The problem of learning

These notes are written in English and refer to the fantastic Caltech online course you can find here: http://work.caltech.edu/telecourse.html. It is also worth noting you can find it on iTunes U, and if you have an iPad this makes a fantastic combo :-) The book is here: http://www.amazon.com/gp/product/1600490069. I make ample use of Yaser’s material, including linking the relevant lecture PDF in each note and using his pictures (this is much faster than drawing stuff myself). Please note this is just for personal use and I give ALL the due credit to him!

A disclaimer on my notes in general: they are meant for my personal use and, although I’d love to know they helped somebody else, they are not written with clarity for third parties in mind. However, if you do use them and you would like to point out any errors, please write me an email!

#The problem of learning

RELEVANT SLIDES: Lecture 1

The problem of learning can be defined as that of learning some function1 we don’t know but of which we have some data points.

There are some basic hypotheses that need be satisfied before one can (or should) rely to a machine learning approach:

  1. A pattern must exist in the data we have. If there is no pattern, there is nothing to learn.
  2. We cannot pin this pattern down mathematically (for example, why should you use machine learning to predict how long an object takes to fall in the vacuum? You already have formulas for that).
  3. We have enough data on it. This is the less trivial of the three and it is worth a thorough discussion because we don’t know what “enough” means. How much is it? A function has an infinite number of points in its domain and we cannot possibly comprehend all of them, therefore if we were to resort to binary logic we should always say that there is never enough data. However, if we accept a more fuzzy approach, we can say that the performance of the machine learning model is strongly correlated with the amount of data we have. If (and only if) we accept the possibility of getting some predictions (slightly) wrong, then we can learn. This approach falls under the name of PAC, “Probably, Approximately Correct”2.

To express our notation and pin down some concepts, let us focus on a simple example: say we are a bank and we want to have a machine learning system handle credit card applicants.

In this problem, we have:

  • Each applicant is formalized as an input vector
  • The decision for each applicant is a scalar (reject, accept)
  • The target function is , where and .

Following this formalism, all the data relevant to previous customers is in the form

Classification of machine learning problems

In general, a machine learning problem is regarded as:

  • A regression problem when is real-valued and can therefore assume an infinite amount of values. Time-series predictions in stocks fall in this category.
  • A classification problem when the domain of is composed by a finite and pre-determined number of “buckets” (for example, you may want to classify stuff among 5 classes. The number “5” is pre-determined). A common case is binary classification, like the one in this example.

The problem is also classified with respect to the form of the input:

  • We have a supervised learning problem when we have data in the form i.e. we have information about in the training set.
  • We have a unsupervised learning problem when we have data in the form i.e. we have no information about - “Here is my previous data, please find some patterns in it”. Clustering and anomaly detection fall in this category.
  • We have a reinforced learning problem when we don’t have direct information about , but we do have an indication of how we performed, like “good” or “bad”. This is used in chess AI, where you learn whether a move was good or bad by judging what happened later, or maybe in robotics when you have to make a robotic arm “learn” how to move.

Needless to say, there categories are orthogonal. For example, the credit card application problem is a supervised learning, binary classification problem. In general, we will focus on supervised learning and classification because they have generally more interesting applications.

This brings us to our first learning framework diagram.

diagram1

There is an extra element we did not talk about, which is the learning algorithm. There are multiple learning algorithms and we can expect some of them to be better than others for a certain application (this is the case). The choice of the algorithm involves making some a priori assumptions about the form of your data (smoothness being probably the most common and the most general3). All these assumptions restrict the domain in which we pick our hypotheses (i.e. our candidate functions) to a certain domain . We pick many candidate functions and feed them to the learning algorithm which will elect (what it thinks it is) the best function and our final hypothesis, .

  1. Later, this will be expanded to any general distribution.

  2. The definition comes from Prof. Leslie Valiant. Look at http://www.amazon.com/dp/0465032710/ref=rdr_ext_sb_ti_sims_2.

  3. Long story short, smoothness requires that infinitesimal variations in the domain reflect in infinitesimal variations in the codomain. This can be expressed as a requirement of differentiability. For a more formal definition, look at Wikipedia: http://en.wikipedia.org/wiki/Smoothness.