1、with softmax regression to solve multi-classificationIn these notes, we describe the Softmax regression model. This model generalizes logistic regression to classification problems where the class label y can take on more than two possible values. This will be useful for such problems as MNIST digit
2、 classification, where the goal is to distinguish between 10 different numerical digits. Softmax regression is a supervised learning algorithm, but we will later be using it in conjuction with our deep learning/unsupervised feature learning methods.Recall that in logistic regression, we had a traini
3、ng set of m labeled examples, where the input features are . (In this set of notes, we will use the notational convention of letting the feature vectors x be n+ 1 dimensional, with x0 = 1 corresponding to the intercept term.) With logistic regression, we were in the binary classification setting, so
4、 the labels were . Our hypothesis took the form:and the model parameters were trained to minimize the cost functionIn the softmax regression setting, we are interested in multi-class classification (as opposed to only binary classification), and so the label y can take on k different values, rather
5、than only two. Thus, in our training set , we now have that . (Note that our convention will be to index the classes starting from 1, rather than from 0.) For example, in the MNIST digit recognition task, we would have k = 10 different classes.Given a test input x, we want our hypothesis to estimate
6、 the probability that p(y = j | x) for each value of . I.e., we want to estimate the probability of the class label taking on each of the k different possible values. Thus, our hypothesis will output a k dimensional vector (whose elements sum to 1) giving us our k estimated probabilities. Concretely
7、, our hypothesis h(x) takes the form:Here are the parameters of our model. Notice that the term normalizes the distribution, so that it sums to one.For convenience, we will also write to denote all the parameters of our model. When you implement softmax regression, it is usually convenient to repres
8、ent as a k-by-(n + 1) matrix obtained by stacking up in rows, so thatCost FunctionWe now describe the cost function that well use for softmax regression. In the equation below, is the indicator function, so that 1a true statement = 1, and 1a false statement = 0. For example, 12 + 2 = 4 evaluates to
9、1; whereas11 + 1 = 5 evaluates to 0. Our cost function will be:Notice that this generalizes the logistic regression cost function, which could also have been written:The softmax cost function is similar, except that we now sum over the k different possible values of the class label. Note also that i
10、n softmax regression, we have that .There is no known closed-form way to solve for the minimum of J(), and thus as usual well resort to an iterative optimization algorithm such as gradient descent or L-BFGS. Taking derivatives, one can show that the gradient is:Recall the meaning of the “ “ notation
11、. In particular, is itself a vector, so that its l-th element is the partial derivative of J() with respect to the l-th element of j.Armed with this formula for the derivative, one can then plug it into an algorithm such as gradient descent, and have it minimize J(). For example, with the standard i
12、mplementation of gradient descent, on each iteration we would perform the update (for each ).When implementing softmax regression, we will typically use a modified version of the cost function described above; specifically, one that incorporates weight decay. We describe the motivation and details b
13、elow.Properties of softmax regression parameterizationSoftmax regression has an unusual property that it has a “redundant“ set of parameters. To explain what this means, suppose we take each of our parameter vectors j, and subtract some fixed vector from it, so that every j is now replaced withj (fo
14、r every ). Our hypothesis now estimates the class label probabilities asIn other words, subtracting from every j does not affect our hypothesis predictions at all! This shows that softmax regressions parameters are “redundant.“ More formally, we say that our softmax model is overparameterized, meani
15、ng that for any hypothesis we might fit to the data, there are multiple parameter settings that give rise to exactly the same hypothesis function h mapping from inputs x to the predictions.Further, if the cost function J() is minimized by some setting of the parameters , then it is also minimized by
16、 for any value of . Thus, the minimizer of J() is not unique. (Interestingly, J() is still convex, and thus gradient descent will not run into a local optima problems. But the Hessian is singular/non-invertible, which causes a straightforward implementation of Newtons method to run into numerical pr
17、oblems.)Notice also that by setting = 1, one can always replace 1 with (the vector of all 0s), without affecting the hypothesis. Thus, one could “eliminate“ the vector of parameters 1 (or any other j, for any single value of j), without harming the representational power of our hypothesis. Indeed, r
18、ather than optimizing over the k(n + 1) parameters (where ), one could instead set and optimize only with respect to the (k 1)(n + 1)remaining parameters, and this would work fine.In practice, however, it is often cleaner and simpler to implement the version which keeps all the parameters , without
19、arbitrarily setting one of them to zero. But we will make one change to the cost function: Adding weight decay. This will take care of the numerical problems associated with softmax regressions overparameterized representation.Weight DecayWe will modify the cost function by adding a weight decay ter
20、m which penalizes large values of the parameters. Our cost function is nowWith this weight decay term (for any 0), the cost function J() is now strictly convex, and is guaranteed to have a unique solution. The Hessian is now invertible, and because J() is convex, algorithms such as gradient descent,
21、 L-BFGS, etc. are guaranteed to converge to the global minimum.To apply an optimization algorithm, we also need the derivative of this new definition of J(). One can show that the derivative is: By minimizing J() with respect to , we will have a working implementation of softmax regression.Relations
22、hip to Logistic RegressionIn the special case where k = 2, one can show that softmax regression reduces to logistic regression. This shows that softmax regression is a generalization of logistic regression. Concretely, when k = 2, the softmax regression hypothesis outputsTaking advantage of the fact
23、 that this hypothesis is overparameterized and setting = 1, we can subtract 1 from each of the two parameters, giving usThus, replacing 2 1 with a single parameter vector , we find that softmax regression predicts the probability of one of the classes as , and that of the other class as , same as lo
24、gistic regression.Softmax Regression vs. k Binary ClassifiersSuppose you are working on a music classification application, and there are k types of music that you are trying to recognize. Should you use a softmax classifier, or should you build k separate binary classifiers using logistic regressio
25、n?This will depend on whether the four classes are mutually exclusive. For example, if your four classes are classical, country, rock, and jazz, then assuming each of your training examples is labeled with exactly one of these four class labels, you should build a softmax classifier with k = 4. (If
26、therere also some examples that are none of the above four classes, then you can set k = 5 in softmax regression, and also have a fifth, “none of the above,“ class.)If however your categories are has_vocals, dance, soundtrack, pop, then the classes are not mutually exclusive; for example, there can
27、be a piece of pop music that comes from a soundtrack and in addition has vocals. In this case, it would be more appropriate to build 4 binary logistic regression classifiers. This way, for each new musical piece, your algorithm can separately decide whether it falls into each of the four categories.
28、Now, consider a computer vision example, where youre trying to classify images into three different classes. (i) Suppose that your classes are indoor_scene, outdoor_urban_scene, and outdoor_wilderness_scene. Would you use sofmax regression or three logistic regression classifiers? (ii) Now suppose y
29、our classes are indoor_scene, black_and_white_image, and image_has_people. Would you use softmax regression or multiple logistic regression classifiers?In the first case, the classes are mutually exclusive, so a softmax regression classifier would be appropriate. In the second case, it would be more appropriate to build three separate logistic regression classifiers.
