The constraints of artificial intelligence
Flawed Reality with Tyrone Grandison on what factors hold AI back
Constraints are limiting factors. They are dimensions or building blocks that create a bounding box around a problem or solution space and force us to focus. Currently, I define AI or Artificial Intelligence as the field in computer science where a computer has the ability to perform, or rather mimic, functions typically associated with human beings. This includes tasks like perceiving, learning and interacting with your environment. For this discussion, AI technology includes everything from autonomous vehicles to the application of probabilistic and statistical methods to the spaces of natural language processing, image processing, recommendation creation, predictions, and virtual agent construction. The constraints of AI become apparent when one tries to apply AI. The restrictions of the domain, algorithms, user workflow, and acceptable solution form factor quickly present themselves and focus our end result.
AI method of practice
Every single one of the fields under the AI umbrella has been around in some form or the other for a while now. However, applying AI techniques to specific use cases that are popular is a very recent phenomenon. The method or practice of how do you actually apply AI and apply it responsibly– that's new. The core implicit constraint that we have to recognize is that at the very core of all of it is math. Everything that we're doing is the application of math– it is not magic, and it is very dependent on how people actually apply their mathematical knowledge. The art of an AI practitioner applying math to a particular problem is what drives value.
The best thing that you can actually do for your organization if you're looking to see if a problem can be solved using AI, is to get someone that has a wealth of knowledge and background in a myriad of different methods and techniques and that has applied them in multiple different scenarios. That experience ensures that the person in question is aware of what needs to be done, when it needs to be done, what needs to be used, or if the problem just cannot even be answered with Artificial Intelligence.
That leads you to the recognition that you may not want a practitioner that knows or thinks that their one method or specialization is going to be applicable to every scenario before them. If you have a hammer, everything should not be a nail. Sometimes, all you may need is a tiny screwdriver.
Constraint 1: AI talent
The first constraint is the practitioner, or the person who's actually implementing AI on your behalf. Whether you want a generalist or a specialist is completely up to you and your knowledge of the problem space that you want to explore. But for most scenarios, you initially need a generalist to come in, look around, and figure out what works and what doesn't work.
Constraint 2: AI perception
The second constraint is the current perception of AI. It's the same perception that software engineering had three to four decades ago – of being magical and people just handing over trust, faith, control to whatever is actually built. At the core, AI is math. It is simply a set of methods that take input and spit out some output variables.
You then have to interpret the output. You have to make sure that whatever method you're actually using is relevant, i.e. that the assumptions of that method actually align well with the assumptions of the problem space. You can't haphazardly just apply any random method to any particular problem, and the 'magical AI box' is going to make it all happen. Sadly, it just doesn't work that way.
Constraint 3: AI resources
If you're talking about robotics and vehicles, the other constraint is going to be resources. It's going to be cost and storage and how do you build everything ethically and efficiently. For software applications, the explosion of AI solutions has been fueled by algorithmic advances, cloud computing, and data explosion. All these factors lowered the total cost of development and deployment of software-based AI solutions. Hardware-based AI solutions are where software-based AI solutions were a decade or two ago.
Constraint 4: AI assumptions and unintended consequences
The next constraint is the conception and validation of the end results – not just interpreting the algorithm’s results and verifying the correctness of domain assumptions, but also the factoring in of unintended consequences. This realm or category includes ethics, privacy, impact, and other thoughtful and realistic considerations of the technology’s use. You should understand how the AI solution itself perpetrates the negatives and positives of the current environment and facilitates the future.
Constraint 5: Target workflow
Another important constraint is the identification of where you are actually going to deploy AI. It's imperative to realize how your AI will integrate into your system, technical and non-technical, and to understand the best way to present it to the end user. You must outline the least intrusive and most ethical way for AI to be most effective. And to do all of that, you need to truly understand the business and how it works.
Artificial intelligence is an art form, because it includes assumptions in math and is dependent on the way a practitioner chooses to wield their math toolkit. This art form is on its path to maturity now. We're trying to construct a future without the benefit of a well-defined past. We're learning as we're going at this point. There have been a few success cases over the years. The recommendation engines from LinkedIn and NetFlix are big ones. Google’s Image Search is another significant success.
But when you look underneath the covers and you actually examine those examples, they're really very focused solutions to very focused problems, that can be implemented really easily by inspecting and interrogating their mathematical form. You have models that actually gel with domain assumptions, that give you really defined outputs. Those are the scenarios that if you scale them well are going to get all the praise and the hoopla.
However, the magic that everyone assumes is there, that everyone assumes is happening is, at this point, not really happening. It just looks and feels like magic.