# How meaningful is the connection between MLE and cross entropy in deep learning?

I understand that given a set of $m$ independent observations $\mathbb{O}=\{\mathbf{o}^{(1)}, . . . , \mathbf{o}^{(m)}\}$ the Maximum Likelihood Estimator (or, equivalently, the MAP with flat/uniform prior) that identifies the parameters $\mathbf{θ}$ that produce the model distribution $p_{model}\left(\,\cdot\, ; \mathbf{θ}\right)$ that best matches those observations will be

$$\mathbf{θ}_{ML}(\mathbb{O})= p_{model}\left(\mathbb{O}; \mathbf{θ}\right) = \underset{\mathbf{θ}}{\arg\max}‎‎\prod_{i=1}^{m} p_{model}\left(\mathbf{o}^{(i)}; \mathbf{θ}\right)$$

or, more conveniently

$$\mathbf{θ}_{ML}(\mathbb{O})= \underset{\mathbf{θ}}{\arg\min}\sum_{i=1}^{m} -\log p_{model}\left(\mathbf{o}^{(i)}; \mathbf{θ}\right)$$

and see the role that $\mathbf{θ}_{ML}$ can play in defining a loss function for multi-class deep neural networks, in which $\mathbf{θ}$ corresponds to the the network's trainable parameters (e.g., $\mathbf{θ} = \{\mathbf{W}, \mathbf{b}\} )$ and the observations are the pairs of input activations $\mathbf{x}$ and corresponding correct class labels $y \in [1, k]$, $\mathbf{o}^{(i)}$ = {$\mathbf{x}^{(i)}, y^{(i)}$}, by taking

$$p_{model}\left(\mathbf{o}^{(i)}; \mathbf{θ}\right) \equiv p_{model}\left(y^{(i)} | \mathbf{x}^{(i)}; \mathbf{θ}\right)$$

What I don't understand is how this relates to the so called "cross entropy" of the (vectorized) correct output, $\mathbf{y}^{(i)}$, and the corresponding output activations of the network, $\mathbf{a}(\mathbf{x}^{(i)}; \mathbf{θ})$ $$H(\mathbf{o}^{(i)}; \mathbf{θ}) = -\mathbf{y}^{(i)}\cdot \mathbf{log}\,\mathbf{a}(\mathbf{x}^{(i)}; \mathbf{θ})‎$$ that is used in practice when measuring error/loss during training. There are several related issues:

### Activations "as probabilities"

One of the steps in establishing the relationship between MLE and cross entropy is to use the output activations "as if" they are probabilities. But it's not clear to me that they are, or at least that they $all$ are.

In calculating training error — specifically, in calling it a "cross entropy loss" — it is assumed that (after normalizing activations to sum to 1)

$$p_{model}\left(\mathbf{o}^{(i)}; \mathbf{θ}\right) \equiv a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ})\tag{1}\label{1}‎‎$$

or

$$\log p_{model}\left(\mathbf{o}^{(i)}; \mathbf{θ}\right) = \log a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ})‎‎$$

so that we can write

$$-\log p_{model}\left(\mathbf{o}^{(i)}; \mathbf{θ}\right) = -\mathbf{y}^{(i)}\cdot \mathbf{log}\,\mathbf{a}(\mathbf{x}^{(i)}; \mathbf{θ})‎\tag{3}\label{3}$$

and thus

$$\mathbf{θ}_{ML}(\mathbb{O})=\underset{\mathbf{θ}}{\arg\min}\sum_{i=1}^{m} H(\mathbf{o}^{(i)}; \mathbf{θ})$$

But while this certainly makes $a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$ a probability (to the extent that anything is), it places no restrictions on the other activations.

Can the $\mathbf{a}_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$ really be said to be PMFs in that case? Is there anything that makes the $a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$ not in fact probabilities (and merely "like" them)?

### Limitation to categorization

The crucial step above in equating MLE with cross-entropy relies entirely on the "one-hot" structure of $\mathbf{y}^{(i)}$ that characterizes a (single-label) multi-class learning problem. Any other structure for the $\mathbf{y}^{(i)}$ would make it impossible to get from $\eqref{1}$ to $\eqref{3}$.

Is the equation of MLE and cross-entropy minimization limited to cases where the $\mathbf{y}^{(i)}$ are "one-hot"?

### Different training and prediction probabilities

During prediction, it is almost always the case that

$$p_{model}\left(y^{(i)} | \mathbf{x}^{(i)}; \mathbf{θ}\right) \equiv P\left(\underset{j\in[1,k]}{\arg\max}\,a_j(\mathbf{x}^{(i)}; \mathbf{θ}) = y^{(i)}\right)\tag{2}\label{2}$$

which results in correct prediction probabilities that are different from the probabilities learned during training unless it is reliably the case that

$$a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML}) = P\left(\underset{j\in[1,k]}{\arg\max}\,a_j(\mathbf{x}^{(i)}; \mathbf{θ}_{ML}) = y^{(i)}\right)$$

Is this ever reliably the case? Is it likely at least approximately true? Or is there some other argument that justifies this equation of the value of the learned activation at the label position with the probability that the maximum value of learned activations occurs there?

### Entropy and information theory

Even assuming that the above concerns are addressed and the activations are valid PMFs (or can meaningfully be treated as such), so that the role played by cross entropy in computing $\mathbf{θ}_{ML}$ is unproblematic, it's not clear to me why it is helpful or meaningful to talk about the entropy of the $\mathbf{a}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$, since Shanon entropy applies to a specific kind of encoding, which is not the one being used in training the network.

What role does information theoretic entropy play in interpreting the cost function, as opposed to simply providing a tool (in the form of cross entropy) for computing one (that corresponds to MLE)?

## 4 Answers

Neural nets don't necessarily give probabilities as outputs, but they can be designed to do this. To be interpreted as probabilities, a set of values must be nonnegative and sum to one. Designing a network to output probabilities typically amounts to choosing an output layer that imposes these constraints. For example, in a classification problem with $k$ classes, a common choice is a softmax output layer with $k$ units. The softmax function forces the outputs to be nonnegative and sum to one. The $j$th output unit gives the probability that the class is $j$. For binary classification problems, another popular choice is to use a single output unit with logistic activation function. The output of the logistic function is between zero and one, and gives the probability that the class is 1. The probability that the class is 0 is implicitly one minus this value. If the network contains no hidden layers, then these two examples are equivalent to multinomial logistic regression and logistic regression, respectively.

Cross entropy $H(p, q)$ measures the difference between two probability distributions $p$ and $q$. When cross entropy is used as a loss function for discriminative classifiers, $p$ and $q$ are distributions over class labels, given the input (i.e. a particular data point). $p$ is the 'true' distribution and $q$ is the distribution predicted by the model. In typical classification problems, each input in the dataset is associated with an integer label representing the true class. In this case, we use the empirical distribution for $p$. This simply assigns probability 1 to the true class of a data point, and probability 0 to all other classes. $q$ is the distribution of class probabilities predicted by the network (e.g. as described above).

Say the data are i.i.d., $p_i$ is the empirical distribution, and $q_i$ is the predicted distribution (for the $i$th data point). Then, minimizing the cross entropy loss (i.e. $H(p_i, q_i)$ averaged over data points) is equivalent to maximizing the likelihood of the data. The proof is relatively straightforward. The basic idea is to show that the cross entropy loss is proportional to a sum of negative log predicted probabilities of the data points. This falls out neatly because of the form of the empirical distribution.

Cross entropy loss can also be applied more generally. For example, in 'soft classification' problems, we're given distributions over class labels rather than hard class labels (so we don't use the empirical distribution). I describe how to use cross entropy loss in that case here.

To address some other specifics in your question:

Different training and prediction probabilities

It looks like you're finding the output unit with maximum activation and comparing this to the class label. This isn't done for training using the cross entropy loss. Instead, the probabilities output by the model are compared to the 'true' probabilities (typically taken to be the empirical distribution).

Shanon entropy applies to a specific kind of encoding, which is not the one being used in training the network.

Cross entropy $H(p,q)$ can be interpreted as the number of bits per message needed (on average) to encode events drawn from true distribution $p$, if using an optimal code for distribution $q$. Cross entropy takes a minimum value of $H(p)$ (the Shannon entropy of $p$) when $q = p$. The better the match between $q$ and $p$, the shorter the message length. Training a model to minimize the cross entropy can be seen as training it to better approximate the true distribution. In supervised learning problems like we've been discussing, the model gives a probability distribution over possible outputs, given the input. Explicitly finding optimal codes for the distribution isn't part of the process.

• "This isn't done for training using the cross entropy loss." This is exactly what APIs like TensorFlow's softmax_cross_entropy_with_logits do: they calculate $\underset{\mathbf{θ}}{\arg\min}\sum_{i=1}^{m} H(\mathbf{o}^{(i)}; \mathbf{θ})$ and thus $\mathbf{θ}_{ML}(\mathbb{O})$ which defines a network "designed to" produce probabilities (at least at the label location). No? – orome Aug 14 '17 at 12:43
• Yes, cross entropy is minimized and likelihood is maximized (at least locally). In that sentence, I was referring to the equations in the section "different training and prediction probabilities". Looking at it again, it's not clear to me exactly what you mean with those equations, so I'll just say this: $p_{model}(y^{(i)} = j \mid x^{(i)}; \theta) = a_j(x^{(i)}; \theta)$ if you're using an output layer where each unit gives a class probability (e.g. softmax). The model probabilities are the same during training and prediction. – user20160 Aug 14 '17 at 22:02
• I understand that the same values are used — that is, the learned $\mathbf{a}$ are used in prediction — but they're used in different ways. The probability that the model learns for $p_{model}\left(y^{(i)} | \mathbf{x}^{(i)}; \mathbf{θ}_{ML}\right)$ is indeed $a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$, but the probability that $y^{(i)}$ will be predicted by the trained model in response to the same input, $\mathbf{x}^{(i)}$ is $P\left({\arg\max}_{j\in[1,k]}\,a_j(\mathbf{x}^{(i)}; \mathbf{θ}_{ML}) = y^{(i)}\right)$. These are not the same unless (2) is true. – orome Aug 15 '17 at 14:15
• And (first question) I understand that because because of the role, defined in eq. (1), played by the $a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ})$ in maximizing $p_{model}\left(\mathbb{O}; \mathbf{θ}\right)$, the $a_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$ values are probabilities (not because of softmax, which only assures they will add to 1). But that places no constraints on the other $a_j$; $j\neq y^{(i)}$ (other than they sum to $1-a_{y^{(i)}}$). So I don't see how $\mathbf{a}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$ as a hole can be considered a PMF. – orome Aug 15 '17 at 14:15
• Another way of making the point of the first question is that only the $a_{y^{(i)}}$ ever participate in the ML process, and thus only they can be considered probabilities. And while a suitable activation function (e.g. softmax) ensures that the sum of the remaining activations will be a probability, the relationships among any of them has no meaning. – orome Aug 17 '17 at 15:25

I'll answer from a slightly more general perspective, concerning the nature of how, when, and why we can consider NN outputs to be probability distributions.

In the sense that the softmax enforces the outputs to sum to 1 and also be non-negative, the output of the network is a discrete probability distribution over the classes, or at least can be interpreted as such. Hence it is perfectly reasonable to talk about cross-entropies and maximum likelihoods.

However, what I think you are seeing (and it is correct), is that the output "probabilities" may have nothing to do with the actual probability of correctness. This is a well-known problem in ML, called calibration. For instance, if your classifier $f_\theta$ of dogs $D$ and cats $C$ says $f_\theta(x_i,C) = P(x_i = C|\theta) = 0.7$, then you would expect that if you took a set of examples $S=\{x_j\}$ all of which had $P(x_j = C|\theta) = 0.7$, then roughly 30% of the inputs would be misclassified (since it was only 70% confident).

However, it turns out that modern training methods do not enforce this at all! See Guo et al, On the Calibration of Modern Neural Networks to see some discussion of this.

In other words, the "probability" of the output from the softmax may well have nothing to do with the actual model confidence. And this is no surprise: we merely want to maximize our accuracy, and every input example has a probability of 1 of being its target class. There is little incentivizing the model to get this right. If it doesn't need to estimate uncertainty then why should it? Cross-entropy does not rectify this issue; indeed, you are telling it to go to a delta function every time!

Lots of recent work on Bayesian neural networks strive to rectify this issue. Such models employ a distribution over parameters given the data $P(\theta|X) = P(X|\theta)P(\theta)/P(X)$, which can be integrated to get an actual probability distribution $P(y_i|x_i,X)=\int P(y_i|\theta,x_i) P(\theta|X) \,d\theta$. This helps guarantee useful uncertainty measurements and better calibration. However, it is more problematic computationally.

Hopefully I didn't misunderstand your question!

Feed-forward neural networks approximate the true class probabilities when trained properly.

In 1991, Richard & Lippmann proved that feed-forward neural networks approach posterior class probabilities, when trained with {0,1} class-indicator target patterns [Richard M. D., & Lippmann R. P. (1991). Neural network classifiers estimate bayesian a posteriori probabilities. Neural Computation, 3, 461– 483.]. In their line of proof, they use one-hidden layer feed-forward neural networks.

In the mathematical annotation of Duda & Hart [Duda R.O. & Hart P.E. (1973) Pattern Classification and Scene Analysis, Wiley], define the feature distributions provided as input vector to the feed-forward neural network as $P({\bf{\it x}}\,\mid\,\omega_i)$, where for example the data vector equals ${\bf{\it x}}=(0.2,10.2,0,2)$, for a classification task with 4 feature-variables. The index $i$ indicates the possible $n$ classes, $i \in \{1,\ldots,n\}$.

The feed-forward neural network classifier learns the posterior probabilities, ${\hat P}(\omega_i\,\mid\,{\bf{\it x}})$, when trained by gradient descent. The desired output pattern needs for example to be ${\bf {\it o}}=(0,1)$, for a two-class classification problem. The feed-forward neural network has one output node per class. The vector $(0,1)$ indicates that the observed feature-vector belongs to the 2'nd class.

• That's not the question. – orome Apr 10 '18 at 0:16

The log-likelihood is not directly linked to the entropy in the context of your question. The similarity is superficial: both have the sums of logarithms of probability-like quantities.

The logarithm in log-likelihood (MLE) is done purely for numerical calculation reasons. The product of probabilities can be a very small number, especially if your sample is large. Then the range of likelihoods goes from 1 to disappearingly small value of a product. When you get the log, the product becomes a sum, and the log function compresses the range of values to a smaller more manageable domain. Logarithm is a monotonous function, so the max (min) of log-likelihood will produce the same answer of the likelihood itself. Hence, the presence of the log in MLE expression is not important in mathematical sense, and is simply a matter of convenience.

The presence of a logarithm function in the entropy is more substantial, and has its roots in statistical mechanics, a branch of physics. It's linked to Boltzmann distribution, which is used in theory of gases. You could derive the air pressure as a function of the altitude using it, for instance.

• Can you highlight what part of the question this addresses? – orome Apr 10 '18 at 18:57
• As I say in the OP, it's clear that the use of the log in the second way of expressing MLE is mere convenience (your first two paragraphs). And your last paragraph just seems to say that the presence of the log in the expression for entropy is meaningful — in the context of entropy (notably physics). But what's missing (and this is the question) is a justification for linking these two distinct (and true) observations. I don't see one, other than the equation after (3) being a useful way to express the second equation for MLE. Perhaps that's what you're saying? – orome Apr 10 '18 at 19:22
• @orome, you can make NN to calculate the entropy, of course, but that's not how the cross entropy function is actually used in most cases. You can think of it as another kind of a cost function, that's all to it here. It seems to have desired properties, and is nicely symmetrical. – Aksakal Apr 10 '18 at 19:25
• Yes, so calling it entropy or suggesting that the $\mathbf{a}_{y^{(i)}}(\mathbf{x}^{(i)}; \mathbf{θ}_{ML})$ are meaningful distrubutions (for which "entropy" provides any insight) is misleading. – orome Apr 10 '18 at 19:30
• @orome, I wouldn't obsess about the name. It's like "hinge loss" function has little to do with hinges. They call this "entropy loss" because its functional form is exactly like an information entropy equation. – Aksakal Apr 10 '18 at 19:43