I know we are converting the tensor in scaler than applying backward(), but when to sum and when to mean?
some_loss_function.sum().backward()
-OR-
some_loss_function.mean().backward()
There is no canonical answer to your question. Essentially what you're asking is should I average or sum my loss, as readers we have no knowledge of your problem and what this loss function corresponds to. It all depends on your use case.
Generally though, you would average over summation because you often don't wish the loss value to scale with the dimensionality of the output. Indeed high dimensionality of your output would lead to a higher value to your loss than a summation which is meant to be constant w.r.t. the dimensions of your output tensor. If you sum your loss you will end up scaling your loss value and the gradients that are inferred from it uncontrollably.
Related
I wonder how MAE loss is optimized with SGD optimizer? I mean how the derivative of absolute values sum is calculated. Is there used any numerical solution or something else?
I've found out that in sklearn.linear_model.SGDRegressor MAE loss is a special case of 'epsilon_insensitive' loss with epsilon equal to 0. And according to source code of this loss we simply apply sign(x) function to difference of ground truth and predicted values in order to calculate derivative.
Is there a way to calculate Accuracy instead of Error metrics for neural networks when doing regression (prediction of continuous variable) the same way we do when classifying categorical variables?
Though, the concept of accuracy comes in the classification, but you can print the predicted values and check them with dependent variables.
The problem with continuous variable, is that the probability to reproduce exactly a given value is (practically) zero. For instance if your neural network produces 2.000001 and the actual value is 2, then this would count as a wrong prediction as both values are different (although they are very close). Error metric like the root mean square, measure therefore at the average difference (squared).
However, depending on your application, you could introduce a threshold value ϵ and consider a given output of your neural network as correct if the absolute value of the difference between the observed value and the output is smaller than ϵ and compute the percentage of correct prediction.
In practice such a metric is not minimized directly, because it is difficult to compute its gradient, but it is still a useful quantity to compute.
Can anyone please explain in simple words and possibly with some examples what is a loss function in the field of machine learning/neural networks?
This came out while I was following a Tensorflow tutorial:
https://www.tensorflow.org/get_started/get_started
It describes how far off the result your network produced is from the expected result - it indicates the magnitude of error your model made on its prediciton.
You can then take that error and 'backpropagate' it through your model, adjusting its weights and making it get closer to the truth the next time around.
The loss function is how you're penalizing your output.
The following example is for a supervised setting i.e. when you know the correct result should be. Although loss functions can be applied even in unsupervised settings.
Suppose you have a model that always predicts 1. Just the scalar value 1.
You can have many loss functions applied to this model. L2 is the euclidean distance.
If I pass in some value say 2 and I want my model to learn the x**2 function then the result should be 4 (because 2*2 = 4). If we apply the L2 loss then its computed as ||4 - 1||^2 = 9.
We can also make up our own loss function. We can say the loss function is always 10. So no matter what our model outputs the loss will be constant.
Why do we care about loss functions? Well they determine how poorly the model did and in the context of backpropagation and neural networks. They also determine the gradients from the final layer to be propagated so the model can learn.
As other comments have suggested I think you should start with basic material. Here's a good link to start off with http://neuralnetworksanddeeplearning.com/
Worth to note we can speak of different kind of loss functions:
Regression loss functions and classification loss functions.
Regression loss function describes the difference between the values that a model is predicting and the actual values of the labels.
So the loss function has a meaning on a labeled data when we compare the prediction to the label at a single point of time.
This loss function is often called the error function or the error formula.
Typical error functions we use for regression models are L1 and L2, Huber loss, Quantile loss, log cosh loss.
Note: L1 loss is also know as Mean Absolute Error. L2 Loss is also know as Mean Square Error or Quadratic loss.
Loss functions for classification represent the price paid for inaccuracy of predictions in classification problems (problems of identifying which category a particular observation belongs to).
Name a few: log loss, focal loss, exponential loss, hinge loss, relative entropy loss and other.
Note: While more commonly used in regression, the square loss function can be re-written and utilized for classification.
I am trying to understand how the gradients are computed when using miinibatch SGD. I have implemented it in CS231 online course, but only came to realize that in intermediate layers the gradient is basically the sum over all the gradients computed for each sample (the same for the implementations in Caffe or Tensorflow). It is only in the last layer (the loss) that they are averaged by the number of samples.
Is this correct? if so, does it mean that since in the last layer they are averaged, when doing backprop, all the gradients are also averaged automatically?
Thanks!
It is best to understand why SGD works first.
Normally, what a neural network actually is, a very complex composite function of an input vector x, a label y(or target variable, changes according to whether the problem is classification or regression) and some parameter vector, w. Assume that we are working on classification. We are actually trying to do a maximum likelihood estimation (actually MAP estimation since we are certainly going to use L2 or L1 regularization, but this is too much technicality for now) for variable vector w. Assuming that samples are independent; then we have the following cost function:
p(y1|w,x1)p(y2|w,x2) ... p(yN|w,xN)
Optimizing this wrt to w is a mess due to the fact that all of these probabilities are multiplicated (this will produce an insanely complicated derivative wrt w). We use log probabilities instead (taking log does not change the extreme points and we divide by N, so we can treat our training set as a empirical probability distribution, p(x) )
J(X,Y,w)=-(1/N)(log p(y1|w,x1) + log p(y2|w,x2) + ... + log p(yN|w,xN))
This is the actual cost function we have. What the neural network actually does is to model the probability function p(yi|w,xi). This can be a very complex 1000+ layered ResNet or just a simple perceptron.
Now the derivative for w is simple to state, since we have an addition now:
dJ(X,Y,w)/dw = -(1/N)(dlog p(y1|w,x1)/dw + dlog p(y2|w,x2)/dw + ... + dlog p(yN|w,xN)/dw)
Ideally, the above is the actual gradient. But this batch calculation is not easy to compute. What if we are working on a dataset with 1M training samples? Worse, the training set may be a stream of samples x, which has an infinite size.
The Stochastic part of the SGD comes into play here. Pick m samples with m << N randomly and uniformly from the training set and calculate the derivative by using them:
dJ(X,Y,w)/dw =(approx) dJ'/dw = -(1/m)(dlog p(y1|w,x1)/dw + dlog p(y2|w,x2)/dw + ... + dlog p(ym|w,xm)/dw)
Remember that we had an empirical (or actual in the case of infinite training set) data distribution p(x). The above operation of drawing m samples from p(x) and averaging them actually produces the unbiased estimator, dJ'/dw, for the actual derivative dJ(X,Y,w)/dw. What does that mean? Take many such m samples and calculate different dJ'/dw estimates, average them as well and you get dJ(X,Y,w)/dw very closely, even exactly, in the limit of infinite sampling. It can be shown that these noisy but unbiased gradient estimates will behave like the original gradient in the long run. On the average, SGD will follow the actual gradient's path (but it can get stuck at a different local minima, all depends on the selection of the learning rate). The minibatch size m is directly related to the inherent error in the noisy estimate dJ'/dw. If m is large, you get gradient estimates with low variance, you can use larger learning rates. If m is small or m=1 (online learning), the variance of the estimator dJ'/dw is very high and you should use smaller learning rates, or the algorithm may easily diverge out of control.
Now enough theory, your actual question was
It is only in the last layer (the loss) that they are averaged by the number of samples. Is this correct? if so, does it mean that since in the last layer they are averaged, when doing backprop, all the gradients are also averaged automatically? Thanks!
Yes, it is enough to divide by m in the last layer, since the chain rule will propagate the factor (1/m) to all parameters once the lowermost layer is multiplied by it. You don't need to do separately for each parameter, this will be invalid.
In the last layer they are averaged, and in the previous are summed. The summed gradients in previous layers are summed across different nodes from the next layer, not by the examples. This averaging is done only to make the learning process behave similarly when you change the batch size -- everything should work the same if you sum all the layers, but decrease the learning rate appropriately.
I am using libsvm for my document classification.
I use svm.h and svm.cc only in my project.
Its struct svm_problem requires array of svm_node that are non-zero thus using sparse.
I get a vector of tf-idf words with lets say in range [5,10]. If i normalize it to [0,1], all the 5's would become 0.
Should i remove these zeroes when sending it to svm_train ?
Does removing these would not reduce the information and lead to poor results ?
should i start the normalization from 0.001 rather than 0 ?
Well, in general, in SVM does normalizing in [0,1] not reduces information ?
SVM is not a Naive Bayes, feature's values are not counters, but dimensions in multidimensional real valued space, 0's have exactly the same amount of information as 1's (which also answers your concern regarding removing 0 values - don't do it). There is no reason to ever normalize data to [0.001, 1] for the SVM.
The only issue here is that column-wise normalization is not a good idea for the tf-idf, as it will degenerate yout features to the tf (as for perticular i'th dimension, tf-idf is simply tf value in [0,1] multiplied by a constant idf, normalization will multiply by idf^-1). I would consider one of the alternative preprocessing methods:
normalizing each dimension, so it has mean 0 and variance 1
decorrelation by making x=C^-1/2*x, where C is data covariance matrix