1_biases = tf.Variable(tf.zeros([num_hidden])) variable_summaries(layer1_biases, 'Layer1' + '/biases')
Before the call to initialize_all_variables we need to call:
merged = tf.merge_all_summaries() test_writer = tf.train.SummaryWriter('/tmp/tf/test', session.graph )
And then during training:
summary, _, l, predictions = session.run( [merged, optimizer, loss, train_prediction], feed_dict=feed_dict)
TensorBoard is quite a good tool to give you insight into what is going on in your model. Whether the program is correctly doing what you think and whether there is any sanity to the data. It also lets you tune the various parameters to ensure you are getting best results.