On the future of AI with Pandata's Cal Al-Dhubaib
Cal Al-Dhubaib is the Founder and CEO of Cleveland-based Pandata where they help innovative organizations design and develop trustworthy, human-centered AI.
1. Name: Cal Al-Dhubaib
2. Company and Title: Pandata, CEO & AI Strategist
3. Hometown: Dhahran, Saudi Arabia
4. Current City: Cleveland
5. In two sentences or less, tell us about Pandata.
We help companies design and develop practical AI applications. We're focused on making AI useful, trustworthy, and explainable.
6. When it comes to AI, why is there *still* a disparity between the hype and results?
A statistic that has remained steady for the past few years is the failure rate of AI projects. You'll hear some variation of fewer than 80% of projects making it to production, and of those, fewer than a third have a demonstrable ROI. These numbers change year to year, but the gist is a disparity between hype and result. Broadly speaking, AI is software that can learn to recognize and react to complex patterns in a wide variety of data. The three bolded words are crucial:
Recognizing complex patterns: If we accept that data fuels the patterns that AI systems recognize and react to, and that most data is rooted in human behavior, we must acknowledge that AI requires fine-tuning over time.
Reacting to complex patterns: Whether you're building a truly automated process like a recommender system (think Amazon and Netflix - you may like X item) or alerting a human to take action (for example, detecting potential fraud or identify a potential product defect in an assembly line), there's generally a human at the receiving end of the solution. The intentional design of this interaction must be taken into consideration to ensure buy-in and result in the behavior or action that in-turn yields ROI.
In my experience, most AI projects must go through a 'building resilience' phase. This may look like addressing potentially negative unintended consequences, a shift in human behavior, fine-tuning an assumption, or navigating change management related to the adoption of the solution. Many AI projects are killed too soon as a result. Alternatively, some are not killed soon enough because these inevitable challenges were not taken into consideration at the onset.
7. What are some barriers to AI development and how have they changed over the last few years?
Building AI solutions used to be the exclusive domain of 'deep tech' companies of the scale of Facebook, Google, IBM, and Microsoft, among others. In the past few years, the building blocks to produce AI, such as pre-trained models and pipelines to work with multimedia data such as text and images, are becoming more accessible. In the past, you would need teams of data scientists and engineers to build these blocks and then integrate them into end-to-end solutions.
Let's say you're a healthcare system and you're trying to get a better handle on preventing readmissions, especially with respect to underserved populations. You may need to connect disparate datasets on these patients that can very rarely leave organizational walls - patient demographics, clinical notes, and patient medical information. You may need to take into consideration labs (documents), medical images, physician notes, in addition to more traditionally structured data. And then you would need to build a model that takes these data into consideration to predict the occurrence of a readmission. The field of AI Development refers to the practice of developing new AI-powered capabilities either as software or back-end business processes.
Some organizations fret about having 'good enough' data. Data quality and governance are important, but to date, I've yet to encounter a 'pristine' dataset. The discipline of AI Development has evolved in the past few years from the world of research to be more of an applied discipline using the building blocks described above. You still need to do fine-tuning of assumptions, but the effort to prepare data for building powerful models has dramatically decreased.
8. There has been a lot of talk about trustworthy AI. What is it and why does designing AI with it in mind matter?
You may have heard of Apple's credit card product discriminating against women, or hospitals leveraging a readmissions algorithm that prioritized white patients over black patients, or AI-driven chatbots coming up with 'toxic' content. In many cases, these unintended consequences are a function of data complexity and goal complexity. This is an extension of AI as software that recognizes and reacts to complex patterns.
Complexity in taking action: When we think about 'taking action', it's relatively straight forward in some applications to determine 'good' and 'bad' performance. If you're predicting a customer is at risk for churning - after a certain period of time, they either do or don't. Either way, you have a record of truth to compare a prediction against. When it comes to more complex actions like recommending a product or creating a narrative, the definition of 'correct' gets a little more complicated. In many cases, two equally trained professionals may not agree on the same definition of manufacturing defect or the correct phrasing for a story being relayed in an article. This makes it harder to audit AI systems.
Complexity in patterns: The recent decrease in barriers is leading to more AI Development on broader types of data - including images, video, and text - and at greater scales than in the past. GPT-3, for example has 175 billion parameters. That's essentially incomprehensible. Even if you know which parameter is responsible for bias, it's impossible to describe in a human-intelligible way. This makes it harder to inspect data for biases, toxicity, or blatant errors, especially when you consider the scale of thousands, millions, and even billions of observations.
Trustworthy AI development takes these factors into consideration with the goal of developing AI that is more fair, more secure, and easier to explain and audit.
9. The future of AI... What does that mean to you?
An interesting thing about AI is its definition constantly shifts, at least in the perception of the public. Things we may have once considered "AI" like autocorrect, today we take for granted as basic software functionalityThe blocks to produce AI, such as pre-trained models and pipelines to work with multimedia data are becoming easier to use. This means more time spent designing and building applications and less time on enabling infrastructure.
As more AI solutions become used in production, the requirements to maintain these solutions are only going to grow. And unlike traditional software that is static - AI is powered by models and patterns of human behavior that shift over time. The future of AI means more intelligent applications working across both traditional and multimedia data sources, but it also means we're going to require even more trained professionals dedicated to the care and maintenance of these systems.
Cal Al-Dhubaib is the Founder and CEO of Cleveland-based Pandata where they help innovative organizations design and develop trustworthy, human-centered AI.