While machine learning has been around for a long time it has only been in the latest increases in computing power that we’ve seen machine learning become more applicable to smaller companies and startups.
With this rise, it can be easy to dismiss machine learning as a too-hyped tech that will fall away in a few years. In one of our recent conversations with a client they mentioned they knew “machine learning is a hot topic right now and we don’t want to get caught up in the hype.” In reality, machine learning was the answer to their issues rather than just hype.
A lot of this confusion surfaces around the scope of machine learning. While there is a lot of hype around deep learning, it is just a subfield of machine learning. The simplest definition one can give is: if you are using statistics to solve a problem you can reasonably argue you are using machine learning to solve your problem.
In machine learning there are two main “camps” that a method or algorithm will fall into.
Classical Machine Learning
In the world of classic machine learning a machine learning engineer / data scientist would take data and find features of interest and possible engineer new features from existing data. They would then find the best model and parameters to get the best predictions, or classifications, for new data coming in.
Popular algorithms in this field would be Logistic Regression, Naive Bayes, and Decision Tree to name a few. Solutions for the world of classic machine learning can be sentiment analysis, spam filtering, fraud detection, and forecasting. The great thing about classical methods is you can start with a small amount of data and get decent results in most cases. Enough at least to get a proof of concept up and running while more data collection takes place.
If there is any hype around machine learning it is most certainly around the subfield named deep learning. This is where we use a neural network using some linear algebra and calculus to solve many of the same problems as classic machine learning. The leverage that deep learning gives us is that with enough data features, engineering is not needed. Over time the model will find commonalities in the data. This is great as it allows startups and other companies to implement machine learning models with a single data scientist or machine learning engineer instead of a full team.
Deep learning also has the ability to use what is called transfer learning. This is where we take a model from Google that is trained on all of their data, then specialize the model to look at the kind of data we are looking at. This can help lower the amount of data needed and save time in training a proof of concept model.
While the world of Deep Learning is an interesting one, many companies’ problems can be solved in the area of classical learning. Thanks to the continued drop of computer power, building platforms capable of running thousands of inferences cheaply, or even offloading this inferencing to mobile devices, is actually possible now. It is this lowering cost we are seeing more companies and startups asking how machine learning could give them a competitive advantage.
This isn’t to say that everything in the world of machine learning is going well. The compute cost needed to push the industry forward continues to increase, making it harder and more expensive to be on the bleeding edge of research. Do not let this scare you away from machine learning, though. Even new and evolving complex problems like object detections can be implemented in a proof of concept quality for what is relativity a small investment even a small startup can afford.
In machine learning, we have three main groups of algorithms.
Simply defined, this kind of learning is when we give our algorithms an answer key during their training. The two main areas in this field are classification and regression.
Here we are asking a model to train with no answer key and are mostly looking to find like items. There are three main areas in this subfield; Clustering (group similar items), association (find sequences), and dimension reduction (find hidden dependents).
In some systems, you will find unsupervised models grouping items so that later, when an answer is provided, the label can be applied to a group of images and a supervised model can be trained on the results. An example of this is face detectors on sites like FaceBook or Google Photos.
Most of the bleeding-edge research today takes place in this field. The idea in this arena is you give AI some constraints and scoring and allow them to learn over time what is good and bad. In the world of self-driving cars this can be overly simplified as “Stay on the road +10 points, hit a curb – 20 points, ” and the agents are programmed to try and achieve a high score. While this is mostly talked about in the world of self-driving cars and games, a more realistic place to find these in the wild is automated stock trading or enterprise resource management. Although take a look at AlphaStar to see a reinforcement AI that was created by Alphabet’s Deepmind to play StarCraft against players online.
Like in many fields there are specializations. The two main ones are computer vision and natural language processing. Currently, in the world of machine learning, natural language processing is getting the most attention. The world of chatbots and AI assistants continues to be a big area of funding for large tech companies as they try to create the ultimate helper to keep you using their system.
Computer vision itself is not to be overlooked as AR and VR continue to gain steam and on-device computer vision algorithms gain in popularity. While computer vision and natural language processing are very different in terms of what you need to know to be successful, the two paired together can create amazingly powerful tools. One that is always brought up is Google’s translator application that can not only read what is on signs but can actually put the translation over the sign, in realtime.
With the lowering cost of powerful hardware and knowledge requirements needed to create machine learning solutions, it is no surprise machine learning has been taking off in recent years. The large tech companies now have a data scientist embed on every project to see if machine learning can give each of their projects an advantage. However, you no longer need a large tech company’s R&D budget to leverage machine learning at your company. Here at Big Nerd Ranch, we are here to help, be it discovery, building proof-of-concept, or building out a production machine learning pipeline. We can give your company a competitive edge and bring delight to your application that feels like magic.