AI’s Next Mission
by Yuchi Huang, ACTNext
Today you can read any publication from Time magazine to peer-reviewed scientific journals and you will find debate about the benefits and future of artificial intelligence (AI). Some experts think it will save the world. Others think it could be our downfall.
For example, a recent article by Professor Melanie Mitchell in The New York Times, Artificial Intelligence Hits the Barrier of Meaning, presents a typical view of the future of AI. It lists many examples of the failure of current AI applications and claims that we underestimate the difficulty in creating a machine that is as intelligent as humans. The author believes that to achieve a real breakthrough, we should introduce human-like understanding into machines through study of human cognition
As a practitioner of machine learning and AI, I am familiar with the frustrating failures detailed in this article and others. I don’t even deny that almost all AI applications are flawed to some extent. However, I believe there is great potential in AI and through continued investment it has the potential to change our world in positive ways.
AI is indeed in its infancy, but early AI has already begun to make people’s lives much more convenient or greatly increase productivity in industry. You can ask Amazon’s Alexa to play music, request recommendations, order coffee at Starbucks or book a ride using Uber or Lyft with voice commands. You can even ask what song is playing on the radio (if you want to know the name so badly).
Another example is Google’s translation app: nowadays tourists could be bold enough to visit a foreign country without knowing a word of the local language. By using this app, you are allowed to listen to or speak in over 90 different languages; you can even use embedded camera-based translation to understand foreign road signs and notices. Although such systems may produce errors, you can still infer the correct information from imperfect results.
Mitchell gives some examples of the shortcomings of AI and she’s not wrong. Clearly you can trick speech recognition through the use of homophones and homographs or get phrases that are lost in translation. We’re not even close to grasping idiomatic speech. It is easy to find examples of document reader mistakes and self-driving cars have a long way to go.
Serious scientists have never underestimated the difficulty of AI. The constant challenge and reflection of past research has formed the driving force for endless improvement. The facial recognition technology of several years ago cannot differentiate between a real face and a photo (or a clip of video) of a face; with unremitting efforts on the development of the anti-spoofing technology, It is much harder to fool the state-of-the-art high-end facial recognition system by disguising a specific person using only photos or videos. You may still deceive it by wearing a realistic latex mask scanned and 3D-printed beforehand (like Tom Cruise did in Mission Impossible) but that is quite expensive to make.
The term Machine Learning is apt and Mitchell has a point. We are not even close to Deep Learning, or what might be called simply “learning.”
Click-bait articles overplay the significance of many aspects of AI and its development. One possible reason is that entrepreneurs in this field exaggerate the prospects of AI on media to attract enough eyeballs and investment.
In the absence of an understanding of the difficult knowledge in this field, people are willing to believe in these promises and ignore the potential obstacles to technological advancement.
Twenty years ago, over-optimism for the Web technology generated an Internet bubble. Some similarities may be found between 2000 and today’s enthusiasm for AI. Is this dot-com déjà vu? We don’t know; but we do know is that we have indeed entered the Information age in the two decades after the dot-com era. Likewise, AI will continue to develop, although the progress may move forward with zigs and zags.
We’re using blunt tools, the steam shovels of our day, to excavate meaning from vast amounts of data. In that respect, machine learning has much value.
Mitchell asks “What is the nature of intelligence?” But a better next step question might be: “How we can learn more from data?”
We apply self-correcting algorithms to mountains of data. To really improve AI, research on human cognition is certainly a path, but not necessarily the only one. Animal and human intelligence evolved and improved over hundreds of millions of years; the biological foundation is carbon-based life. On the contrary, computers have different physical compositions.
The research progress of biological cognitive systems is quite slow, and our knowledge in this area is still extremely limited. Even if we understand more about the relevant principles, can this knowledge guide us to build intelligence on a different material basis? Not necessarily.
Humans think through electrochemical reactions in brains; but to make our machines think, we might have to invent a different type of thinking. In addition to algorithm improvements, what AI research has achieved so far relies on the tremendous increase of computing power and storage capacity. Another major breakthrough in artificial intelligence could be triggered by revolutionary changes in computer architecture such as quantum computers.
What’s next for AI? It is truly hard to say what we should expect from an area that is advancing with unlimited possibility. AI techniques driven by deep learning algorithms have already proved their capabilities in well-defined perception problems (such as visual/speech recognition). The next big challenge is to push the boundaries of AI in reasoning tasks, such as using common sense, dealing with changing situations, planning and making decisions in virtually every industry: Healthcare, Finance and Education. To gradually overcome the limitation of AI systems and avoid bias, one solution is to place humans — especially subject matter experts in the loop and take advantage of knowledge provided by experts in specific problem domains. As an AI researcher working for Education, I believe the combination of human reasoning and machine learning will make the previously impossible possible.
