DevOps Blog

Stay in the know with in-depth articles about DevOps, micro-services and cloud native topics, delivered to you weekly.

How Does Machine Learning Work?

Written by Marius Rimkus
on May 19, 2020

All businesses rely on data when making critical decisions; even individuals need timely info when they search for pictures, blogs, memes, groceries to buy, tech news, and just about anything else imaginable.

Similarly, machine learning applies artificial intelligence to help devices and systems learn from available data to improve users’ experiences.

If someone, for instance, visits an ecommerce website and they’re recommended products saying “People Who Bought This, Also Saw These Products,” they are watching the site’s machine learning algorithm at work.

All you have to do is take a look at the likes of Google and Amazon to see how this technology is playing a role in enhancing our lives – both personally and professionally. In fact, Netflix saves $1 billion a year by using machine learning to make personalized recommendations – this also keeps a significant number of its users from switching to rival platforms.

There’s plenty of hype surrounding this and other emerging technologies but it’s safe to say that ML is only going to get smarter and better with time.


Machine Learning – How Did It All Begin?

The term machine learning was first used by Arthur Samuel, an artificial intelligence and computer gaming expert back in 1959.  Since then, it’s grown into a data analytics technique that teaches computers to think like humans, enabling them to learn from experience.

It allows computers to develop their own programs without the interference of humans. Machines can take information from data by algorithms; they no longer need to depend on an equation to make decisions. These algorithms increase their performance as the data set they’re exposed to, grows.

It uses pure mathematics to provide calculated predictions based on mathematical formulae.


Benefits Of Machine Learning

It is estimated that 20% of executive-level managers across the world are already using machine learning in some form. It is at the very heart of many advances, attracting nearly 60% of all outside investment from outside the Artificial Intelligence industry.

It attracts brilliant minds from science and business alike because of its ability to represent reality as closely as possible using pure mathematics. It solves real-world problems in a scalable way, disrupting industries by automating processes and providing actionable data.

Live vehicle tracking, product recommendations, image searches wouldn’t be where they are if it wasn’t for machine learning algorithms and models.

With machine learning, Amazon managed to reduce its ‘click to ship’ time to only 15 minutes – that’s how long it takes for them to ship a product once a customer clicks on the buy button!

It has also helped create credit scores, stock and investor trading, face recognition, motion, and object detection as well as voice recognition. In the field of medicine, it’s helping doctors detect tumors and early signs of diseases, research vaccines for different viruses, and DNA sequencing.

Load forecasting for the energy sector and advanced security systems, EBD, cruise control, automatic parking, and ABS cars are all applications of machine learning in the real world.


How Does Machine Learning Work?

Machine learning involves the following steps:

  • Learning From The Training Data Set: Creating parameters and plotting them on a graph to find out the relationship between them. In real life, there are way more than two parameters representing the complex interactions between different data parameters in the training set
  • Measuring Errors: When the model is trained on the set, it should be checked for inaccuracies and problems. There can be four situations here:
    • True Positive: Model predicts a condition correctly when it’s there
    • True Negative: Model doesn’t predict a condition when it isn’t there
    • False Positive: Model predicts a condition when it’s not there
    • False Negative: Model doesn’t predict a situation when it’s there
  • Managing Noise: In real life, many parameters will contain abnormal data that affects the integrity of the set and hinders machine learning. You’ll have to consider data labeling errors, input errors and hidden attributes that are not considered because of insufficient data
  • Testing And Extrapolating: Finding out whether the information is an over-fit or under-fit for the algorithm, through rigorous testing the model and generalizing the results

Machine learning allows several ways to learn from data – your expected output and the kind of input you give determines the right method.

Supervised Machine Learning

This style of machine learning helps make evidence-based predictions with uncertainties. The input and the output for the data are known, and the model is trained by the algorithm to predict responses for new, unknown data. As training progresses, results become more accurate.

Classification and regression are used in supervised learning. Classification can predict discrete responses like whether an email is spam or sent by a genuine user. This is used in speech recognition, medical imaging, and credit scores.

If your data can be categorized, tagged, or grouped, classification is the technique to use. Algorithms used here include Naïve Bayes, K-nearest neighbor, discriminant analysis, and neural networks.

On the other hand, a regression can be used for continuous, non-exact responses. This is used for trading, energy, and electric demand forecasting to see how much power is required for extraction and distribution, respectively.

Algorithms used here are nonlinear, regularization, stepwise regression, and linear regression.

Self-Supervised Machine Learning

It’s where the output is not known, and only the input is given. The algorithm itself takes over and discovers a pattern or structure in the data.

Unsupervised or self-supervised learning aims to uncover the mathematical pattern in the data to learn more about the data.

It can be used for clustering or grouping variables with similar characteristics, for instance, users based on search history. It can also be used for discovering the rules that show how variables in the data sets interact with each other - called association. For instance, people looking for bed sets online will also be interested in table lamps.

Which One Should You Use

Firstly, one size doesn’t fit all. The right algorithm for the job depends on trial and error. Data scientists working in the field can’t tell for sure which algorithm will work without testing it first. Determining factors, however, include:

  • Quantity and type of data you’re working with
  • Insights you’re expecting to get from it
  • How you plan to use those insights

As a general rule, you can choose between supervised and unsupervised learning based on certain conditions. If, for example, you’re looking to make predictions such as the future value of stocks or to spot car manufacturers from webcam footage, supervised learning is suitable.

Unsupervised learning, on the other hand, is better when you want to train your model to split the data into clusters to glean a meaningful pattern from it.

If you need to implement AI and test machine learning, Cherry Servers provides a simple, yet cost-effective cloud infrastructure to help your business grow and scale positively.