On April 27th, more than 200 defense industry stakeholders and national security leaders representing 15 different countries gathered at HyperWerx for SparkCognition Government Systems’ “Time Machine Interactive: AI and the Future of Defense” event. Over the course of the day, attendees participated in panels and watched demos on topics ranging from computer vision to autonomy to battlefield operations and cybersecurity and more. To make sure everyone shared an understanding of the foundational technologies driving AI’s hockey stick-like growth trajectory today, Dr. Bruce Porter, SparkCognition’s Chief Science Officer and a two-time chair of the University of Texas Computer Science Department, was on hand to provide a level-set on the breadth and depth of AI capabilities.
With a career in artificial intelligence and computer science that spans over four decades, Dr. Porter offered a professorial perspective on where we are now on this technological journey. He circled back several times to an interesting question: Why now?
What explains the tremendous uptick in success AI seems to be having all at once? Dr. Porter reminded the audience that the field of AI has been around since the 1950s. He noted that Moore’s Law and the increasing availability of training data are significant factors driving the rate of progress. But, in his view, these factors “raise all the boats” and don’t fully explain why so many AI applications are seemingly breaking through all at once.
“Think of another technology that’s had this kind of impact on the world”
So why now? Dr. Porter’s entry point into answering this question began with a brief history of AI itself. He explained how scientists worked methodically to understand the inventory of human cognition and organize fields of AI to approximate human abilities:
- Knowing (knowledge representation and memory)
- Reasoning (inference)
- Using Language (natural language processing)
- Learning (machine learning)
- Seeing (computer vision)
- Navigating (robotics)
Then he talked about how breakthroughs in each of the above AI fields have changed markets and how we move about our daily lives, citing milestones like AI beating the world’s best human players in Go and chess, how e-commerce disrupted retail forever, how Google did the same for information retrieval, how computer vision is changing security and safety surveillance, and the list goes on. Offering a prediction on the next AI-enabled disruption, he named drug discoveries and fully-autonomous driving on the near horizon and acknowledged that Generative AI and large language models are demonstrating sweeping potential across industries.
Dr. Porter confessed that even he struggles to project exactly how far its impact will be felt.
And then, he challenged the audience to “think of another technology that’s had this kind of impact on the world.” Answering the silence, he said: “I can’t think of any.”
He continued, “Maybe technology is successful if it disrupts one industry. But look at where we are today. Why is AI having this outsized impact? And why is the cadence of success picking up so significantly—not only in frequency but amplitude and impact, and the size of the impact?”
Dr. Porter listed the divergent problems addressed by AI, for example, computer vision, predictive maintenance, and cybersecurity, and the AI technologies that solve them, like machine learning, neural networks, deep learning, etc. But how can the same technologies solve all of these different problems? Dr. Porter offered an answer:
What’s really having an impact of late is machine learning. The idea with machine learning is simply this: The AI community—in fact, no humans—know how to write the computer program to automate most of human cognition. It’s just unknown and undoable and too hard.
We spent decades trying to write a parser for natural languages […], and we couldn’t figure out the parser! Decades went into that, and hundreds of people’s lives went into that problem.
The idea with machine learning is don’t try to write those programs. We don’t know how to. Let’s, instead, write a machine learning program and give it lots of examples of correct behavior. And let the machine learning program induce or infer what that algorithm must be. And that’s turned out to be very successful.
There’s one technology in particular for machine learning called neural networks that I’ll talk a little bit more about, and you probably already know. And there’s a subset of that called deep learning, which sounds really fancy and whatnot, but all that really means is a neural net that has more than two hidden layers in it. So deep learning is just a category of neural nets, which is a category of machine learning. But this is the thing that’s really hit home runs of late.
“Breadth and depth” of AI: General purpose deep learning algorithm plus scaffolding of capabilities
Dr. Porter talked about two forces convening to enable AI’s rapid pace of evolution today, which he posits as:
- More mature AI scaffolding throughout the field, creating depth of capabilities
- The versatility of general-purpose deep learning algorithms creating a breadth of coverage
He elaborated:
I think what we’ve seen is that the problems that AI can solve well have been growing […] significantly, and especially in recent years, because of two forces.
By one force, we’re finding ways (or machine learning and deep learning) has exhibited a capability of building a solution to one problem as a kind of foundational solution and then incrementally building or bootstrapping or scaffolding more complex solutions on top of that […] solving harder and harder problems given an initial germ of a simple problem.
It’s what we all do is as, as humans, right? We learn basic knowledge. And first grade that builds on second grade, and so on. And pretty soon, you have a domain expert.
That’s working in machine learning—and it hasn’t worked before. This is a new development in the field. We figured out how to do this scaffolding of capabilities.
The second force is causing this breadth. You know […] I talked about how to look at those […] problems that are solved by AI like cyber security, computer vision, and machine translation. These seem to have nothing to do with one another. Why would one technology solve them all?
It’s this breadth of coverage. That’s what we’re experiencing. What we found (or maybe what we have stumbled into) is this algorithm of deep learning that surprisingly has this breadth of coverage. It’s the same algorithm that solves multiple problems in multiple verticals. Problems that you think have nothing to do with one another.
The ever-expanding realm of problems that AI solves well
Dr. Porter illustrated his point about scaffolding capabilities with a computer vision example. He talked about how training a computer vision model starts with basic challenges (e.g., detecting the presence of a single thing in a single image) that progress toward deeper and more complex capabilities (e.g., understanding situational clues across a series of images). Like kindergarteners practicing and mastering new skills that lead to higher-level lesson work, the models add layers of understanding and fluidity by learning from examples.
- Single object detection and identification
- In-context object detection and identification
- Situation identification
- Activity recognition
- Image search and querying
- Super-resolution and AR
To contextualize the breadth of problems that AI can solve well, he then walked the audience through several examples of one general-purpose deep learning algorithm that be leveraged for many different applications.
- Task: Given an image of a face, determine the name (deep learning for face recognition)
- Task: Given a maintenance log, extract or infer the essential details of the resolution (deep learning for maintenance support)
- Task: Given sensor data, detect when the equipment deviates from normal behavior (deep learning for predictive maintenance)
- Task: Given a file or packet, determine whether it contains malware (deep learning for cybersecurity)
Dr. Porter marveled at how far AI has come with deep learning in recent years: “You’ve got to admit that it’s a striking realization that this swiss-army knife of algorithms here is solving end-to-end problems—generating complete solutions. The same algorithm is [serving] all these domains with just a little bit of tweaking by [today’s] data scientists. That is amazing, and it’s one of the reasons why machine learning is working now to such a startling degree.”
How Generative AI is “rewriting the storyline”
Wrapping up his short presentation, Dr. Porter spoke about how generative AI is “rewriting the storyline.” In his words, “There’s a hockey stick move here that has disrupted everything.” Noting how there are two general classes of AI: discriminative, which encompasses almost all of the familiar AI fieldwork to date and aims to ‘find’ things, and generative AI, which accelerates and automates users to ‘produce’ new things. “This class is going to take off in importance—no doubt about it.” Dr. Porter broke down the current generative AI platform projects that SparkCognition is working on, from groundbreaking seismic imaging techniques to enabling highly intuitive chat-based dialogs inside our solutions to automating programming for accelerating AI discoveries.
Dr. Porter’s level set helped the TMI23 attendees put the rapid expansion of AI capabilities into perspective, showing how this phenomenal ascent is the product of two key factors creating synergistic ‘breadth and depth’ effects today. Over years of steady research and learning, we’re witnessing akin to ‘mastery’ come into play in the ability to stack complex solutions on top of foundational ones. This is being reinforced by powerful and malleable deep learning algorithms that have displayed an astonishing range to address rather unrelated issues across various industries. And as the field continues to evolve, the growing importance of generative AI will further accelerate AI’s pace of progress, opening up exciting possibilities for the future of artificial intelligence.
