There’s a lot of hype surrounding AI. Unfortunately, a lot of it is hyperbolic warnings about how we’ll lose our humanity and the machines will be smarter than we are. Just about every field and institution is preparing for an AI revolution. There’s even a brand new Minister of AI in the UAE to help deal with the coming changes. Do you really know what’s changing though? At ODSC East 2019, Dr. Pieter Abbeel, professor at UC Berkeley and co-founder of Covariant.ai, helped the audience understand what’s happening in the field of AI and how you might need to respond. Let’s take a look at these trends in AI.
What’s Driving AI Breakthroughs?
We’re getting these breakthroughs because the world is messy. Our internal AI systems are coming up against real models of the world and having to adapt to those variables that we can’t ever predict. No system of data is perfect unless it’s built in the confines of a lab, but how accurate is data like that?
These breakthroughs are possible because we’ve begun creating systems that can process patterns the way humans do but on a more massive scale. One of humanity’s great adaptations was the ability to extrapolate patterns and apply those patterns to different situations. Now, our neural networks mimic the way we learn.
Major Advances in AI
We’re on the precipice of a huge leap in our AI systems and here are breakthroughs leading the way:
Traditional programming happens line by line and is time intensive. The amount of coding just to translate an image to the virtual mind is breathtaking. Instead of that, we’re building networks that operate more closely to the human brain. It cuts our training time down and gives machines the power to harness pattern recognition and learn without us.
Now, the program lives inside the neural network. If you want to change the program, you change neural connections, much the way our neural connections fire when we’re learning new information. Changing lines of code to adapt requires too much human intervention to be practical for the amount and types of data we have now, so our neural networks can take the reigns and adapt to new stimuli.
Neural networks made deep learning possible. This type of data processing halved the ImageNet competition time in half recently and ultimately ended the competition entirely. Access to deep learning models makes harnessing unstructured data that much more realistic.
The biggest deal happening now is that programs aren’t just training, they’re actually learning. These neural networks can process massive possibilities simultaneously and reduce the time needed to learn skills. It’s now more possible to run hands-off simulations with big data without intensive input from the human side.
For example, a neural network tries to learn to run. In the first few tries, it falls immediately. However as tries new things and fall over just a fraction of a second later, it begins to combine what works until the model is running in virtual space.
In the real world, this is responsible for the success of AlphaGo Zero. In another recent experiment, a machine learned in the virtual world how to manipulate a block with human-like fingers, and was able to translate that to a physical representation.
The big breakthrough here is that machines are learning unsupervised. They’re taking this data and forming patterns around it without the aid of human intervention. This could make it possible to deploy robots in business situations that can be given a different task every day and successfully learn how to do it.
So What Happens Now?
AI will be present everywhere. If your business can harness it to make better predictions, you’ll have the leading edge. The next wave of integration will be those deep learning insights that can consider unstructured data with minimal human intervention.
Now that robots are getting pretty good at learning and handling our menial tasks, robots that can actually see are on the horizon. Once robots have eyes, making them teachable is a bigger possibility, which could revolutionize manufacturing.
Companies such as Covariant AI are building the new generation of AI automation, expanding what’s possible. Their target is flexible solutions that fit into a specialized context, rather than trying to fit a generalized solution into every setting.
So Is AI Here To Stay?
Short answer, yes. We have more data, better computation, and more innovative execution. It’s beginning to unlock solutions to persistent problems. Those citing Moore’s Law were right up to a certain point. Even there, advances in how we design the chips – i.e., building dedicated neural chips – have reopened possibilities for our computational power.
And what about humans? The good news is that even with advances in AI, humans aren’t obsolete. We’re still the reigning champions for things like creativity. Even more, we’re beginning to build value aligned neural networks anyway.
In the future, we’re looking towards some of the worst of AI (wealth accumulation, fake news, and other unintentional bad incomes) and building more robust systems for discouraging those outcomes. The flip side is that our computation power is nearing human levels. We’ll be able to make bigger advances in medical diagnostics, work on big projects like real Mars travel, and maybe one day even build extensions of our very own brains. It’s a wide world out there, and we’re only just beginning to explore.