The Azimuth Project
Blog - Science, models and machine learning (Rev #12)

This page is a blog article in progress, written by David Tweed. To see discussions of this article while it was being written, visit the Azimuth Forum.

Please remember that blog articles need HTML, not Markdown.

Science, models and machine learning

The members of the Azimuth Project have been working on both predicting and understanding the El Niño phenomenon, along with expository articles such as this. So far we've mostly talked about the physics and data of the El Niño, along with looking at one method of actually trying to predict El Niño events. Since there's going to more data exploration using methods more typical of machine learning, it's an opportune time to briefly describe the mindset and highlight some of differences between different kinds of predictive models. Here we'll concentrate on the concepts rather than the fine details and particular techniques.

We also stress there's not a fundamental distinction between machine learning (ML) and statistical modelling and inference. There are certainly differences in culture, background and terminology, but in terms of the actual algorithms and mathematics used there's a great commonality. Throughout the rest of the article we'll use the talk about "machine learning models", but could equally have used "statistical models".

For our purposes, a model is any systematic procedure for taking some input data and providing a prediction of some output data. There's a spectrum of models, ranging from physically based models at one end to purely data driven models at the other. As a very simple example, suppose you have a commute by car from your place of work to your home and you want to leave work in order to arrive home at 6.30pm. You can tackle this by building a model which takes as input the day of the week and gives you back a time to leave.

  • There's the data driven approach, where you try various leaving times on various days and record whether or not you get home by 6.30pm. You might find that the traffic is lighter on weekend days so you can leave at 6.10pm while on weekdays you have to leave at 5.45pm, except on Wednesdays when you have to leave at 5.30pm. Since you've just crunched on the data you have no idea why this works, but it's a very reliable rule when you use it to predict when you need to leave.

  • There's the physical model approach, where you attempt to infer how many people are doing what on any given day and then figure out what that implies for the traffic levels and thence what time you need to leave. (Of course this is just an illustrative example; in climate modelling a physical model would be based upon actual physical laws such as conservation of energy, conservation of momentum, Boyle's law, etc.)

    In this case you find out that there's a mid-week sports game on Wednesday evenings which leads to even higher traffic. This not only predicts that you've got to leave at 5.30pm on Wednesday's but also lets you understand why.

There are trade-offs between the two types of approach. Data driven modelling is a relatively simple process. In contrast, by proceeding from first principles you've got a more detailed framework which is equally predictive but at the cost of having to investigate a lot of complicated underlying effects. Physical models have one advantage: nothing in the data driven model prevents it violating physical laws (eg, not conserving energy, etc) whereas a physically based model obeys the physical laws by design. This is seldom a problem in practice, but worth keeping in mind.

The situation with data driven techniques is analogous to one of those American medication adverts: there's the big message about how "using a purely data driven techniques can change your life for the better" while the voiceover gabbles out all sorts of small print. The remainder of this post will try to cover some of the basic principles in that small print.

Preprocessing and feature formation

There's a popular misconception that machine learning works well when you simply collect some data and throw it into a machine learning algorithm. In practice that kind of approach yields a model that is often quite poor. Almost all successful machine learning applications are preceded by some form of preprocessing. Sometimes this is simply rescaling variables so that different variables have similar magnitudes, are zero centred, etc. However, there are often steps that are more involved. For example, many machine learning techniques have what are called kernel variants which involves (in a way whose details don't matter here) choosing a nonlinear mapping from the original data to a new space which is more amenable to the core algorithm. There are various kernels with the right mathematical properties, and the choice of a good kernel frequently happens either by experimentation or knowledge of the physical principles. Here's an example (from Wikipedia's entry on the SVM) of how a good choice of kernel can convert a not linearly separable dataset into a linearly separable one:

An extreme example of preprocessing is explicitly forming new features in the data. For example, in the work by Ludescher et al that we've been looking at, the correlation between different points are taken as the basic features to consider. While these could theoretically be learned by the ML algorithm, this is quite a complicated function. By explicitly choosing to represent the dat using this feature the amount the algorithm has to discover is reduced and hence the likelihood of it finding an excellent model is dramatically increased.

Limited amounts of data for model development

Some of the problems that we describe below would vanish if we had unlimited amounts of data to use for model development. However, in real cases we often have a strictly limited amount of data we can use for development. Consequently we need methodology to address the issues that arise when data is limited.

Training sets and test sets

The most common way to work with collected data is to split it into a training set and a test set. (Sometimes there is a division into three sets, the third being a validation set.) The training and validation sets are used in the process of determining the best model parameters, while the test set – which is not used in any way in determining the best model parameters – is then used to see how effective the model is likely to be on new, unseen data. This division of data into multiple sets acts to further reduce the effective amount of data used to setting model parameters.

After we've made this split we have to be careful how much of the test data we scrutinise in any detail since once it has been investigated it can't meaningfully be used for testing again, although it can still be used for future training. (Informally this is often known as burning data.) That only applies to detailed inspection however; one common way to develop a model is to look at some training data and then train the model (also known as fitting the model) to that training data. It can then be evaluated on the test data to see how well it does. It's also then OK to purely mechanically train the model on the test data and evaluate it on the training data to see how "stable" the performance is. (If you get dramatically different scores then your model is probably flaky!) However, once we start to look at precisely why the model failed on the test data — in order to change the form of the model — the test data has now become training data and can't be used as test data for future variants of that model. (Remember, the real goal is to accurately predict the outputs for new, unseen inputs.)

Random patterns in small samples

Suppose we're modelling a system where some aspect has a true probability distribution $ P() $. We can't directly observe this, but we have some samples $ S $ obtained from observation of the system and hence come from $ P() $. Clearly there are problems if we generate this sample in a way that will bias the area of the distribution we sample from: it wouldn't be a good idea to try to get training data featuring heights in the American population by only handing out surveys in the locker rooms of basketball facilities. But if we take care to avoid (as much as possible) any sampling bias, then we can make various kinds of estimates of the distribution that we think $ S $ comes from; let's consider the estimate $ P'() $ implied for $ S $ by some particular technique. It would be nice if $ P = P' $, wouldn't it? And indeed many good estimators have the property that as the size of $ S $ tends to infinity $ P' $ will tend to $ P $. However, for finite sizes of $ S $, and especially for small sizes, $ P' $ may have some spurious detail that's not present in $ P $.

As a simple illustration of this, my computer has a pseudo-random number generator which generates essentially uniformly distributed random numbers between 0 and 32767. I just asked for 8 numbers and got

  • 2928, 6552, 23979, 1672, 23440, 28451, 3937, 18910.

Note that we've got one subset of 4 values (2928, 6552, 1672, 3937) within the interval of length 5012 between 1540 and 6552 and another subset of 3 values (23440, 23979 and 28451) the interval of length 5012 between 23440 and 28451. For this uniform distribution the expected value of the number of values falling within a given range of that size is about 1.2. Readers will be familiar with how the expectation of a random quantity for a small sample will have a large amount of variation around its value that only reduces as the sample size increases, so this isn't a surprise. However, it does highlight that even with completely unbiased sampling from the true distribution will typically give rise to extra "structure" within the distribution implied by the samples. For example, here's results from one way of estimating the probability from the samples:
The green line is the true density while the red curve shows the probability density obtained from the samples, with clearly spurious extra structure.

Generalization

Almost all modelling techniques, while not necessarily estimating a full model of the probability distributions from the training samples, can be seen as building functions that are related to the probability distribution: for example, a thresholding classifier for dividing input into two output classes will place the threshold at the optimal point for the distribution implied by the samples. As a consequence, one important aim in building machine learning models is to try to estimate the features that are present in the true probability distribution while not learning such fine details that they are likely to be spurious features due to the small sample size. If you think about this, it's a bit counter-intuitive: you deliberately don't want to perfectly reflect every single the pattern in the training data. Indeed, specialising a model too closely to the training is given the name over-fitting.

This brings us to generalization. Strictly speaking generalization is the ability of a model to work well upon unseen instances of the problem (which may be difficult for a variety of reasons). In practice however one tries hard to get representative training data so that the main issue in generalization is in preventing overfitting, and the main way to do that is – as discussed above – to split the data into a set for training and a set only used for testing.

One factor that's often related to generalization is regularization and in particular sparsity. Sparsity refers to the degree to which a model has empty elements, typically represented as 0 coefficients. There are various reasons for wanting sparse models: the obvious one is speed of model evaluation, although that is much less significant with modern computing power. It's often possible to incorporate a prior into the modelling procedure which will encourage the model to be sparse. (Recall that in Bayesian modelling the prior represents our initial ideas of how likely various different parameter values are.) There are some cases where we have various detailed Bayesian priors about sparsity for problem specific reasons. However the more common case is having a "general modelling" belief, based upon general experience in modelling, that sparser models have a better generalization performance.

There are a couple of other reasons for wanting sparse models. The obvious one is speed of model evaluation, although that is much less significant with modern computing power. A less obvious reason is that one can often only effectively utilise a sparse model, eg, if you're attempting to see how the input factors should be physically modified in order to affect the real system in a particular way. In this case one might want a good sparse model rather than an excellent dense model.

Inferring a physical model from a ML model

It is certainly possible to take a predictive model obtained by machine learning and use it to figure out a physically based model; this is one way of performing what's known a data mining. However in practice there are a couple of reasons why it's necessary to exhibit some care when doing this:

  • The variables in the training set may be related by some non-observed latent variables which may be difficult to reconstruct without knowledge of the physical laws that are in play. (There are machine learning techniques which attempt to reconstruct unknown latent variables but this is a much more difficult problem than estimating known but unobserved latent variables.)

  • Machine learning models have a maddening ability to find variables that are predictive due to the way the training data was gathered. For example, in an early vision system aimed at finding tanks all the images of tanks were taken during one day on a military base when there was accidentally a speck of grime on the camera lens, while all the images of things that weren't tanks were taken on other days. A neural net cunningly learned that to decide if it was being shown a tank it should see if the shadow from the grime was in place or not.

  • It's common to have some groups of very highly correlated input variables. In that case a model will generally learn a function which utilises an arbitrary linear combination of the correlated variables and an equally good model would result from using any other linear combination. Certain sparsity encouraging priors have the useful property of encouraging the model to select only one representative from a group of correlated variables. However, even in that case it's still important not to assign too much importance to the particular division of model parameters in groups correlated variables.

  • One can often come up with good machine learning models even when physically important variables haven't been collected in the training data. A related issue is that if all the training data is collected from a particular subspace factors that aren't important there won't be found. For example, if in a collision system to be modelled all data is collected at low speeds the machine learning model won't learn about relativistic effects that only have a big effect at a substantial fraction of the speed of light.

Utility functions and decision theory

While there are some situations where a model is sought purely to develop knowledge of the universe, in many cases we are interested in models in order to direct actions. For example, having forewarning of El Niño events would enable all sorts of mitigation actions. However, these actions are costly so they shouldn't be undertaken when there isn't an upcoming El Niño. When presented with an unseen input the model can either match the actual output (ie, be right) or differ from the actual output (ie, be wrong). While it's impossible to know in advance if a single ouput will be right or wrong, from the training data it's generally possible to estimate the fractions of predictions that will be right and will be wrong in a large number of trials. So we want to link these probabilities with the effects of actions taken in response to model predictions.

We can do this using a utility function and a loss function. The utility maps each possible output to a numerical value proportional to the benefit from taking actions when that output was correctly anticipated while the loss maps outputs to a number proportional to the loss from the actions when the output was incorrectly predicted by the model. (There is a evidence that human beings often have inconsistent utility/loss functions, but that's straying from the main point here...)

There are three common ways the utility and loss functions are used:

  • Maximising the expected value of the utility minus the loss for each output.

  • Minimising the expected loss while ensuring that the expected utility is at least some value.

  • Maximising the expected utility while ensuring that the expected loss is at most some value.

Once we've chosen which one we want, it's often possible to actually tune the fitting of the system model to ensure that criterion is optimized.

Of course sometimes when building a model we don't know enough details of how it will be used to get accurate utility and loss functions (or indeed know how it will be used at all).

Conclusions

All of the ideas discussed above are really just ways of making sure that work on statistical/machine learning models are producing meaningful results in situations where the training data is scarce. As Bob Dylan (almost) sang, "to live outside the physical law, you must be honest; I know you always say that you agree".

category: blog