Bing Liu was road testing a self-driving car, when suddenly something went wrong. The vehicle had been operating smoothly until it reached a T-junction and refused to move. Liu and the car’s other occupants were baffled. The road they were on was deserted, with no pedestrians or other cars in sight. “We looked around, we noticed nothing in the front, or in the back. I mean, there was nothing,” says Liu, a computer engineer at the University of Illinois Chicago.
Stumped, the engineers took over control of the vehicle and drove back to the laboratory to review the trip. They worked out that the car had been stopped by a pebble in the road. It wasn’t something a person would even notice, but when it showed up on the car’s sensors it registered as an unknown object — something the artificial intelligence (AI) system driving the car had not encountered before.
The problem wasn’t with the AI algorithm as such — it performed as intended, stopping short of the unknown object to be on the safe side. The issue was that once the AI had finished its training, using simulations to develop a model that told it the differences between a clear road and an obstacle, it could learn nothing more. When it encountered something that had not been part of its training data, such as the pebble or even a dark spot on the road, the AI did not know how to react. People can build on what they’ve learnt and adapt as their environment changes; most AI systems are locked into what they already know.
In the real world, of course, unexpected situations inevitably arise. Therefore, Liu argues that any system aiming to perform learnt tasks outside a lab needs to be capable of on-the-job learning — supplementing the model it’s already developed with new data that it encounters. The car could, for instance, detect another car driving through a dark patch on the road with no problem, and decide to imitate it, learning in the process that a wet bit of road was not a problem. In the case of the pebble, it could use a voice interface to ask the car’s occupant what to do. If the rider said it was safe to continue, it could drive on, and it could then call on that answer for its next pebble encounter. “If the system can continually learn, this problem is easily solved,” Liu says.
Such continual learning, also known as lifelong learning, is the next step in the evolution of AI. Much AI relies on neural networks, which take data and pass them through a series of computational units, known as artificial neurons, which perform small mathematical functions on the data. Eventually the network develops a statistical model of the data that it can then match to new inputs. Researchers, who have based these neural networks on the operation of the human brain, are looking to humans again for inspiration on how to make AI systems that can keep learning as they encounter new information. Some groups are trying to make computer neurons more complex so they’re more like neurons in living organisms. Others are imitating the growth of new neurons in humans so machines can react to fresh experiences. And some are simulating dream states to overcome a problem of forgetfulness. Lifelong learning is necessary not only for self-driving cars, but for any intelligent system that has to deal with surprises, such as chatbots, which are expected to answer questions about a product or service, and robots that can roam freely and interact with humans. “Pretty much any instance where you deploy AI in the future, you would see the need for lifelong learning,” says Dhireesha Kudithipudi, a computer scientist who directs the MATRIX AI Consortium for Human Well-Being at the University of Texas at San Antonio.
Continual learning will be necessary if AI is to truly live up to its name. “AI, to date, is really not intelligent,” says Hava Siegelmann, a computer scientist at the University of Massachusetts Amherst who created the Lifelong Learning Machines research-funding initiative for the US Defense Advanced Research Projects Agency. “If it’s a neural network, you train it in advance, you give it a data set and that’s all. It does not have the ability to improve with time.”
In the past decade, computers have become adept at tasks such as classifying cats or tumours in images, identifying sentiment in written language, and winning at chess. Researchers might, for instance, feed the computer photos that have been labelled by humans as containing cats. The computer receives the photos, which it interprets as numerical descriptions of pixels with various colour and brightness values, and runs them through layers of artificial neurons. Each neuron has a randomly chosen weight, a value by which it multiplies the value of the input data. The computer runs the input data through the layers of neurons and checks the output data against validation data to see how accurate the results are. It then repeats the process, altering the weights in each iteration until the output reaches a high accuracy. The process produces a statistical model of the values and the placement of pixels that define a cat. The network can then analyse a new photo and decide whether it matches the model — that is, whether there’s a cat in the picture. But that cat model, once developed, is pretty much set in stone.
One way to get the computer to learn to identify many objects would be to develop lots of models. You could train one neural network to recognize cats and another to recognize dogs. That would require two data sets, one for each animal, and would double the time and computing power needed to develop each model. But suppose you wanted the computer to distinguish between pictures of cats and dogs. You would have to train a third network, either using all the original data or comparing the two existing models. Add other animals into the mix and yet more models must be developed.
Training and storing more models requires greater resources, and this can quickly become a problem. Training a neural network can take reams of data and weeks of time. For instance, an AI system called GPT-3, which learnt to produce text that sounds as if it was written by a human, required almost 15 days of training on 10,000 high-end computer processors1. The ImageNet data set, which is often used to train neural networks in object recognition, contains more than 14 million images. Depending on the subset of the total number of images that is used, it can take from a few minutes to more than a day and a half to download. Any machine that has to spend days re-learning a task each time it encounters new information will essentially grind to a halt.
One system that could make the generation of multiple models more efficient is Self-Net2, created by Rolando Estrada, a computer scientist at Georgia State University in Atlanta, and his students Jaya Mandivarapu and Blake Camp. Self-Net compresses the models, to prevent a system with a lot of different animal models from growing too unwieldy.
The system uses an autoencoder, a separate neural network that learns which parameters — such as clusters of pixels in the case of image-recognition tasks — the original neural network focused on when building its model. One layer of neurons in the middle of the autoencoder forces the machine to pick a tiny subset of the most important weights of the model. There might be 10,000 numerical values going into the model and another 10,000 coming out, but in the middle layer the autoencoder reduces that to just 10 numbers. So the system has to find the ten weights that will allow it to get the most accurate output, Estrada says.
The process is similar to compressing a large TIFF image file down to a smaller JPEG, he says; there’s a small loss of fidelity, but what is left is good enough. The system tosses out most of the original input data, and then saves the ten best weights. It can then use those to perform the same cat-identification task with almost the same accuracy, without having to store enormous amounts of data.
To streamline the creation of models, computer scientists often use pre-training. Models that are trained to perform similar tasks have to learn similar parameters, at least in the early stages. Any neural network learning to recognize objects in images, for instance, first needs to learn to identify diagonal and vertical lines. There’s no need to start from scratch each time, so newer models can be pre-trained with the weights that already recognize those basic features. To make models that can recognize cows or pigs or kangaroos, Estrada can pre-train other neural networks with the parameters from his autoencoder. Because all animals share some of the same facial features, even if the details of size or shape are different, such pre-training allows new models to be generated more efficiently.
The system is not a perfect way to get networks to learn on the job, Estrada says. A human still has to tell the machine when to switch tasks; for example, when to start looking for horses instead of cows. That requires a human to stay in the loop, and it might not always be obvious to a person that it’s time for the machine to do something different. But Estrada hopes to find a way to automate task switching so the computer can learn to identify characteristics of the input data and use that to decide which model it should use, so it can keep operating without interruption.
It might seem that the obvious course is not to make multiple models but rather to grow a network. Instead of developing two networks for recognizing cats and horses respectively, for instance, it might appear easier to teach the cat-savvy network to also recognize horses. This approach, however, forces AI designers to confront one of the main issues in lifelong learning, a phenomenon known as catastrophic forgetting. A network trained to recognize cats will develop a set of weights across its artificial neurons that are specific to that task. If it is then asked to start identifying horses, it will start readjusting the weights to make it more accurate for horses. The model will no longer contain the right weights for cats, causing it to essentially forget what a cat looks like. “The memory is in the weights. When you train it with new information, you write on the same weights,” says Siegelmann. “You can have a billion examples of a car driving, and now you teach it 200 examples related to some accident that you don’t want to happen, and it may know these 200 cases and forget the billion.”
One method of overcoming catastrophic forgetting uses replay — that is, taking data from a previously learnt task and interweaving them with new training data. This approach, however, runs head-on into the resource problem. “Replay mechanisms are very memory hungry and computationally hungry, so we do not have models that can solve these problems in a resource-efficient way,” Kudithipudi says. There might also be reasons not to store data, such as concerns about privacy or security, or because they belong to someone unwilling to share them indefinitely.
Siegelmann says replay is roughly analogous to what the human brain does when it dreams. Many neuroscientists think that the brain consolidates memories and learns things by replaying experiences during sleep. Similarly, replay in neural networks can reinforce weights that might otherwise be overwritten. But the brain doesn’t actually review a moment-by-moment rerun of its experiences, Siegelmann says. Rather, it reduces those experiences to a handful of characteristic features and patterns — a process known as abstraction — and replays just those parts. Her brain-inspired replay tries to do something similar; instead of reviewing mountains of stored data, it selects certain facets of what it has learnt to replay. Each layer in a neural network, Siegelmann says, moves the learning to a higher level of abstraction, from the specific input data in the bottom layer to mathematical relationships in the data at higher layers. In this way, the system sorts specific examples of objects into classes. She lets the network select the most important of the abstractions in the top couple of layers and replay those. This technique keeps the learnt weights reasonably stable — although not perfectly so — without having to store any previously used data at all.
Because such brain-inspired replay focuses on the most salient points that the network has learnt, the network can find associations between new and old data more easily. The method also helps the network to distinguish between pieces of data that it might not have separated easily before — finding the differences between a pair of identical twins, for example. If you’re down to only a handful of parameters in each set, instead of millions, it’s easier to spot the similarities. “Now, when we replay one with the other, we start looking at the differences,” Siegelmann says. “It forces you to find the separation, the contrast, the associations.”
Focusing on high-level abstractions rather than specifics is useful for continual learning because it allows the computer to make comparisons and draw analogies between different scenarios. For example, if your self-driving car has to work out how to handle driving on ice in Massachusetts, Siegelmann says, it might use data that it has about driving on ice in Michigan. Those examples won’t exactly match the new conditions, because they’re from different roads. But the car also has knowledge about driving on snow in Massachusetts, where it is familiar with the roads. So if the car can identify only the most important differences and similarities between snow and ice, Massachusetts and Michigan, instead of getting bogged down in minor details, it might come up with a solution to the specific, new situation of driving on ice in Massachusetts.
Looking at how the brain handles these issues can inspire ideas, even if they don’t replicate what’s going on biologically. To deal with the need for a neural network that can learn tasks without overwriting the old, scientists take a cue from neurogenesis — the process by which neurons are formed in the brain. A machine can’t grow parts the way a body can, but computer scientists can replicate new neurons in software by generating connections in parts of the system. Although the mature neurons have learnt to react to only certain data inputs, these ‘baby neurons’ can respond to all the input. “They can react to new samples that are fed into the model,” Kudithipudi says. In other words, they can learn from new information while the already-trained neurons retain what they’ve learnt.
Adding more neurons is just one way to enable a system to learn new things. Estrada has come up with another approach, on the basis of the fact that a neural network is only a loose approximation of a human brain. “We call the nodes in a neural network ‘neurons’. But if you see what they’re actually doing, they’re basically computing a weighted sum. It’s an incredibly simplified view of real, biological neurons, which perform all sorts of complex nonlinear signal processing.”
In an effort to mimic some of the complicated behaviours of real neurons more successfully, Estrada and his students developed what he calls deep artificial neurons (DANs)3. A DAN is a small neural network that is treated as a single neuron in a larger neural network.
DANs can be trained for one particular task — for instance, Estrada might develop one for identifying handwritten numbers. The model in the DAN is then fixed, so it can’t be changed and will always provide the same output to other neurons in the still-trainable network layers surrounding it. That larger network can go on to learn a related task, such as identifying numbers written by someone else — but the original model is not forgotten. “You end up with this general-purpose module that you can reuse for similar tasks in the future,” Estrada says. “These modules allow the system to learn to perform the new tasks in a similar way to the old tasks, so that the features are more compatible with each other over time. So that means that the features are more stable and it forgets less.”
So far, Estrada and his colleagues have shown that this technique works on fairly simple tasks, such as number recognition. But they’re trying to adapt it to more challenging problems, including learning how to play old video games such as Space Invaders. “And then, if that’s successful, we could use it for more sophisticated things,” says Estrada. It might, for instance, prove useful in autonomous drones, which are sent out with basic programming but have to adapt to new data in the environment, and will have to do any on-the-fly learning within tight power and processing constraints.
There’s a long way to go before AI can function as people do, dealing with an endless variety of ever-changing scenarios. But if computer scientists can develop the techniques to allow machines to make the continual adaptations that living creatures are capable of, it could go a long way towards making AI systems more versatile, more accurate and more recognizably intelligent.
This article is part of Nature Outlook: Robotics and artificial intelligence, an editorially independent supplement produced with the financial support of third parties. About this content.