Fruit Loops and Learning - The LUPI Paradigm and SVM+

Here’s a short story you might know: you have a black box, whose name is Machine Learning Algorithm. It’s got two modes: training mode and testing mode. You set it to training mode, and throw in a lot (sometimes a lot a lot) of ordered pairs \((x_i, y_i), 1 \leq i \leq l\). Here, the \(x_i\) are called the examples and the \(y_i\) are called the targets. Then, you set it to testing mode and throw in some more examples, for which you don’t have the corresponding targets. You hope the \(y_i\)s that come out are in some sense the “right” ones.

Generally speaking, this is a parable of supervised learning. However, Vapnik (the inventor of the SVM) recently described a new way to think about machine learning (e.g. here): learning using privileged information, or LUPI for short.

This post is meant to introduce the LUPI paradigm of machine learning to people who are generally familiar with supervised learning and SVMs, and are interested in seeing the math and intuition behind both things extended to the LUPI paradigm.

What is LUPI?

The main idea is that instead of two-tuples \((x_i, y_i)\), the black box is fed three-tuples \((x_i, x_i^{*}, y_i)\), where the \(x^{*}\)s are the so-called privileged information that is only available during training, and not during testing. The hope is that this information will train the model to better generalize during the testing phase.

Vapnik offers many examples in which LUPI can be applied in real life: in bioinformatics and proteomics (where advanced biological models, which the machine might not necessarily “understand”, serve as the privileged information), in financial time series analysis (where future movements of the time series are the unknown at prediction time, but are available retrospectively), and in the classic MNIST dataset, where the images were converted to a lower resolution, but each annotated with a “poetic description” (which was available for the training data but not for the testing data).

Vapnik’s team ran tests on well-known datasets in all three application areas and found that his newly-developed LUPI methods performed noticeably better than classical SVMs in both convergence time (i.e. the number of examples necessary to achieve a certain degree of accuracy) and estimation of a good predictor function. In fact, Vapnik’s proof-of-concept experiments are so whacky that they actually make for an entertaining read !

Classical SVMs (separable and non-separable case)

There are many ways of thinking about SVMs, but I think that the one that is most instructive here is to think of them as solving the following optimization problem:

Minimize \(\frac{1}{2} \|w\|^2\)

subject to \(y_i [ w \cdot x_i + b ] \geq 1, \; 1 \leq i \leq l\).

Basically all this is saying is that we want to find the hyperplane that separates our data by the maximum margin. More technically speaking, this finds the parameters (\(w\) and \(b\)) of the maximum margin hyperplane, with \(l_2\) regularization.

In the non-separable case, we concede that our hyperplane may not classify all examples perfectly (or that it may not be desireable to do so: think of overfitting), and so we introduce a so-called slack variable \(\xi_i \geq 0\) for each example \(i\), which measures the severity of misclassification of that example. With that, the optimization becomes:

Minimize \(\frac{1}{2} \|w\|^2 + C\sum_{i=1}^{l}{\xi_i}\)

subject to \(y_i [ w \cdot x_i + b ] \geq 1 - \xi_i, \; \xi_i \geq 0, 1 \leq i \leq l\).

where \(C\) is some regularization parameter.

This says the same thing as the previous optimization problem, but now allows points to be (a) classified properly (\(\xi_i = 0\)), (b) within the margin but still classified properly (\(0 < \xi_i < 1\)), or (c) misclassified (\(1 \leq \xi_i\)).

In both the separable and non-separable cases, the decision rule is simply \(\hat{y} = \text{sign}(w \cdot x + b)\).

An important thing to note is that, in the separable case, the SVM uses \(l\) examples to estimate the \(n\) components of \(w\), whereas in the nonseparable case, the SVM uses \(l\) examples to estimate \(n+l\) parameters: the \(n\) components of \(w\) and \(l\) values of slacks \(\xi_i\). Thus, in the non-separable case, the number of parameters to be estimated is always larger than the number of examples: it does not matter here that most of slacks may be equal to zero: the SVM still has to estimate all of them.

The way both optimization problems are actually solved is fairly involved (they require Lagrange multipliers), but in terms of getting an intuitive feel for how SVMs work, I think that examining the optimization problems suffice!

What is SVM+?

In his paper introducing the LUPI paradigm, Vapnik outlines SVM+, a modified form of the SVM that fits well into the LUPI paradigm, using privileged information to improve performance. It should be emphasized that LUPI is a paradigm - a way of thinking about machine learning - and not just a collection of algorithms. SVM+ is just one technique that interoperates with the LUPI paradigm.

The innovation of the SVM+ algorithm is that is uses the privileged information to estimate the slack variables. Given the training three-tuple \((x, x^{*}, y)\), we map \(x\) to the feature space \(Z\), and \(x^{*}\) to a separate feature space \(Z^{*}\). Then, the decision rule is \(\hat{y} = \text{sign}(w \cdot x + b)\) and the slack variables are estimated by \(\xi = w^{*} \cdot x^{*} + b^{*}\).

In order to find \(w\), \(b\), \(w^{*}\) and \(b^{*}\), we solve the following optimization problem:

Minimize \(\frac{1}{2} (\|w\|^2 + \gamma \|w^{*}\|^2) + C \sum_{i=1}^{l}{(w^{*} \cdot x_i^{*} + b^{*})}\)

subject to \(y_i [ w \cdot x_i + b ] \geq 1 - (w^{*} \cdot x^{*} + b^{*}), \; (w^{*} \cdot x^{*} + b^{*}) \geq 0, 1 \leq i \leq l\).

where \(\gamma\) indicates the extent to which the slack estimation should be regularized in comparison to the SVM. Notice how this optimization problem is essentially identical to the non-separable classical SVM, except the slacks \(\xi_i\) are now estimated with \(w^{*} \cdot x^{*} + b^{*}\).

Again, the method of actually solving this optimization problem involves Lagrange multipliers and quadratic programming, but I think the intuition is captured in the optimization problem statement.

Interpretation of SVM+

The SVM+ has a very ready interpretation. Instead of a single feature space, it has two: one in which the non-privileged information lives (where decisions are made), and one in which the privileged information lives (where slack variables are estimated).

But what’s the point of this second feature space? How does it help us? Vapnik terms this problem knowledge transfer: it’s all well and good for us to learn from the privileged information, but it’s all for naught if we can’t use this newfound knowledge in the test phase.

The way knowledge transfer is resolved here is by assuming that examples in the training set that are hard to separate in the privileged space, are also hard to separate in the regular space. Therefore, we can use the privileged information to obtain an estimate for the slack variables.

Of course, SVMs are a technique with many possible interpretations, of which my presentation (in terms of the optimization of \(w\) and \(b\)) is just one. For example, it’s possible to think of SVMs in terms of kernels functions, or as linear classifiers minimizing hinge loss. In all cases, it’s possible and worthwhile to understand that interpretation of SVMs, and how the LUPI paradigm contributes to or extends that interpretation. I’m hoping to write a piece later to explain these exact topics.

Vapnik also puts a great emphasis on analyzing SVM+ based on its statistical learning theoretic properties (in particular, analyzing its rate of convergence via the VC dimension). Vapnik was one of the main pioneers behind statistical learning theory, and has written an entire book on this stuff which I have not read, so I’ll leave that part aside for now. I hope to understand this stuff one day.

Implementation of SVM+

There’s just one catch: SVM+ is actually an fairly inefficient algorithm, and definitely will not scale to large data sets. What’s so bad about it? It has \(n\) training examples but \(2n\) variables to estimate. This is twice as many variables to estimate as the standard formulation of the vanilla SVM. This isn’t something that we can patch: the problem is inherent to the Lagrangian dual formulation that Vapnik and Vashist proposed in 1995.

Even worse, the optimization problem has constraints that are very different from those of the standard SVM. In essence, this means that efficient libraries out-of-the-box solvers for the standard SVM (e.g. LIBSVM and LIBLINEAR) can’t be used to train an SVM+ model.

Luckily, a recent paper by Xu et al. describes a neat mathematical trick to implement SVM+ in a simple and efficient way. With this amendment, the authors rechristen the algorithm as SVM2+. Essentially, instead of using the hinge loss when training SVM+, we will instead use the squared hinge loss. It turns out that changing the loss function in this way leads to a tiny miracle.

This (re)formulation of SVM+ becomes identical to that of the standard SVM, except we replace the Gram matrix (a.k.a. kernel matrix) \(\bf K\) by \(\bf K + \bf Q_\lambda \odot (\bf y y^t)\), where

• \(\bf y\) is the target vector
• \(\odot\) denotes the Hadamard product
• \(\bf{Q_\lambda}\) is given by \(Q_\lambda = \frac{1}{\lambda} (\tilde{K} (\frac{\lambda}{C} I_n + \tilde{K})^{-1} \tilde{K})\), and
• \(\bf \tilde{K}\) is the Gram matrix formed by the privileged information

So by replacing the hinge loss with the squared hinge loss, the SVM+ formulation can now be solved with existing libraries!

Extensions to SVM+

In his paper, Vapnik makes it clear that LUPI is a very general and abstract paradigm, and as such there is plenty of room for creativity and innovation - not just in researching and developing new LUPI methods and algorithms, but also in implementing and applying them. It is unknown how to best go about supplying privileged information so as to get good performance. How should the data be feature engineered? How much signal should be in the privileged information? These are all open questions.

Vapnik himself opens up three avenues to extend the SVM+ algorithm:

1. a mixture model of slacks: when slacks are estimated by a mixture of a smooth function and some prior
2. a model where privileged information is available only for a part of the training data: where we can only supply privileged information on a small subset of the training examples
3. multiple-space privileged information: where the privileged information we can supply do not all share the same features

Clearly, there’s a lot of potential in the LUPI paradigm, as well as a lot of reasons to be skeptical. It’s very much a nascent perspective of machine learning, so I’m interested in keeping an eye on it going forward. I’m hoping to write more posts on LUPI in the future!