What We Overlook To Rush AI Deployment
Contributed by Adnan Khaleel.

I have often written about how AI, in spite of being a promising technology, is still in its infancy from both a technological and even from a societal readiness perspective. Anyone who needs convincing of this fact needn’t look any further than this recent article from the Washington post which shows how a medical wellness algorithm, even with race discounted, was favoring white patients over sicker, more needy African-Americans (see link).
Although the immediate implications of this flaw are not clearly indicated in the article, it doesn’t take much for us to reach a conclusion that some people’s lives might have been adversely affected by a serious flaw that decided which individuals likely needed healthcare. Worse still, experts agree that this flaw is not just isolated to this single vendor, and likely exists in systems in wide usage by other vendors.
Well-intentioned as the developers may have been, the algorithm intentionally discarded race as a factor to precisely avoid this sort of an outcome, and consequently, backlash. Despite this careful planning, it’s interesting to see that the outputs weren’t race-neutral as common-sense would predicate. I recommend you read the full article for the details, but the simplistic underlying cause of the flaw is (our old friend) “causation”.
Let me elaborate: just removing the primary parameter, race in this case, is insufficient if you still include other parameters that are co-dependent, cost in this case. And clearly cost is not a neutral parameter, especially when you are mixing population segments that clearly possess very different purchasing powers.
And herein lies the crux of the argument that I’ve often made with widescale deployment of some of the latest AI technologies without fully understanding how they could affect the lives of millions of people with unintended consequences.
In our rush to optimize (or improve) life expectancy (or corporate profits) with AI, might we be actually doing the opposite?
Knowingly or not, careful analysis of this flaw is a standard example of prejudice bias. This is the case when the historical training data contains artifacts that haven’t been normalized for the population to ensure fairness. There are two other types of data bias called measurement and sample, and you can find a more detailed article about the various types of data bias here.
Almost all sources of historically collected social data likely contains some bias. In some instances, say criminal justice data, we are more aware of the racial bias and can account for it. But in other instances, as the case above suggests, racial bias can appear in subtle unobvious ways. This brings to mind a popular joke amongst statisticians that, if you torture data enough, it will tell you anything you want it to — but I digress.
Hidden data biases are just the tip of the iceberg when it comes to peeling back the onion, and really understanding why, or rather how, an AI neural net arrived at its conclusions. The models today are very accurate indeed, but also so horribly complicated that human minds are simply incapable of deciphering how “exactly” a bunch of nodes with weighted edges can disambiguate a cat from a dog.
Let’s set aside for a moment that we don’t exactly understand this in our own brains either, but at least we didn’t create them from scratch like we have artificial neural networks. They literally are the proverbial “black box” that performs a function and yet provides little transparency into how it does so. Neural Network transparency is an active area of research and you can find some details in this link.
I think by now, you might have begun to put the pieces together, and how between not completely understanding the data with its various biases, coupled with incomplete comprehension of how neural nets arrive at their decisions we are metaphorically handing over our lives to a system that does not lend itself to accountability.
The good news is that it doesn’t have to be this way.
Before you jump to a conclusion and think that government regulation is the answer, my thoughts are that forcing companies to greater accountability might not be a bad idea. However, we also want to be careful of a heavy-handed approach that might stifle technological innovation. But, we also know that relying entirely on corporations to devise standards is also the reason we’re in this quagmire of social media fake news. Although a single commercial entity is unlikely to devise a fair standard, a group of corporations and social thought-leaders can jointly devise a set of standards that benefits everyone, at least until such time that a rogue corporation forces government regulation.
The standards would encompass industry best practices for avoiding and/or detecting these social biases, as well as create an industry standard for AI model exploitability at the very least. And I can see such a foundation having a deep, long lasting and profound effect on the industry in much the same way the Free Software Foundation (fsf.org) has had on open source software.
Which brings me to my last point. This standards body is just one of the objectives that we’re considering as part of the Austin AI Ecosystem. As the name suggests, the mandate is broader than that, but with strong community support and industry backing, we’re hoping to create a model forum that could form the basis for such an organization. We are at a very early stage of creating the AI-Ecosystem, and we are soliciting suggestions from the community and industry at large through public forums.
For more details, please get in touch with me.
About The Author:
Adnan is an experienced business development professional with over 20 years of experience in High Performance Computing and Machine Learning. Currently, Adnan is a Global HPC Sales Strategist at Dell EMC, for HPC & AI, and with his strong technical background, Adnan works very closely with customers in helping them choose the optimal HPC solutions for their specific problems.

