explorations, AI, artificial intelligence, emerging tech
event recap

Event Recap: What Can an Algorithm Actually Know?

A recap of our Demystifying Algorithms event.

· 3 min read

This is a recap of the virtual event component of our Demystifying Algorithms series. Check out the first piece here, second piece here, and the third one here.

We closed out Explorations by asking two AI experts an ambitious question: Can an algorithm actually know anything?

In last week’s live event, Ryan and Hayden spoke with Mark Riedl, professor of computer science at Georgia Tech and Saška Mojsilović, Head of Trusted AI Foundations at IBM Research, to get the answer.

See below for the full event recording, or keep reading if you prefer a written recap.

It’s all about the transfer

Emerging Tech Brew

Riedl told us that an AI system’s ability to generalize—to apply its specific training to new, unfamiliar areas—is a critical measure of whether it can be said to know anything.

An example: Riedl used gameplay to illustrate the different levels of sophistication AI can have.

  • “Can you play the game? Can you win the game? Well, can you win the game in a more interesting way? Can you play a slightly different game than the one you were trained on?”
  • “This is called transfer—taking what you might ‘know,’ and putting it into a new situation.”

His research uses storytelling to test an AI system’s ability to apply concepts in this way. He gave the example of asking a neural network to tell a story about a mundane aspect of human life, like the experience of eating in a restaurant.

  • When an AI system trips over the details of a humdrum request, it indicates that “the models that these neural nets have learned from reading things on the internet are not as complete as we would expect them, or want them to be.”

Big picture: Riedl said AI researchers are currently focused on“broadening the input space, without broadening the output space.” So while a system like OpenAI’s image generation program Dall-E can take in a ton of information, it still can only output one specific thing: an image.

  • “We haven’t broadened the output. We haven’t gone to a system that can draw, and tell stories, and drive a car.”

Get practical

Emerging Tech Brew

So, we’ll ask again: Do AI systems know anything?

Keep up with the innovative tech transforming business

Tech Brew informs business leaders about the latest innovations, automation advances, policy shifts and more to help them make smart decisions.

According to Mojsilović, “realistically, if we look at the systems of today, they are kind of dumb. They’re very, very far away from the cognitive capabilities of human beings. Or, at least, they are far less intelligent than what people think from what they read in the news.”

Like Riedl, she said the focus has been on developing more narrow applications—where an algorithm learns to do a single task, like speech recognition, really well. But, again, that expertise doesn’t transfer: If you take a system that is trained to diagnose breast cancer, Mojsilović noted, that very same system cannot diagnose skin cancer.

But narrowness isn’t necessarily bad. Mojsilović pointed out that singularly focused AI can help solve important problems, like discovering new drugs and creating new molecules.

  • “I think we may want to take a pause from obsessing over artificial general intelligence and maybe think about how we create AI solutions for these kinds of problems.”
  • Mojsilović and her IBM team created architectures that recently produced “the very first antibiotics that were created by AI.”

Big picture: Mojsilović said that “we need moonshots” like AGI because they are “how we advance science and technology,” but that AI researchers also need to keep practical, socially useful applications top of mind.

And despite the power of certain narrow AI systems, she thinks we’re “still at the very beginning of being able to create capabilities that even remotely mimic what humans can do.”

Keep up with the innovative tech transforming business

Tech Brew informs business leaders about the latest innovations, automation advances, policy shifts and more to help them make smart decisions.