THE VISIONARY >>
START TRAINING
THE MACHINE
What is machine learning? Does it differ from AI or deep learning? Try our interactive workout to get up to speed on one of today’s most talked-about (and implemented) technologies.
Intro to Machine Learning
TRAINING ROOM
Machine learning is essentially any use of algorithms that enable computers to learn from data. If this were a gym workout, algorithms would be any of your physiological attributes that enable you to be healthier, more alert and more energetic. Without data, they’re nothing.
1
OVERVIEW
LESSON
5
4
3
2
Machine Learning
NEXT: LESSON 2
FALSE. As stated in Lesson 1, the term ‘machine learning’ is often conflated with deep learning, but the latter is really just a subset of the former. Study up--the next quiz won’t be so easy.
CORRECT.
INCORRECT.
F
T
OR
Machine learning and deep learning are the same thing.
TRUE FALSE?
Algorithm is a nation
LESSON 1
There are many different machine learning methods, including decision tree learning (essentially a more expansive tree-branch-like version of “if this, then that”), reinforcement learning (broadly speaking, a machine that learns from the consequences of its calculations, trial and error style) and deep learning (uses complex neural networks loosely modeled on the human brain).
That last one is particularly popular at the moment because it is so well-suited to helping build and expand AI applications such as computer vision, where it helps power self-driving cars, face recognition in everything from photo management to airport security, social media analysis, and medical screening.
The term ‘machine learning’ is often conflated with deep learning, but the latter is really just a subset of the former.
Deep learning is powered by artificial neural networks (ANNs), and the two terms can be used interchangeably. Loosely modeled after the human brain and its interconnected neurons, neural networks are fed data (say, pictures of koala bears) and progressively make sense of it by repeatedly dissecting and processing it through various mathematically analytical layers until its predictions (i.e. the ability to identify koala bears in pictures) are more or less accurate.
How many neural networks does it take to screw in a lightbulb?
LESSON 2
Different types of neural networks are used for (and best-suited to) different applications. Neural networks types include Feedforward, Radial basis function, Kohonen self organizing, modular, recurrent and convolutional. The recurrent and convolutional methods are among the most used today, with the latter best applied to computer vision tasks.
Convolutional neural networks (CNNs) are a faster, more efficient way to detect patterns in words or images. In image recognition, for example, a CNN uses a matrix of pixels (called a 'filter') that slides across a part of the image, multiplying the values in each group of pixels by those in an overlapping group to produce a third value (a process known as 'convolution'). By running a series of filters on the same data, a CNN can gradually refine its understanding of the image by detecting edges, then shapes, then more complex objects.
Neural networks were invented within the past decade.
NEXT: LESSON 3
FALSE. The artificial neural network was invented in 1958 by psychologist Frank Rosenblatt, but it didn’t explode in popularity until recently. In the past decade, the increased ubiquity of inexpensive yet powerful computing and exponentially growing mounds of data, along with the exploding AI market demand, have ignited a renaissance in the machine learning method.
Data is a many-splendored thing.
LESSON 3
Most of the world’s data is structured.
NEXT: LESSON 4
FALSE. According to Gartner, about 80 percent of the world’s data is unstructured.
Just as a child learns what, say, a cat is from seeing cats in real life, machines learn from looking at pictures of cats (and presumably a child could, too). In a machine learning context, those pictures might also be called data.
Data, of course, is the basis for any machine learning, and there is no bottom to how specific data examples can get depending on the goals for a machine learning model: pictures of specific people (for facial recognition), CT scans (for medical diagnosis), large moving animals, not to be confused with street lights or mailboxes (self-driving cars), spam emails (spam filters), users saying “OK, Google” (to operate Google Assistant), online behavior (for targeted advertising, to name literally the tiniest tip of the iceberg.
Zooming out from specific topic areas, we can identify two types of data--structured and unstructured--that are used to train machines.
Structured data refers to information that has been organized into readily searchable patterns, the kind of stuff you’d see in databases and spreadsheets: think phone or social security numbers, zip codes, and, to a certain extent, photos and images that have been tagged or labeled. Bottom line: This data can be easily read by a machine right out of the box.
Unstructured data is pretty much everything else—so, Word documents, and text files, but also social media posts, chats, emails, photos, video, and audio recordings. Unless previously tagged, labeled or organized--and big outsourced teams of humans in back offices across the globe are hired to do just that and “structure” inherently unstructured data--it can’t easily be ready by a machine. (One caveat: While you can loosely treat Instagram posts with images and texts as some form of ‘structured data,’ it tends to be inconsistent and fraught with numerous machine learning pitfalls such as implicit meaning, sarcasm, irony, slang, and more.)
When it comes to training machines so that they can be put to work and perform complex AI tasks competently, access to structured data is essential and generally the best and first approach.
A particular grouping of data--say, 2,000 images of people fighting to train a drone AI to spot brawls on the ground or snapshots of people with their eyes accidentally closed--is called a dataset.
6
7
The AI Revolution Will Be Supervised (and also Unsupervised)
LESSON 4
As with data types, general machine learning methods generally fall into three categories. Supervised, unsupervised, and semi-supervised learning.
Supervised learning consists mainly of feeding structured data samples into machine learning algorithms so that they learn from confirmed examples. In other words, if you’re training your neural network to identify the presence of, say, koala bears in photos, then you’ll feed your algorithm thousands of images of koala bears from every which angle, and the neural network will take those images and find patterns in the pixels in order to gradually come up with “koala bear/not koala bear” predictions in pictures.
Unsupervised learning means feeding the model completely unstructured or labeled data and letting the network come up with patterns on its own. This can be useful to trudge through piles of unstructured data, which is most of the world’s data, as well as to do basic “clustering,” or grouping of data into general categories that you may not have thought of. It’s also good for finding anomalies--such as patterns around fraud--that you may not have considered. These findings can often be used for later drill downs using supervised methods.
Semi-supervised learning involves a mix of structured or labeled data with unstructured data, usually more of the latter. It’s particularly useful when getting a large dataset is prohibitively expensive or time-consuming, as the labeled data can be used to automatically label the unstructured data. The outcomes of semi-supervised learning outcomes aren’t always consistently accurate, requiring a decent amount of actual human intervention.
In supervised learning, humans are sometimes directly involved in tweaking algorithms.
TRUE. Though technically supervised learning refers to the use of structured or labeled data, it also means that human data scientists are involved in tweaking the training goals, parameters and datasets and making sense of the outcomes.
NEXT: LESSON 5
Where the student is becoming the master: machine learning’s biggest impacts in the wild
LESSON 5
Medicine: Computer vision-powered medical scanning trained on images of specific medical anomalies, is reducing the instance of false positives. A recent breast cancer study found that machine learning-trained computer vision software by PathA improved the biopsy diagnosis accuracy rate from 85 percent (humans only) to 99.5 percent.
Autonomous cars: This may go without saying, but in everything from current collision avoidance systems in existing consumer vehicles to the fully self-driving cars and trucks of the future, computer vision is playing an integral part. Cars are not only recording driving on roads in all weather conditions on a daily basis across the globe, but some companies are even training models on computer-generated, video game-like virtual driving models to improve self-driving accuracy.
Video: From spotting the presence of logos or winning plays in sports footage to drama-worthy moments in reality-show video, computer vision trained on examples of sponsor logos, fist pumps and other winning moves, and dramatic or romantic activity, respectively, is making highlight videos and sponsorship management easier than ever.
Security: Facial recognition has finally gone big time, though not without controversy. In China, crime suspects have been apprehended from crowds of thousands with closed circuit TVs and high schools are analyzing students’ attention spans for teacher feedback. Meanwhile, airports and airlines across the globe are testing out facial recognition for boarding passes, security and customs clearances, and departure alerts.
Agriculture: Companies such as Blue River have created a series of smart, camera-enabled crop dusters that use computer vision trained on datasets of images of plants in varying states of infestation to analyze crops and decide which ones need pesticides, saving farmers money and reducing the amount unnecessary chemicals in food. It’s no wonder that tractor giant John Deere scooped up the company for $305 million earlier this year, demonstrating the faith that an old-school tractor giant has in a new-school AI idea--all thanks to machine learning.
END OF LESSON
By now you should have a cocktail party conversation level of knowledge to explain what computer vision is, how it’s developed, and how neural networks (a.k.a.) deep networks and A.I. fit into the overall CV picture.
You’ve completed The Machine Training Room’s Computer Vision 101 course.
Watch this space for more courses which will cover some of the cool ways that computer vision is being used by consumers and businesses alike today. For regular updates, make sure to subscribe to The Visionary newsletter.
CONGRATULA-TIONS!
THANK YOU FOR SUBSCRIBING TO OUR NEWSLETTER
THANK YOU!
