Q&A: An Introduction to Deep Learning
We explore what deep learning is and the benefits it offers, its relationship with AI, and where it’s headed with Martin Ford.
- By James E. Powell
- January 2, 2019
In his new book, Architects of Intelligence: The Truth about AI from the People Building It (Packt Publishing, November 23, 2018) Martin Ford interviews AI experts about the future of AI, its impact on society, and what should concern us about the technology. Upside spoke with Ford about his perspective, AI’s benefits, and where AI and particularly deep learning are headed.
Upside: In a nutshell, how would you describe deep learning to someone unfamiliar with the technology? How is it being used?
Martin Ford: Deep learning refers to software that roughly emulates the way biological neurons in the brain operate and interact. Neural networks were invented in the 1950s, but it is only in the last decade that sophisticated versions of these networks with many layers of neurons have been built. The word "deep" refers to the fact that these networks have many layers of artificial neurons (in some cases over 100).
Deep learning has been the primary technology that has driven the remarkable progress we've seen in AI over the past few years. Deep neural networks power the speech recognition and generation that makes Amazon Alexa possible. It has also revolutionized machine vision. Computers can now outperform humans at recognizing visual images, and this power is already being deployed in applications ranging from self-driving cars to medical systems that can diagnose cancer based on visual analysis of images. The technology is also being used to give robots better perception and dexterity. The number of applications for deep learning is virtually limitless in areas such as medicine, science, business, factory automation, and transportation.
The technology continues to advance rapidly. This is occurring in part because computer hardware is getting faster and also better. Companies such as NVIDIA and Intel now produce computer chips that are specifically designed for the technology. There is also a great deal of research being done on improving the learning algorithms used in deep learning. Huge tech companies are making massive investments in the technology and this will in all likelihood continue to drive progress going forward.
What’s the relationship between deep learning and AI? How does DL advance AI?
Deep learning (and more generally machine learning, which includes approaches other than neural networks) is just one branch of artificial intelligence. However, it is far and away the single area that is experiencing the most explosive progress.
There is actually something of a debate among top experts about the role and importance of deep learning in AI going forward. Architects of Intelligence includes interviews with the pioneers of deep learning -- including Geoff Hinton, Yann LeCun, and Yoshua Bengio -- and they tend to believe that neural networks will ultimately prove capable of doing everything that really matters in AI. Other experts disagree. They think deep learning will have to be combined with other approaches in order to continue to drive the field forward.
What are the benefits of deep learning? What are the drawbacks?
The primary benefit of deep learning is that systems have power: the (often super-human) ability to recognize patterns in data. The patterns might be objects or people in a visual image, words in human speech, important insights in business data, or an early warning that a particular part in a machine is likely to fail.
There are a number of potential drawbacks with the current state of the technology.
- Deep learning systems are relatively inflexible or "brittle." If the initial assumptions change, the system will generate incorrect data. Unlike humans, these systems cannot adapt.
- Deep learning systems can lack transparency or act as "black boxes." They give results but no explanation as to how those results were produced. This can be a big problem when the systems are used in areas such as criminal justice, where it is critical to understand the rationale that leads to decisions.
- Bias on the basis of race, gender, or other parameters has been detected in some cases. This arises because of bias in the data used to train the system. For example, if a neural network is used in a facial recognition system and the network is trained on mostly white faces, then the system may be less effective at recognizing non-white faces. This could result in "false positives" so that non-white people are more often incorrectly identified.
- Security is always a major concern with any autonomous systems, and there is evidence that deep learning systems may be susceptible to hacking.
Increasingly, regulations require enterprises to explain their decisions, and in AI that may be difficult because the algorithms are hidden from view – hence the push for algorithmic transparency. Is DL part of the solution?
The tendency of deep learning systems to be "black boxes" is a major concern in the field, and there is a lot of ongoing research into building systems that are more transparent. Companies such as Google have teams working both on transparency and to address bias in these systems.
One of the people I interviewed, David Ferrucci (who led the team that built IBM Watson), has started a company specifically focused on building systems that can explain themselves. He is utilizing the latest advances in natural language processing and ultimately envisions a system that can answer question in a way much like a human analyst might.
What are some common misconceptions about DL? What do people think it can do that it can’t?
The most important misconception is that people associate true intelligence with deep learning. These systems are currently limited to being extremely effective pattern recognizers. They do not exhibit anything like true human intelligence. This misconception can be amplified when the media describes neural networks at "brain-like" computers.
Although many of the people I interviewed (including, for example, Demis Hassabis of DeepMind and Ray Kurzweil at Google) are very interested in making progress toward true machine intelligence, most researchers agree that this remains a long way off.
For now, DL systems can do very specific things with amazing proficiency -- but they are not "thinking" like people.
Where is DL headed in the next year or two? Where is it headed longer term (say, in 5 to 10 years)?
The next year or two will see continued breakthroughs in the core technology, but perhaps more important, an explosion of applications for DL in sectors throughout the economy. One of the people I interviewed, Andrew Ng, has several initiatives designed to accelerate this progress: Landing.ai is bringing deep learning to manufacturing, while the AI Fund is incubating new ideas that will eventually be scaled up to startup companies focused on a variety of areas.
Over the next 5-10 years, DL will likely evolve into a true general-purpose technology. It will be almost like electricity, and it will become increasingly indispensable in nearly every aspect of the economy, science and culture.
We should also be prepared for the possibility of truly disruptive breakthroughs over this longer time frame. Ray Kurzweil, for example, believes we may be able to achieve human-level machine intelligence within 11 years. That is an aggressive prediction, but we should expect remarkable progress with huge implications for the economy and society. The potential for a great many more routine and predictable jobs to be automated is one area that I have focused on a great deal both in my previous book Rise of the Robots, and in Architects of Intelligence. I believe we need to begin a conversation about how we can adapt our society and economy to the implications of these technologies.
What best practices can you recommend to enterprises today to prepare for DL?
First, ensure that executives have an understanding of the technology. There are many resources available. Andrew Ng has just introduced a new course called "AI for Everyone" which is focused on this challenge. Business managers need to understand that, in the near future, not leveraging DL will be like not using electricity -- it will be a clear recipe for failure.