I am stuck with a supposedly simple problem. I have a mixture of experts system, consisting of multiple neural networks for classification, whose mixture weights are determined by the data as well. For the posterior probability of label y, given data x and K different expert networks, we have:
In this scheme, p(y|z_k,x) are expert network posterior probabilities, which are Softmax functions applied to network outputs. p(z_k|x) are the network weights.
My problem is the following. Usually, in Pytorch we feed the outputs of the last layer (logits) into the cross entropy loss function. Pytorch handles the numerical stability issues with the Logsumexp trick (How is log_softmax() implemented to compute its value (and gradient) with better speed and numerical stability?). In my case here however, my model output is not the logits, the probabilities, directly instead, due to the nature of the mixture model. Taking the logarithm of the mixture probabilities and feeding into the NLL loss crashes after a couple of iterations since some probabilities quickly become very close to 0 and underflow-overflow issues start to appear. The calculation would be very unstable, numerically.
In this particular case, what would be the correct way to calculate CE (or NLL) loss, without losing the numerical stability?
I have a few set of questions related to the usage of various activation functions used in neural networks? I would highly appreciate if someone could give good explanatory answers.
Why ReLU is used only on hidden layers specifically?
Why Sigmoid is a not used in Multi-class classification?
Why we do not use any activation function in regression problems having all negative values?
Why we use "average='micro','macro','average'" while calculating performance metric in multi_class classification?
I'll answer to the best of my ability the 2 first questions:
Relu (=max(0,x)) is used to extract feature maps from data. This is why it is used in the hidden layers where we're learning what important characteristics or features the data holds that could make the model learn how to classify for example. In the FC layers, it's time to make a decision about the output, so we usually use sigmoid or softmax, which tend to give us numbers between 0 and 1 (probability) that can give an interpretable result.
Sigmoid gives a probability for each class. So, if you have 10 classes, you'll have 10 probabilities. And depending on the threshold used, your model would predict for example that the image corresponds to two classes when in multi-classification you want just one predicted class per image. That's why softmax is used in this context: It chooses the class with the maximum probability. So it'll predict just one class.
Keras model.fit supports per-sample weights. What is the range of acceptable values for these weights? Must they sum to 1 across all training samples? Or does keras accept any weight values and then perform some sort of normalization? The keras source includes, e.g. training_utils.standardize_weights but that does not appear to be doing statistical standardization.
After looking at the source here, I've found that you should be able to pass any acceptable numerical values (within overflow bounds) for both sample weights and class weights. They do not need to sum to 1 across all training samples and each weight may be greater than one. The only sort of normalization that appears to be happening is taking the max of 2D class weight inputs.
If both class weights and samples weights are provided, it provides the product of the two.
I think the unspoken component here is that the activation function should be dealing with normalization.
After training any classifier, the classifier tells the probability of data point belonging to a class.
y_pred = clf.predict_proba(test_point)
Does the classifier predicts the class with the max probability or does it considers the probabilities as a distribution draws according to distribution?
In other words, suppose the output probability is -
C1 - 0.1 C2 - 0.2 C3 - 0.7
Will the output be C3 always or only 70% of the times?
When clf predict it won’t calculate the probably of each class . It will use the full connect get a array like [itemsnum ,classisnum] then you can use max output[1] get the items class
by the way when clf training it use softmax to get the probably of each class which is more smooth to optimize you can find some doc about softmax if you are interested about train process
How to go from class probability scores to a class is often called the 'decision function', and is often considered separate from the classifier itself. In scikit-learn, many estimators have a default decision function accessible via predict() for multi-class problems this generally just returns the largest value (argmax function).
However this may be extended in various ways, depending on needs. For instance if the effects of one prediction one of the classes is very costly, then one might weight those probabilities down (class weighting). Or one can have a decision function that only gives a class as output if the confidence is high, else returns an error or a fallback class.
One can also have multi-label classification, there the output is not a single class but a list of classes. [ 0.6, 0.1, 0.7, 0.2 ] -> (class0, class2) These can then use a common threshold, or a per-class threshold. This is common in tagging problems.
But in almost all cases the decision function is a deterministic function, not a probabilistic one.
Most examples of neural networks for classification tasks I've seen use the a softmax layer as output activation function. Normally, the other hidden units use a sigmoid, tanh, or ReLu function as activation function. Using the softmax function here would - as far as I know - work out mathematically too.
What are the theoretical justifications for not using the softmax function as hidden layer activation functions?
Are there any publications about this, something to quote?
I haven't found any publications about why using softmax as an activation in a hidden layer is not the best idea (except Quora question which you probably have already read) but I will try to explain why it is not the best idea to use it in this case :
1. Variables independence : a lot of regularization and effort is put to keep your variables independent, uncorrelated and quite sparse. If you use softmax layer as a hidden layer - then you will keep all your nodes (hidden variables) linearly dependent which may result in many problems and poor generalization.
2. Training issues : try to imagine that to make your network working better you have to make a part of activations from your hidden layer a little bit lower. Then - automaticaly you are making rest of them to have mean activation on a higher level which might in fact increase the error and harm your training phase.
3. Mathematical issues : by creating constrains on activations of your model you decrease the expressive power of your model without any logical explaination. The strive for having all activations the same is not worth it in my opinion.
4. Batch normalization does it better : one may consider the fact that constant mean output from a network may be useful for training. But on the other hand a technique called Batch Normalization has been already proven to work better, whereas it was reported that setting softmax as activation function in hidden layer may decrease the accuracy and the speed of learning.
Actually, Softmax functions are already used deep within neural networks, in certain cases, when dealing with differentiable memory and with attention mechanisms!
Softmax layers can be used within neural networks such as in Neural Turing Machines (NTM) and an improvement of those which are Differentiable Neural Computer (DNC).
To summarize, those architectures are RNNs/LSTMs which have been modified to contain a differentiable (neural) memory matrix which is possible to write and access through time steps.
Quickly explained, the softmax function here enables a normalization of a fetch of the memory and other similar quirks for content-based addressing of the memory. About that, I really liked this article which illustrates the operations in an NTM and other recent RNN architectures with interactive figures.
Moreover, Softmax is used in attention mechanisms for, say, machine translation, such as in this paper. There, the Softmax enables a normalization of the places to where attention is distributed in order to "softly" retain the maximal place to pay attention to: that is, to also pay a little bit of attention to elsewhere in a soft manner. However, this could be considered like to be a mini-neural network that deals with attention, within the big one, as explained in the paper. Therefore, it could be debated whether or not Softmax is used only at the end of neural networks.
Hope it helps!
Edit - More recently, it's even possible to see Neural Machine Translation (NMT) models where only attention (with softmax) is used, without any RNN nor CNN: http://nlp.seas.harvard.edu/2018/04/03/attention.html
Use a softmax activation wherever you want to model a multinomial distribution. This may be (usually) an output layer y, but can also be an intermediate layer, say a multinomial latent variable z. As mentioned in this thread for outputs {o_i}, sum({o_i}) = 1 is a linear dependency, which is intentional at this layer. Additional layers may provide desired sparsity and/or feature independence downstream.
Page 198 of Deep Learning (Goodfellow, Bengio, Courville)
Any time we wish to represent a probability distribution over a discrete variable with n possible values, we may use the softmax function. This can be seen as a generalization of the sigmoid function which was used to represent a probability
distribution over a binary variable.
Softmax functions are most often used as the output of a classifier, to represent the probability distribution over n different classes. More rarely, softmax functions can be used inside the model itself, if we wish the model to choose between one of n different options for some internal variable.
Softmax function is used for the output layer only (at least in most cases) to ensure that the sum of the components of output vector is equal to 1 (for clarity see the formula of softmax cost function). This also implies what is the probability of occurrence of each component (class) of the output and hence sum of the probabilities(or output components) is equal to 1.
Softmax function is one of the most important output function used in deep learning within the neural networks (see Understanding Softmax in minute by Uniqtech). The Softmax function is apply where there are three or more classes of outcomes. The softmax formula takes the e raised to the exponent score of each value score and devide it by the sum of e raised the exponent scores values. For example, if I know the Logit scores of these four classes to be: [3.00, 2.0, 1.00, 0.10], in order to obtain the probabilities outputs, the softmax function can be apply as follows:
import numpy as np
def softmax(x):
z = np.exp(x - np.max(x))
return z / z.sum()
scores = [3.00, 2.0, 1.00, 0.10]
print(softmax(scores))
Output: probabilities (p) = 0.642 0.236 0.087 0.035
The sum of all probabilities (p) = 0.642 + 0.236 + 0.087 + 0.035 = 1.00. You can try to substitute any value you know in the above scores, and you will get a different values. The sum of all the values or probabilities will be equal to one. That’s makes sense, because the sum of all probability is equal to one, thereby turning Logit scores to probability scores, so that we can predict better. Finally, the softmax output, can help us to understand and interpret Multinomial Logit Model. If you like the thoughts, please leave your comments below.