The Verge published an online article on October 10th entitled, “These are three of the biggest problems facing today’s AI,” where author James Vincent interviewed three influencers in the field of robotics on the future of Artificial Intelligence. There seems to be a rather resounding consensus amongst the experts: AI has yet to be perfected for three reasons: one, artificial intelligence requires a ton of ever-changing hard-to-get information; two, AI is unable to multi-task; and three, there needs to be more focus on how these systems reach their final conclusions.
Similar to a consumer waiting for prices to drop when the first flat screen hit the market, should we be playing the AI waiting game prior to purchasing humanoid robots like JIBO – the world’s first family robot? According to Neil Lawrence, Professor of machine learning at the University of Sheffield and part of Amazon’s AI team, “these systems don’t just require more information than humans to understand concepts or recognize features, they require hundreds of thousands times more.” Lawrence says that huge tech giants like Google, Facebook and Microsoft are perfect resources for AI. “They have abundant data and so can afford to run efficient machine learning systems.”
In the realm of collecting an abundance of data, Lawrence mentions that it becomes exponentially more difficult to harness information in specialized fields like healthcare. AI, for example, is being used for machine vision tasks such as searching for tumors on x-ray scans. Lawrence, referencing the difficulty in securing data in healthcare said, “It’s generally considered unethical to force people to become sick to acquire data.” Assuming that there is a wealth of information available through medical research and clinical trials, I was a bit confused as to why AI needed real-time information on the “sick” in order to be effective with AI vision tasks on x-ray scans. Isn’t there a way to download the wealth of information already available in the x-ray scanning process with regard to identifying tumors? Regardless, data collection will be instrumental in moving AI forward and there’s no better resource than the search engine tech companies, like Google, to generate real-time information to move AI forward.
The second major problem, according to Google DeepMind research scientist Raia Hadsell, is AI’s inability to multi-task. Compared to an idiot savant, “there is no neural network in the world, and no method right now that can be trained to identify objects and images, play Space Invaders, and listen to music,” Hadsell said. With Neural Networks being the building blocks for deep learning systems, there has yet to be a network capable of mastering multiple tasks at once. The solution, according to Hadsell, “connecting separate deep learning systems together so that they can pass on certain bits of information.” In February of 2015, Hadsell’s DeepMind team at Google announced it had created a system that could beat 49 Atari games. Some considered this a breakthrough but according to Hadsell, it reiterated the fact that AI is still way off in the distance. “Each time it beat a game, the system had to be retrained to beat the next one. We can’t even learn multiple games.”
Finally, there needs to be greater insight into how AI reaches its conclusions. In other words, instead of being spoon fed the meaning of objects, systems need to interpret their surroundings objectively. According to Murray Shanahan, Professor of Cognitive Robotics at Imperial College London, “what goes on in the mind can be reduced to basic logic, where the world is defined by a complex dictionary of symbols. By combining these symbols – which represent actions, events, objects, etc. – you basically synthesize thinking.” This summarizes Shanahan’s “Deep Symbolic Reinforcement Learning” methodology that he hopes will break some of AI’s barriers when it comes to deep learning.
In closing, experts are estimating 30+ years all the way to a century before artificial intelligence is close to where it should be. As it stands right now, AI consists of feeding systems data and waiting for them to notice patterns. Deep learning, although in its innovation infancy, happens to be the key ingredient in AI’s breakthrough.