How to Teach Your Computer to See Cats
September 20, 2016
Lessons in Deep Learning: What Developments in Machine Learning Mean for Artificial Intelligence
Artificial Intelligence: From Machine Learning to Deep Learning
Recent developments in deep learning promise to revolutionize applications for artificial intelligence. Before examining such applications, we'll begin with a basic definition of deep learning, and then discuss its background in machine learning. We'll then discuss some of the latest deep learning technology trends and how they've shaped our understanding and expectations for artificial intelligence.
Deep learning is an area of computer science research that uses algorithms to create high-level abstractions for use in tasks like recognizing speech, human faces, or languages. Deep learning accomplishes this by learning features, or physical attributes of a data set, through multiple layers of feature learning. For example, a feature can be a shape that the algorithm "thinks" belongs to a common domesticated animal. This process, also known as feature extraction, is also a core part of the broader field of machine learning.
Deep learning emerged originally as a subfield of machine learning and, overall, shifts machine learning closer to the goal of achieving genuine artificial intelligence.1 Although deep learning is often considered more complex, in order to understand deep learning, we must first understand machine learning. Machine learning is, according to Phil Simon, a "field of study that gives computer the ability to learn without being explicitly programmed" with such knowledge.2 This means that machine learning adapts to data; it learns from the material it studies.
Often used in scenarios when a straightforward algorithm won't suffice, machine learning is found in applications like search engines, product recommendations, computer vision, and predictive analytics. Another common application of machine learning is now a basic component of nearly every inbox: the spam filter. The spam filter is an excellent illustration of machine learning. Rather than hand-coding an exhaustive list of every attribute of spam, the machine uses large data sets containing examples of existing spam.3 You "train" your inbox to learn what spam is.
Seeing (Cats) Clearly: Deep Learning Trends and Innovation
Deep learning can be thought of as kind of higher-order machine learning, whose level of complexity becomes ever closer to that of the human brain. In fact, one type of deep learning architecture is known as artificial neural networks. Neural networks use a series of discrete layers, some of which are hidden, which allow for the modeling of complex and nonlinear relationships through different interconnections between data.4
This kind of network, which is modeled closely on the ways neurons interact in the human brain, allows for the creation of high-level abstractions. Such abstractions may take distinct yet related forms while maintaining a certain family likeness. Such a feature, or set of features, may seem to elude the set of strict rules found in traditional algorithms. Although to the eyes of an intelligent agent, such a feature is plainly evident.
Take, for example, a cat. Everyone knows what a cat looks like. But can a computer pick one out of a series of digital images? Defining for a computer all of the various forms of appearance a cat may take—consider not only the different hair colors, sizes, etc., but also how dramatically lighting and environment can change the way a cat appears in a digital photograph—and the task ultimately becomes impossible.
In a recent experiment by Google's X lab, after researchers exposed a neural network to 10 million thumbnail images and 20,000 videos, the neural network began to consistently recognize images of cats. Remarkably, no extra information was provided to the neural network other than what it learned by watching thousands of YouTube videos. According to Google engineer fellow Jeff Dean, they had never indicated to the neural network, "This is a cat."5
But is this true artificial intelligence?
What does Deep Learning Mean for Artificial Intelligence?
Identifying real-world objects is, perhaps, an important first step in developing general-purpose intelligent agents. Deep learning may, in fact, prove to be a crucial step in the advance toward the development of genuine artificial general intelligence (AGI), an agent capable of executing tasks of any variety and at any intellectual level. Many share the conception of deep learning as one part of a goal to move "Machine Learning closer to one of its original goals: Artificial Intelligence.6" So why haven't we seen functional human-level artificial intelligence, or AGI?
Before fully autonomous, walking and talking AI beings roam among us, there will be some important intervening stages. Specifically, we'll see a series of dramatic impacts on existing business and technologies that will be ushered in through AI technologies like deep learning. Importantly, the applications of deep learning and neural networks have already been used for autonomous driving, facial detection, speech recognition, language processing, and recognizing objects or other features embedded in images.
Although some have expressed concern over artificial intelligence stealing human jobs, these applications exhibit clear value for society, automating tasks in aging countries where humans may be less available for such work. For example, speech detection and language processing is used in call centers and storefronts to address such problems. And automated facial detection is used at security checkpoints and monitoring facilities such as large commercial complex and stadiums.
Not surprisingly, deep learning has also attracted deep pockets. Citing deep learning's applications for big data and cloud computing, for example, Intel recently purchased deep learning startup Nervana, for more than $400 million.7 Such an acquisition may be seen as part of a tidal shift toward global adoption of AI-related technology. Deep learning is only one lesson in how these AI technologies are radically changing our world.
1 Deep Learning Tutorials
2 Phil Simon, Too Big to Ignore: The Business Case for Big Data (Hoboken, NJ: Wiley, 201
3,4 What's the Difference Between Artificial Intelligence, Machine Learning, and Deep Learning?
5 Google's Artificial Brain Learns to Find Cat Videos
6 Welcome to Deep Learning
7 Intel is paying more than $400 million to buy deep-learning startup Nervana Systems