I started off learning about neural networks with the neuralnetworksanddeeplearning dot com tutorial. In particular in the 3rd chapter there is a section about the cross entropy function, and defines the cross entropy loss as:

$C = -\frac{1}{n} \sum\limits_x \sum\limits_j (y_j \ln a^L_j + (1-y_j) \ln (1 - a^L_j))$

However, reading the Tensorflow introduction, the cross entropy loss is defined as:

$C = -\frac{1}{n} \sum\limits_x \sum\limits_j (y_j \ln a^L_j)$ (when using the same symbols as above)

Then searching around to find what was going on I found another set of notes: (https://cs231n.github.io/linear-classify/#softmax-classifier) that uses a completely different definition of the cross entropy loss, albeit this time for an softmax classifier rather than for a neural network.

Can someone explain to me what is going on here? Why are there discrepancies btw. what people define the cross-entropy loss as? Is there just some overarching principle?


2 Answers 2


These three definitions are essentially the same.

1) The Tensorflow introduction, $$C = -\frac{1}{n} \sum\limits_x\sum\limits_{j} (y_j \ln a_j).$$

2) For binary classifications $j=2$, it becomes $$C = -\frac{1}{n} \sum\limits_x (y_1 \ln a_1 + y_2 \ln a_2)$$ and because of the constraints $\sum_ja_j=1$ and $\sum_jy_j=1$, it can be rewritten as $$C = -\frac{1}{n} \sum\limits_x (y_1 \ln a_1 + (1-y_1) \ln (1-a_1))$$ which is the same as in the 3rd chapter.

3) Moreover, if $y$ is a one-hot vector (which is commonly the case for classification labels) with $y_k$ being the only non-zero element, then the cross entropy loss of the corresponding sample is $$C_x=-\sum\limits_{j} (y_j \ln a_j)=-(0+0+...+y_k\ln a_k)=-\ln a_k.$$

In the cs231 notes, the cross entropy loss of one sample is given together with softmax normalization as $$C_x=-\ln(a_k)=-\ln\left(\frac{e^{f_k}}{\sum_je^{f_j}}\right).$$


In the third chapter, equation (63) is the cross entropy applied to multiple sigmoids (that may not sum to 1) while in the Tensoflow intro the cross-entropy is computed on a softmax output layer.

As explained by dontloo both formula are essentially equivalent for two classes but it is not when more than two classes are considered. Softmax makes sense for multiclass with exclusive classes (i.e when there is only one label per sample, that allow the one-hot encoding of labels) while (multiple) sigmoids can be used to describe a multilabel problem (i.e with samples that are possibly positive for several classes).

See this other dontloo answer as well.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.