Machine learning has fundamentally changed the way computers work.
Because machine learning is one of the most well-known evolutions of artificial intelligence, it’s natural that the two terms are often used interchangeably by non-engineers.
However, they aren’t quite the same thing.
The distinction is mainly hierarchical, but keeping it in mind helps to better understand the practical applications of both.
Read on for a discussion of the differences between these concepts and a glimpse of what’s coming in the future.
What is Artificial Intelligence?
Artificial intelligence is one of the fastest growing yet least understood tech trends.
Some people fear AI will take jobs from humans or lead to Hollywood-style robot wars.
The reality is that artificial intelligence is both less fanciful and more intriguing than fiction suggests.
Instead of rendering human workers obsolete, it gives them tools to become more effective and frees them to pursue more highly skilled tasks.
The concept is very broad.
Artificial intelligence strives to create machines that can behave in “intelligent” ways.
Given a large subset of data, AI could make its own decisions about relevance and priority rather than relying on subroutines.
AI-driven processes don’t need predetermined guidelines for every possible situation.
They have the ability to judge situations and take the most reasonable action without needing human oversight.
This is a major departure from non-AI programs.
Even the most highly refined logical algorithm can’t account for the millions of tiny elements involved in everyday tasks.
Consider email sorting.
There are very sophisticated algorithms used in evaluating whether a particular email matters to the account holder, yet dozens of mass advertising messages find their way to the inbox.
Too many variables affect the outcome: sender, content, past interactions, and even date. In its ideal state, artificial intelligence could scan a message’s content and combine that with metadata and other factors to create a “living inbox” where the most relevant emails are always listed first.
AI was divided into specific subfields for most of its history.
Each application was treated as a different subject, and there was little interaction between subdisciplines.
These fields covered diverse topics such as:
- Language processing: Understanding human language
- Computer vision: Identifying and handling unlabelled images
- Genetic algorithms: A method for solving optimization problems using the same process evolution
- Decision theory: The study of how decisions are made and why
Game Changing Shift in Focus
Machine learning is a subdiscipline of artificial intelligence that aims to give machines the ability to learn from previous experiences and use that knowledge in future interactions.
Essentially, a computer is given a pile of data and a machine learning algorithm to process it.
The algorithm sorts the data, adjusting itself after mistakes, until it can achieve the desired results with a high degree of accuracy.
Machine learning does require a data scientist to adjust the model, choose the algorithm, and subset the data.
Still, it represents a huge leap forward in teaching machines to think.
The advent of machine learning was a unifying force on AI research.
It’s wide applicability meant it could be used in many fields, and learning has become a much-desired characteristic of artificial intelligence.
There are two major types of machine learning.
In supervised learning, the algorithm begins with a training dataset of labelled input variables and output variables.
An algorithm is used to map the relationship between variables so that for any new input variable, the correct output can be predicted.
“Supervised” refers to how the training dataset is like a teacher checking the algorithm’s answers against itself.
Unsupervised learning deals with data that doesn’t come with a labelled set of outcomes for reference.
It finds patterns among blocks of data.
Cluster modelling, the most widely used unsupervised learning method, groups data points by their most relevant traits.
It’s often applied to sales data during the customer classification and segmentation process.
Other unsupervised methods include:
- Pattern mining
- Data mining
- Image and object recognition
- Sequence analysis
Machine learning in action
Machine learning has an astounding variety of end applications.
There are too many to describe them all, but they can be broken down into a few functions.
Distinguishing relevant features (Classification):
Machine learning finds patterns within data as well as areas where there are no consistent similarities.
These patterns informs an assessment of the relative importance of the data.
It used to take years for a human worker to sort and identify the relevant features of a disordered dataset.
Machine learning “shakes out” these features in a fraction of that time.
Machine learning excels at recognizing trends in data based on relevant features.
It predicts the classification of incoming data according to past outcomes.
This method- using a model to predict future events- is called time series forecasting, and it has powerful implications for business.
Companies can use insight gained through machine learning to prepare for future disruptions, adjust their supply chain in response to anticipated increases in demand, and decide where to focus new campaigns.
Model selection/Fine-tuning parameters:
For any given artificial intelligence process there are millions (sometimes billions) of factors that affect the process’ operation.
Small changes in these factors can increase or reduce the accuracy of an algorithm’s outcome.
There are too many for a human to manually adjust.
Trying to choose the perfect setting for each would take years.
Machine learning techniques can be used to find the optimal setting for each involved variable.
The Limitations of Machine Learning
Machine Learning isn’t a perfect solution for every problem, of course.
As game-changing as the technology is, it hasn’t advanced to a fully autonomous level yet.
There are limitations to how it can be used.
- Machine learning requires a lot of data.Its nature means that machine learning works best on vast amounts of data.The more data is fed through the algorithm, the more refined it will become.That leads to faster processing and higher accuracy.
Gathering and structuring enough of the right sort of data could present a challenge; at least half of a data scientist’s time is preparing data for machine learning.
This is more of a statistics problem than a machine learning problem, and there’s a lot of labelled training data available for most purposes.
- Most algorithms need to be trained for their intended use.With the exception of Neural Networks and similarly versatile examples, machine learning algorithms have to be directed to a specific application.While the core model may be reusable, experience gained in filtering spam isn’t very useful for image clustering.Refining an algorithm takes time, too.
Machine learning requires lengthy offline training before reaching the point where it adds value.
- Machine learning systems are hard to test and debug.To describe machine learning as complicated would be a massive understatement.As a consequence machine learning systems hard to assess and maintain.Traditional software can be tested for functionality using Boolean-based logic (“This program works as expected”), but engineers use degrees of success when evaluating machine learning (“This algorithm produced 85% accurate results and has improved from the last test by 10%”).
As an interesting wrinkle, it isn’t always possible to be absolutely sure whether machine learning has produced the “correct” result.
Its results are often more indicative of what most people would say rather than what is actually true.
Google’s Director of Research Peter Norvig explains the dilemma: “For some problems, we just don’t know what the truth is.
So, how do you train a machine-learning algorithm on data for which there are no set results?”
What else is out there?
It’s hard to draw a line between machine learning and other artificial intelligence fields like computer vision or natural language processing.
Machine learning has become such a useful way of approaching AI that it’s often incorporated into other applications.
Essentially, artificial intelligence systems that can learn from their mistakes and new data involve machine learning.
There are systems that exhibit “intelligence” without learning on their own.
An example would be an expert system – software programmed to function as an expert in a specific domain.
Expert systems use rules, probabilistic reasoning, and logic to reach conclusions rather than relying on past experience.
They’re capable of providing advice, solving problems, demonstrating processes and explaining their logic, and predicting results.
They have trouble working around gaps in its knowledge base, however, and don’t learn or refine themselves.
The Future of Machine Learning
Just as machine learning is the natural successor to artificial intelligence, Deep Learning is on the cutting edge of machine learning.
It’s the next logical step.
Deep Learning deals with neural networks – algorithms designed to mimic the function of the human brain.
These aren’t the primitive neural networks of the 90s, though.
Scale is of paramount importance.
Deep Learning neural networks are huge, fed with as much data and spread across as many machines as possible.
The more layers and data incorporated into the network, the more accurate the results.
There’s a lot of hardware and training time involved in bringing them to a functional maturity.
In essence, Deep Learning is machine learning on an epic scale.
Distinguishing artificial intelligence from machine learning is like differentiating between automobiles and electric cars.
It’s a matter of succession and inclusivity.
In other words: all machine learning is artificial intelligence, but not all artificial intelligence is machine learning.