A Comprehensive Guide to Machine Learning
By 2020, the digital universe will be 40,000 exabytes, or 40tn gigabytes, in comprehensive size. In contrast, the human brain can hold only 1 million gigabytes of memory. Too much data exists for humans to parse, analyze, and understand. Here is where Machine Learning is finding its value: The raw amount and constant growth of data creates a need for methods to make sense of that data overload in ways that can impact an array of professions and lifestyles.
Although it has many uses, Machine Learning usually gets deployed to solve problems by finding patterns in data we can’t see ourselves. Computers give us the power to unearth concepts that are either too complex for humans or would take us longer to than we’d like to practically use them as a solution.
The first step in Machine Learning is identifying the rules. The automation, or machine part, comes secondary. Rules are essentially the logic upon which we build the automation.
The first step in rule creation is finding the basic breakdown of the data you want to learn about. Think of this area as the labels you give your data in an Excel sheet or database.
Google called these labels “Parameters: the signals or factors used by the model to form its decisions” during a Machine Learning 101 event it held in 2015. A good example here would be working with stock prices to see how different variables can affect the market. In our case, the Parameters would be the stock price, the dates, and the company.
Next, identify the positive and negative results your automation looks to unearth. Essentially, think of this idea in terms of programmatic words such as “True” and “False.” You need to essentially “teach” your program how you, and it, should evaluate your data. Google calls this teaching the “Model: the system that makes predictions or identifications.”
Based on our example we used for the “Parameters,” let’s look at how a basic setup of the “model” would look.
• If we want to see how different variables would affect stock prices, a human being would need to assign the logic of dates and prices with the variables that affected them, such as the upticks in stocks postwar and in conflicts.
• Once you create the basic logic, you take that logic and your data parameters begin to grow your data set you intend to use in the learning stage. At this point in the process you may find your data wanting, and this is the reason to not begin the process data first.
• From here, you run the data through algorithms and tools to solve the logic created. Google calls this process the “Learner: the system that adjusts the parameters — and in turn the model — by looking at differences in predictions versus actual outcome.” In our stock example, our learner would be looking for what variables could have possible impacts on the stocks we’re looking to buy and give us predictions about whether the data suggests that the purchase is a short- or long-term buy.
• Gradient learning, or gradient decline, are an important part of the process from here. We’re talking about small adjustments — not large leaps — that the computer makes over time until it gets the results correct. But you have to watch out for anomalies in the data; they can have huge impacts on the results.
For marketers, we can find a clear use case we can talk about for this basic description of Machine Learning: Google. What would our life be like without it?
Even in its early form, Google took indexed web pages and unstructured data points around them and arranged them based on logic created using the original PageRank. All of this arrangement happened because of tools used to create a result: a search engine results page (SERP).
Marketers have been dealing with Machine Learning in one form or another for some time now. Facebook feeds, Twitter trends, and algorithmic adbuying platforms all use a form of Machine Learning to make decisions easier.
One of the largest fallacies with Machine Learning is that it’ll replace the need for humans. But didn’t we just show you above how humans work among several levels of the process? This idea goes beyond basic data scientists and engineers, extending to people who can shape the problems that Machine Learning will solve, extract the results from the learning process, and apply those results in a meaningful way. As we’ve seen with Google quality raters, humans must also qualify the results and help refine the logic used for learning.
Machine Learning is actually a method for human capital enrichment. It can supe-charge the results achieved by marketers and expand the scope of what we even consider positive results and returns.
We can further break apart Machine Learning into two parts: supervised
and unsupervised learning.
•With supervised learning, you deal with labeled data, and you try to
optimize your Machine Learning algorithm to produce the single,
correct output for each input.
•With unsupervised learning, you deal with unlabeled data, and you let
the model to work on its own to discover underlying patterns.
Are you still with us so far? Good, because now we’re getting to some cool stuff.
Artificial Intelligence, Deep Learning, and Natural Language Processing: They’re shock-and-awe words, but what do they mean? Well, they’re related concepts that have perhaps larger implications on human capital replacement and interaction.
• Artificial Intelligence (AI): We can look at this concept as a computer doing something that humans need intelligence to do. Historically, the Turing test was used to determine whether a computer has a level of intelligence equal to that of a human. The Turing test invites a participant to exchange messages, in real time, with an unseen party. In some cases, that unseen party is another human, but in other cases, it’s a computer. If the participant is unable to distinguish the computer from the human, we say that the computer has passed the Turing test and is considered intelligent.
However, deeper concepts guide AI development today than merely parroting of human language via computers. AI research looks to develop technology that takes action based on learned patterns. You can break down AI into the smaller segments of deep learning — sometimes called neural networks — and natural language processing.
• Deep learning uses the human brain as its design model. It layers brain-like neurons created from levels of Machine Learning. Each level does its learning and produces results that get passed onto the next network that takes on another task with that data. This process replicates itself so that the program can look at one set of data from an unlimited number of viewpoints and create an equally unlimited number of solutions.
Google DeepMind’s AlphaGo program — it made headlines when it beat the world’s top-ranked players at the ancient board game Go — is an example of deep learning in action. The complexities of the game Go means that a computer can’t simply brute force patterns as you can with a game like chess. The computer must learn from patterns and use intuition to make choices. This level of operation isn’t something basic Machine Learning around a base data set can do; however, deep learning allows for the layered neurons to rework the data in unlimited ways by looking at every possible solution.
Deep learning is mostly unsupervised and aims to avoid the need for human intervention. At its core, it looks to learn by watching. When you think about the AlphaGo use case, it all starts to make sense
• Natural Language Processing (NLP): Here’s where computers use Machine Learning to communicate in human language. Search engines and grammar checkers are both examples of NLP. This type of technology is what people most often associate with the Turing test. A quality NLP looks to pass the basics of the Turing test, even though some developers would argue about whether their applications really aim to “trick” users into thinking their applications are humans. Think about it this way: No one believes Amazon’s Alexa is a real person answering questions despite its use of NLP at its core.
In its current and expanding variation, Machine Learning is something that is often seen as a confusing topic. However, as we’ve just described, it breaks neatly into some basic concepts.
• Machine Learning has a clean model for collecting parameters of data that it feeds into a human-created model, which the machine learner then uses to create or find a solution. The basic example we used was the original Google PageRank model for creating SERPs.
• Machine Learning can be unsupervised, often associated with the fields of AI, or supervised, where humans must create the Models and test quality for the findings of the Learners.
• Further, advanced Machine Learning, — often identified as AI — breaks into a few subfields of its own, the most notable of these are Deep Learning and NLP. Deep Learning uses layered Machine Learning in an unsupervised way to approach data from different angles and learn as it moves from layer to layer. NLP uses Machine Learning to communicate in languages humans use every day.
We’ll take a deeper look into several of these topics as we move into decoding the industry jargon associated with Machine Learning and its variants