In this article written by David Julian, author of the book Designing Machine Learning Systems with Python, the author wants to state that, he will first introduce the basic machine learning tasks. Classification is probably the most common task, due in part to the fact that it is relatively easy, well understood, and solves a lot of common problems. Multiclass classification (for instance, handwriting recognition) can sometimes be achieved by chaining binary classification tasks. However, we lose information this way, and we become unable to define a single decision boundary. For this reason, multiclass classification is often treated separately from binary classification.
(For more resources related to this topic, see here.)
There are cases where we are not interested in discrete classes but rather a real number, for instance, a probability. These type of problems are regression problems. Both classification and regression require a training set of correctly labelled data. They are supervised learning problems.
Originating from these basic machine tasks are a number of derived tasks. In many applications, this may simply be applying the learning model to a prediction to establish a causal relationship. We must remember that explaining and predicting are not the same. A model can make a prediction, but unless we know explicitly how it made the prediction, we cannot begin to form a comprehensible explanation. An explanation requires human knowledge of the domain.
We can also use a prediction model to find exceptions from a general pattern. Here, we are interested in the individual cases that deviate from the predictions. This is often called anomaly detection and has wide applications in areas such as detecting bank fraud, noise filtering, and even in the search for extraterrestrial life.
An important and potentially useful task is subgroup discovery. Our goal here is not, as in clustering, to partition the entire domain but rather to find a subgroup that has a substantially different distribution. In essence, subgroup discovery is trying to find relationships between a dependent target variable and many independent explaining variables. We are not trying to find a complete relationship but rather a group of instances that are different in ways that are important in the domain. For instance, establishing the subgroups, smoker = true and family history =true, for a target variable of heart disease =true.
Finally, we consider control type tasks. These act to optimize control setting to maximize a pay off is given different conditions. This can be achieved in several ways. We can clone expert behavior; the machine learns directly from a human and makes predictions of actions given different conditions. The task is to learn a prediction model for the expert's actions. This is similar to reinforcement learning, where the task is to learn about the relationship between conditions and optimal action.
Clustering, on the other hand, is the task of grouping items without any information on that group; this is an unsupervised learning task. Clustering is basically making a measurement of similarity.
Related to clustering is association, which is an unsupervised task to find a certain type of pattern in the data. This task is behind movie recommender systems, and customers who bought this also bought .. on checkout pages of online stores.
Data for machine learning
When considering raw data for machine learning applications, there are three separate aspects:
- The volume of the data
- The velocity of the data
- The variety of the data
The volume problem can be approached from three different directions: efficiency, scalability, and parallelism. Efficiency is about minimizing the time it takes for an algorithm to process a unit of information. A component of this is the underlying processing power of the hardware. The other component, and one that we have more control over, is ensuring our algorithms are not wasting precious processing cycles on unnecessary tasks.
Scalability is really about brute force, and throwing as much hardware at a problem as you can. With Moore's law, which predicts the trend of computer power doubling every two years and reaching its limit, it is clear that scalability is not, by its self, going to be able to keep pace with the ever increasing amounts of data. Simply adding more memory and faster processors is not, in many cases, going to be a cost effective solution.
Parallelism is a growing area of machine learning, and it encompasses a number of different approaches from harnessing capabilities of multi core processors, to large scale distributed computing on many different platforms. Probably, the most common method is to simply run the same algorithm on many machines, each with a different set of parameters. Another method is to decompose a learning algorithm into an adaptive sequence of queries, and have these queries processed in parallel. A common implementation of this technique is known as MapReduce, or its open source version, Hadoop.
The velocity problem is often approached in terms of data producers and data consumers. The data transfer rate between the two is its velocity, and it can be measured in interactive response times. This is the time it takes from a query being made to its response being delivered. Response times are constrained by latencies such as hard disk read and write times, and the time it takes to transmit data across a network.
Data is being produced at ever greater rates, and this is largely driven by the rapid expansion of mobile networks and devices. The increasing instrumentation of daily life is revolutionizing the way products and services are delivered. This increasing flow of data has led to the idea of streaming processing. When input data is at a velocity that makes it impossible to store in its entirety, a level of analysis is necessary as the data streams, in essence, deciding what data is useful and should be stored and what data can be thrown away. An extreme example is the Large Hadron Collider at CERN, where the vast majority of data is discarded. A sophisticated algorithm must scan the data as it is being generated, looking at the information needle in the data haystack. Another instance where processing data streams might be important is when an application requires an immediate response. This is becoming increasingly used in applications such as online gaming and stock market trading.
It is not just the velocity of incoming data that we are interested in. In many applications, particularly on the web, the velocity of a system's output is also important. Consider applications such as recommender systems, which need to process large amounts of data and present a response in the time it takes for a web page to load.
Collecting data from different sources invariably means dealing with misaligned data structures, and incompatible formats. It also often means dealing with different semantics and having to understand a data system that may have been built on a pretty different set of logical principles. We have to remember that, very often, data is repurposed for an entirely different application than the one it was originally intended for. There is a huge variety of data formats and underlying platforms. Significant time can be spent converting data into one consistent format. Even when this is done, the data itself needs to be aligned such that each record consists of the same number of features and is measured in the same units.
The goal in machine learning is not to just solve an instance of a problem, but to create a model that will solve unique problems from new data. This is the essence of learning. A learning model must have a mechanism for evaluating its output, and in turn, changing its behavior to a state that is closer to a solution.
A model is essentially a hypothesis: a proposed explanation for a phenomenon. The goal is to apply a generalization to the problem. In the case of supervised learning, problem knowledge gained from the training set is applied to the unlabeled test. In the case of an unsupervised learning problem, such as clustering, the system does not learn from a training set. It must learn from the characteristics of the dataset itself, such as degree of similarity. In both cases, the process is iterative. It repeats a well-defined set of tasks, that moves the model closer to a correct hypothesis.
There are many models and as many variations on these models as there are unique solutions. We can see that the problems that machine learning systems solve (regression, classification, association, and so on) come up in many different settings. They have been used successfully in almost all branches of science, engineering, mathematics, commerce, and also in the social sciences; they are as diverse as the domains they operate in.
This diversity of models gives machine learning systems great problem solving powers. However, it can also be a bit daunting for the designer to decide what is the best model, or models, for a particular problem. To complicate things further, there are often several models that may solve your task, or your task may need several models. The most accurate and efficient pathway through an original problem is something you simply cannot know when you embark upon such a project.
There are several modeling approaches. These are really different perspectives that we can use to help us understand the problem landscape. A distinction can be made regarding how a model divides up the instance space. The instance space can be considered all possible instances of your data, regardless of whether each instance actually appears in the data. The data is a subset of the instance space.
There are two approaches to dividing up this space: grouping and grading. The key difference between the two is that grouping models divide the instance space into fixed discrete units called segments. Each segment has a finite resolution and cannot distinguish between classes beyond this resolution. Grading, on the other hand, forms a global model over the entire instance space, rather than dividing the space into segments. In theory, the resolution of a grading model is infinite, and it can distinguish between instances no matter how similar they are. The distinction between grouping and grading is not absolute, and many models contain elements of both.
One of the most useful approaches to machine learning modeling is through geometry. Geometric models use the concept of instance space. The most obvious example is when all the features are numerical and can become coordinates in a Cartesian coordinate system. When we only have two or three features, they are easy to visualize. Since many machine learning problems have hundreds or thousands of features, and therefore dimensions, visualizing these spaces is impossible. Importantly, many of the geometric concepts, such as linear transformations, still apply in this hyper space. This can help us better understand our models. For instance, we expect many learning algorithms to be translation invariant, which means that it does not matter where we place the origin in the coordinate system. Also, we can use the geometric concept of Euclidean distance to measure similarity between instances; this gives us a method to cluster alike instances and form a decision boundary between them.
Often, we will want our models to output probabilities rather than just binary true or false. When we take a probabilistic approach, we assume that there is an underlying random process that creates a well-defined, but unknown, probability distribution. Probabilistic models are often expressed in the form of a tree. Tree models are ubiquitous in machine learning, and one of their main advantages is that they can inform us about the underlying structure of a problem. Decision trees are naturally easy to visualize and conceptualize. They allow inspection and do not just give an answer. For example, if we have to predict a category, we can also expose the logical steps that gave rise to a particular result. Also, tree models generally require less data preparation than other models and can handle numerical and categorical data. On the down side, tree models can create overly complex models that do not generalize very well to new data. Another potential problem with tree models is that they can become very sensitive to changes in the input data, and as we will see later, this problem can be mitigated by using them as ensemble learners.
A key concept in machine learning is that of the linear model. Linear models form the foundation of many advanced nonlinear techniques such as support vector machines and neural networks. They can be applied to any predictive task such as classification, regression, or probability estimation.
When responding to small changes in the input data, and provided that our data consists of entirely uncorrelated features, linear models tend to be more stable than tree models. Tree models can over-respond to small variation in training data. This is because splits at the root of a tree have consequences that are not recoverable further down a branch, potentially making the rest of the tree significantly different. Linear models, on the other hand, are relatively stable, being less sensitive to initial conditions. However, as you would expect, this has the opposite effect of making it less sensitive to nuanced data. This is described by the terms variance (for over fitting models) and bias (for under fitting models). A linear model is typically low variance and high bias.
Linear models are generally best approached from a geometric perspective. We know we can easily plot two dimensions of space in a Cartesian co-ordinate system, and we can use the illusion of perspective to illustrate a third. We have also been taught to think of time as being a fourth dimension, but when we start speaking of n dimensions, a physical analogy breaks down. Intriguingly, we can still use many of the mathematical tools that we intuitively apply to three dimensions of space. While it becomes difficult to visualize these extra dimensions, we can still use the same geometric concepts (such as lines, planes, angles, and distance) to describe them. With geometric models, we describe each instance as having a set of real-valued features, each of which is a dimension in a space.
Ensemble techniques can be divided broadly into two types.
- The Averaging Method: With this method, several estimators are run independently, and their predictions are averaged. This includes the random forests and bagging methods.
- The Boosting Methods: With this method, weak learners are built sequentially using weighted distributions of the data, based on the error rates.
Ensemble methods use multiple models to obtain better performance than any single constituent model. The aim is to not only build diverse and robust models, but also to work within limitations such as processing speed and return times. When working with large datasets and quick response times, this can be a significant developmental bottleneck. Troubleshooting and diagnostics are important aspects of working with all machine learning models, but they are especially important when dealing with models that might take days to run.
The types of machine learning ensembles that can be created are as diverse as the models themselves, and the main considerations revolve around three things: how we divide our data, how we select the models, and the methods we use to combine their results. This simplistic statement actually encompasses a very large and diverse space.
When we approach the problem of trying to mimic the brain, we are faced with a number of difficulties. Considering all the different things the brain does, we might first think that it consists of a number of different algorithms, each specialized to do a particular task, and each hard wired into different parts of the brain. This approach translates to considering the brain as a number of subsystems, each with its own program and task. For example, the auditory cortex for perceiving sound has its own algorithm that, say, does a Fourier transform on an incoming sound wave to detect the pitch. The visual cortex, on the other hand, has its own distinct algorithm for decoding the signals from the optic nerve and translating them into the sensation of sight. There is, however, growing evidence that the brain does not function like this at all.
It appears, from biological studies, that brain tissue in different parts of the brain can relearn how to interpret inputs. So, rather than consisting of specialized subsystems that are programmed to perform specific tasks, the brain uses the same algorithm to learn different tasks. This single algorithm approach has many advantages, not least of which is that it is relatively easy to implement. It also means that we can create generalized models and then train them to perform specialized tasks. Like in real brains, using a singular algorithm to describe how each neuron communicates with the other neurons around it allows artificial neural networks to be adaptable and able to carry out multiple higher-level tasks.
Much of the most important work being done with neural net models, and indeed machine learning in general, is through the use of very complex neural nets with many layers and features. This approach is often called deep architecture or deep learning. Human and animal learning occurs at a rate and depth that no machine can match. Many of the elements of biological learning still remain a mystery. One of the key areas of research, and one of the most useful in application, is that of object recognition. This is something quite fundamental to living systems, and higher animals have evolved to possessing an extraordinary ability to learn complex relationships between objects. Biological brains have many layers; each synaptic event exists in a long chain of synaptic processes. In order to recognize complex objects, such as people's faces or handwritten digits, a fundamental task is to create a hierarchy of representation from the raw input to higher and higher levels of abstraction. The goal is to transform raw data, such as a set of pixel values, into something that we can describe as, say, a person riding bicycle.
Resources for Article:
- Python Data Structures [article]
- Exception Handling in MySQL for Python [article]
- Python Data Analysis Utilities [article]