Current AI systems are described as Artificial Narrow Intelligence. They can be trained to execute a specific task which is carefully scoped. This is sometimes described as Weak AI but the execution speed of any AI far exceeds what a human can accomplish in the same time. This amplifies human abilities, but outside of the defined task it cannot replace human intellect and has no insights into solving new problems.

Machine Learning is founded on statistical analysis principles. Patterns emerge when a large corpus of data is analyzed to discern trends and characteristic groupings of data values. Taking new data and fitting that to the distilled analytical results suggests inferences that characterize the new data.

What Is Machine Learning (ML)?

The general public perceive AI as one monolithic concept but we already know that the term ‘AI’ embodies many disparate technologies and applications. Those different facets of AI are designed to solve different problems. There is (as yet) no general purpose solution that solves everything.

Machine Learning is an important aspect of AI. It is easier to understand as smaller component parts. Machine Learning trains models that are designed to recognize and categorize new input data based on what they have seen previously. Those models are built with a range of statistical algorithms and training methods. Machine Learning is based on three fundamental concepts:

In AI literature, Machine Learning is often referred to by its initials ML. Let’s use that convention from now on.

Classifying data into different categories is unlike trend analysis or curve fitting a line to arbitrary data but they are both well-known statistical techniques that ML uses.

These are the key topics to understand when learning about ML:

• Who is responsible for each part of the architecture?
• Training methods.
• Data preparation and labelling.
• Parsing algorithms.
• Internal workings.

Roles & Responsibilities

A new term (MLOps) has evolved recently to describe the responsibilities of ML administrators and operators. Here it is in context:

Training Methods

Models need to be trained with one of several alternative methods. That process feeds large amounts of data through a parsing process that classifies it. Enough information is retained inside the model to define patterns that describe the data.

Three of the training methods are basic and easy to understand. Training methods can also be combined and there are other methods that use more advanced techniques:

Let’s examine the basic training methods in more detail.

Supervised Learning

The results from ML systems are based on training the model with pre-labeled data. This creates stored pattern descriptions inside the ML model.

The parsing algorithms are based on statistical mathematics that use classification and regression techniques. Supervised learning requires human intelligence when labeling the training data. MLOps staff will add more training content later so the ML system will understand currently out-of-scope data. There is no inherent ability for the ML to self-train – yet.

The label provided with the training data is returned as a result when new content is matched against the learned patterns.

Unsupervised Learning

More ambitious systems use Unsupervised learning. This can recognize patterns in unlabeled data. The statistical techniques locate clusters of similar data and can also identify anomalous values that don’t fit any of the clusters.

Unsupervised learning must be verified by a human operator to ensure there is no bias in the results. Facial recognition systems have historically struggled to perform consistently across diverse populations, a limitation attributed to training datasets that were initially trained on large numbers of Caucasian facial data and not sufficiently representative.

Incomplete training data will compromise the ability of an AI system to recognize the desired patterns.

Semi-supervised Learning

Semi-supervised learning is a combination of Supervised and Unsupervised learning. This involves an initial supervised human driven manual training process followed by unsupervised self-learning on a larger set of data.

The model is initially trained in a supervised fashion with labeled data to seed it with fundamental ‘knowledge’. This is a practical solution to situations where the training data set is too large to be labeled beforehand. Only a small sub-set needs to be labeled.

Data that fits the labeled criteria can be assimilated without human intervention. Anomalies can be flagged and a human operator can then create new labels as needed. Overall, this is quite an efficient approach but still requires verification to avoid bias. Some commentators suggest this leads to unreliable models being created.

Reinforcement Learning

The remaining kind of basic level training is called Reinforcement learning. This rewards or penalizes the matching results based on their accuracy. Since the machine being taught has no emotional sense of reward or pleasure itself, this is implemented using a scoring or ranking value. Good matches increment the value while bad ones decrement it.

The scoring of results happens in the external environment that provides a context for the model. That scoring may be the result of a human ticking a checkbox to accept or reject the result.

Reinforcement learning simulates human thinking processes and adds primitive memory capabilities to ML systems.

Later on, we will see how that works in an object oriented implementation.

A Taxonomy Of ML Algorithms

These are the fundamental algorithms used in ML systems. They are not to be confused with training methods. Algorithms are used inside the ML engine to construct the model whereas training methods describe how the data is presented:

Statistical regression algorithms are particularly useful for general purpose ML. Because they depend to some extent on human provided labelling they are ideally suited to supervised training methods such as:

• Linear regression.
• Logistic regression.
• Naive Bayes.
• Decision trees.
• Random forest.
• K-Nearest neighbor.

The separate K means algorithm is based on the nearest neighbor technique and can be used by unsupervised methods after some initial training has taken place. We will break these algorithms down here, but refer to the appendices for a longer list of the different parsing algorithms and training methods.

Linear Regression Algorithms

Classic graph drawing helps to illustrate how this parsing algorithm works. Given a set of data points in X and Y, a scatter chart can place them on a grid:

If a new point is introduced, placing it in the graph at its X-Y co-ordinate location identifies the nearest neighbors. They have similar characteristics. The distance away from the diagonal trend line could indicate the confidence factor for how well it matches the regression. That trend line can be extrapolated in either direction to predict other results.

Logistic Regression Algorithms

Logistic regression techniques categorize data into two discrete classes with a binary splitting process that sorts the data into one of two different buckets.

First of all, establish the range of the data points. This may require an offset as well. The range is based on the maximum and minimum data values minus the offset to the minimum. Now, determine the cut-off point. Any values larger than the cut-off-point are members of the upper set and the rest are in the lower set. The cut-off could be any value and the median value is only used here for convenience.

This could be used to detect audio sample values that exceed a pre-determined volume level in an editing tool – useful for locating the loudest passage in a recording for automatic gain control settings.

Other parsing algorithms can be used when more than two discrete categories are required.

Naïve Bayes Algorithms

This algorithm matches the logistic regression data for two or more categories. Bayes theorem uses conditional probabilities that are independent of one another and can return multiple resulting values.

Those results should carry a confidence factor. This is important because the business logic that requested the analysis needs to examine the confidence factors to decide which is the most reliable.

I once saw a demonstration where a photograph of an American president was passed to the AI system. It correctly identified the individual by name with a 95% certainty. It also determined that the photograph featured a male with only a 65% certainty. Inference from the more confident ‘name’ result would deliver a better gender result. This did not happen automatically inside the ML system.

Other hybrid implementations might look for color combinations, size and shape of objects in the scene and any metadata in the image container. A combination of these might yield a more reliable result.

Always check the confidence factor for the result when there are several outcomes and use the most reliable result.

Decision Tree Algorithms

By nesting the selection criteria, a decision tree algorithm implements a more sophisticated result generator based on multiple criteria. In the diagram below the blue boxes are selectors that choose one of the possible outcomes based on their input stimuli. The small circles are the choices. In this abstract example there are several alternative routes to the same results:

Rule based designs can only get you so far in building ML systems. They require additional mechanisms to store the training data and classify it using statistical techniques. Hybrid training methods might be able to alter the decision tree rule set based on the statistical findings.

Random Forest Algorithms

This parsing algorithm is described as a forest because it combines many decision trees into a single process. There may be thousands of individual decision trees in a random forest implementation.

This technique is resilient to one of the problems with AI data training called overfitting. This happens when an algorithm tries to coerce everything to fit a training data set that was too small. New data that is not within the scope of the training data set will deliver inaccurate outcomes.

K-Nearest Neighbor Algorithms

This approach looks for adjacency to clusters of trained data. Since it is possible to define the properties of any axis to describe our data the surface can represent abstract ideas. In this example, semantic tokenization defines the Y-axis and the frequency that the tokens occur in a data set defines the X-axis.  This would create clusters of semantically similar text blocks.  The clusters represent theme or topic islands.

Introducing a new body of text is subjected to the same analysis algorithm. This determines how close it sits to any of the clusters. Items located inside a cluster are described as known outcomes. Items that are close to a cluster but outside it are unknown but classifiable. The cluster outline could be enlarged to encompass them if they are redelivered with a matching label during a training pass.

This algorithm relies on vector algebra to calculate the distance to the nearest cluster. It is useful for Instance Based Learning methods where it depends on prior instances of ingested training data.

It is conceptually possible for clusters to overlap. Any items located in the overlapping area would return two classification values as a result.

Algorithmically, this technique is similar to polygon area filling functions that determine whether a point is inside or outside a shape. Performance can be increased by using other computer graphics techniques to trivially eliminate matches that fall outside of the extent rectangle for each cluster.

The name of this algorithm may be abbreviated to KNN in user interface dashboards.

K Means Algorithm

This is very similar to the K-Nearest Neighbor algorithm but in this case the distance vector is measured from the new data item to the centroid of the cluster area. In physical terms, the centroid is the center of gravity about which an arbitrary shape will rotate without imposing any directional forces (wobble) as it spins. This is not the same as the center point of an extent rectangle surrounding the cluster.

This algorithm is well suited to use in unsupervised training methods.  The distance away from the centroid location is a vector that could describe a confidence factor for the matching result.

How Does It Actually Work?

A common misconception is that the ML engine must store all the multi-terabytes of training data that is used to feed the knowledge base and then use some really clever and very fast ‘thinking’ processes to scan it all and determine the results. This is not true at all!

Actually, the learning process establishes rules and algorithmic constructions for matching patterns instead of storing the actual data. The only data that is stored is a compact pattern description of the distilled data from each training pass.

Conceptually, this fits very nicely with the object oriented paradigm where the algorithm is embedded in the class and the values derived from each training pass are stored in an instantiated object.

Given a technique such as logistic regression, the minimum, maximum and cut-off points are stored with a textual label as object properties while the unchanging algorithmic processing is embodied in class-based static code.

The objects are serialized and stored in an indexed in a database so they can be examined more easily. Objects can be instantiated from these database records to perform tasks and interact with other objects inside the model when necessary.

When a new data item arrives to be characterized, the object catalog in the database can be traversed to compare it with results of the prior training methods. An existing object whose properties potentially describe the new data item yields a match. The corresponding training data label is returned to identify the new data item as a member of that category.

Although this is a simplified description, it suggests that the internals of an AI system are not so complicated after all. They are simply very large repositories of pattern matching descriptions.

Applying ML

Machine Learning can assimilate large amounts of data and distil it to produce patterns. Since media is just another kind of data, ML is useful as a solution in these contexts:

Conclusions

Artificial Intelligence and Machine Learning are big subjects with a lot of fine detail. How far you choose to dive into the detail depends on whether you are improving your foundational knowledge to ensure you understand the systems you are deploying, and the discussions that should be had with systems providers… or you are considering building your own AI powered systems.

For those wanting to dive deeper and explore building systems; Divide your learning process into smaller parts. Each concept that you learn about is another piece of the overall jigsaw puzzle. That makes the subsequent items easier to understand. Learn the terminology and understand what all the buzzwords actually mean. That cuts through the obfuscation.

Start with a basic understanding of what ML and Deep Learning are useful for and when it is appropriate to use them. If you need to understand what the models are doing internally, study the basics of probability theory, statistics and linear algebra.

For practical applications and building models, many of the tools are developed in the Python language. If you already know some coding languages, Python will be somewhat familiar territory. For deeper level engineering, learning the R language will be useful. Find out about the open-source Hadoop framework for large, distributed storage systems.

Once you have mastered the basics, start to develop your knowledge of the learning algorithms that support supervised and unsupervised systems. This will help you tune your models to fit your requirements more closely. The appendices at the end of this book list many of the commonly used algorithms. There are others. Selecting the most appropriate algorithm and training method are important.

These Appendix articles contain additional information you may find useful: