Our CTO, Ahmad Abdulkader, was recently interviewed by Byron Reese as a part of GigaOM’s podcast, Voices in AI. Below are a few excerpts from Ahmad’s interview with Byron. How would you define AI?To me AI is the ability of machines to do or perform cognitive tasks that humans can do, or learn to do rather. And eventually learn to do it in a seamless way. The most important of these that we actually use almost every second of the day are vision, speech understanding, or language understanding.How are you defining importance? If I had an AI that could diagnose any disease, tell us how to generate unlimited energy, fix all the environmental woes, tell us how to do faster than light travel, all of those things, like, feed the hungry, and alleviate poverty and all of those things…I would say that’s pretty importantTo do these tasks that you’re talking about, it probably implies, that you have done or solved, to a great degree, solved vision. It’s hard to imagine that you would be doing diagnosis without actually solving vision. So, perhaps the utility of what you talked about would be much more useful for us, but if you were to define importance as sort of the basic skills that you could build upon, I would say vision would be the most important one. Language understanding perhaps would be the second most important one. And I think doing well in these basic cognitive skills would enable us to solve the problems that you’re talking about.How do you think humans do transfer learning so well, and we have such difficulty getting machines to do it?The brain does a very good job at capturing and abstracting, and only then you need just a very few examples to learn what you’re trying to learn. I think [over] the next few years, the next big challenge for artificial intelligence is unsupervised learning[, as done in strides]. You can see that most of CIFAR-10 is an unsupervised version of computer vision. The next version of ImageNet is going to actually be smaller or maybe the same size, they just supervised a set and a much bigger unsupervised set. I think this is critical for us to actually do reasonable strides in AI, do the significant improvements in unsupervised learning, and even cloning it correctly.Do you believe we will be able to develop strong AI without understanding how it is that we are intelligent and how the brain works? Is cracking that riddle part of strong AI?We will definitely be inspired by it. We are discovering things about the human brain that actually helps us build better AI models. It’s not a condition, I don’t think we have to fully understand how the brain works to achieve that, but definitely discoveries like this enlighten us big time. We’ve known, for example, that the number of feedback connections in the brain are way more than the feed forward connections. Do our models actually exhibit that? Not many of them, and only recently we had the CPU power and the data to show that they actually add value. For a long time, most of our neural networks did not have any feedback connections. It’s things that we discover about the brain can help us I think, and deliver better models. …there are two critical aspects that we need to think about and try to incorporate in the models that we’re going to invent for the future:
  • Current models, they work in discrete time, there’s no continuous time. The brain is not like that, there’s no concept of discrete time, it’s a continuous thing. It’s analog. Maybe our models need to be analog.
  • Also, the human memory is really weird, we fill lots of the gaps. We’re able to abstract things, it’s exhibited in associative memory, like once I start to talk to you about a story, you somehow fill the rest of the details yourself from your past experiences, and you don’t really need to store all of the details. So, there’s a very, very efficient way of storing things, memories, and pictures and our notion of certain concepts and people. It’s stored in a very, very efficient way… something that we’re not doing in our chips today.
What is Voicera’s mission and what excites you about that?AI—I actually like to call it machine learning more than AI—can actually have a significant impact on people’s productivity. That’s something that I’m very passionate about, and the voice remains the only modality in the workspace that is hardly digitized. We’ve digitized everything, documents, e-mail, but voice remains.Eva joins the meeting as simple as just sending an invite. Eva figures out a way to join, the least common denominator is telephony. During the meeting you can actually interact with Eva using voice commands, instruct [Eva] to take notes, or remind you to schedule a meeting and we’re working on expanding these tasks. After the meeting you can actually search for when John said the word ‘enterprise’ for example. If you didn’t attend the meeting, you will get some meeting insights that tell you what the meeting was all about, what are the salient terms that were uttered in the meeting, sort of an x-ray [that’s of] a skeleton of the meeting, or for the terms and the categories of topics that were discussed. The holy grail of what we are trying to build is a meeting summary, five or six sentences, and that you can look at and can get a sense of what the meeting was all about. At the end of the meeting you get very high accuracy transcriptions of the highlights of the meeting, whether they are action items or decisions or something like that. Sign up for Voicera FREE here.View the full transcript or listen to the full podcast here.