We received compelling questions from you following the Alumni Ventures Group webinar, How AI and Machine Learning Will Impact You. Panelist Daeil Kim, co-founder and CEO of AI.Reverie, tackles some of them here.
What is your method for conjuring up synthetic ‘edge case’ training data, since the most useful edge cases are most often those that are the hardest to anticipate?
Oftentimes the edge cases that we deal with are well known to our clients beforehand, and as a result they come to us with a general sense of where their algorithm goes awry when testing it out in the field. In certain cases, there might be situations they have in mind that they don’t have examples for, but those instances are usually general in their scope: they focus on categories of objects they don’t have support for.
Are we going to see models that are more and more uninterpretable? The so-called black boxes that perpetuate bias. When these models are deployed for determining eligibility for health insurance, for example, or criminal justice or medical diagnosis – isn’t that a bigger more present threat to society than time-traveling terminators?
Deep networks are in general quite uninterpretable. If you’re looking for better interpretation, you might want to consider graphical models or Bayesian approaches that are often used in medical systems where interpretability is key. Deep learning is often considered a representational system and just allows for a highly nonlinear mapping between input (image) and output (“it’s a car!”) with a focus on getting that mapping right without consideration for whether the process is interpretable. I would certainly agree that biases are currently more of a present threat than a time-traveling terminator, and the real present threat is the automation of labor without the social safety net to support folks who no longer work.
What can we achieve with natural language processing today, and what will be possible in the next five to 10 years?
I believe the big advancements will come from things like unsupervised learning and reasoning systems that allow for symbolic logical rules to be incorporated into these networks. The ability to incorporate notions of that within the architecture of the network to allow for such reasoning could be a very powerful evolution of NLP systems today.
What is the biggest hurdle to mass adoption of AI and machine learning technology?
I really do think it’s the data costs and finding the skilled labor to train algorithms effectively. The data is expensive not only in its curation and annotation, but also in terms of the labor used to correct errors (often by machine learning experts) which isn’t the best use of their time. In the future though, I do think that these systems will become more and more commoditized and there will be ways of leveraging AI more easily once those tools are within the open source ecosystem. Consider how powerful it was to have tools like TensorFlow and PyTorch in accelerating machine learning these past few years.
There are so many companies in the AI and machine learning space. How do you see the market evolving over the next several years, particularly in terms of mergers/acquisitions and collaboration?
I certainly think there are going to be a lot of acquisitions within the AI space as tools become more consolidated and integrated into one another. It’s challenging to make a business in AI and that’s partially because it’s so new and companies are still trying to figure out how to integrate those processes into their core business. Those that don’t survive this product-fit market process will inevitably be acquired given the scarcity of the labor force, but the ones that make it will become enormous businesses.
For a business-oriented person, what are some of the best sources available to stay informed about these sectors?
It’s a large space, but one of the best general starting points from the business side is to look at Nathan Beniach’s state of AI: https://www.stateof.ai. Hope that helps!
Please share any additional questions at firstname.lastname@example.org