Skip to Content
Artificial intelligence

Neural Networks Are Learning What to Remember and What to Forget

Memory is a precious resource, so humans have evolved to remember important skills and forget irrelevant ones. Now machines are being designed to do the same.

Deep learning is changing the way we use and think about machines. Current incarnations are better than humans at all kinds of tasks, from chess and Go to face recognition and object recognition.

But many aspects of machine learning lag vastly behind human performance. In particular, humans have the extraordinary ability to constantly update their memories with the most important knowledge while overwriting information that is no longer useful.

That’s an important skill. The world provides a never-ending source of data, much of which is irrelevant to the tricky business of survival, and most of which is impossible to store in a limited memory. So humans and other creatures have evolved ways to retain important skills while forgetting irrelevant ones.

Neural networks are now being taught how to forget.

The same cannot be said of machines. Any skill they learn is quickly overwritten, regardless of how important it is. There is currently no reliable mechanism they can use to prioritize these skills, deciding what to remember and what to forget.

Today that looks set to change thanks to the work of Rahaf Aljundi and pals at the University of Leuven in Belgium and at Facebook AI Research. These guys have shown that the approach biological systems use to learn, and to forget, can work with artificial neural networks too. 

The key is a process known as Hebbian learning, first proposed in the 1940s by the Canadian psychologist Donald Hebb to explain the way brains learn via synaptic plasticity. Hebb’s theory can be famously summarized as “Cells that fire together wire together.”

In other words, the connections between neurons grow stronger if they fire together, and these connections are therefore more difficult to break. This is how we learn—repeated synchronized firing of neurons makes the connections between them stronger and harder to overwrite.

So Aljundi and co have developed a way for artificial neural networks to behave in the same way. They do this by measuring the outputs from a neural network and monitoring how sensitive they are to changes in the connections within the network.

This gives them a sense of which network parameters are most important and should therefore be preserved. “When learning a new task, changes to important parameters are penalized,” say the team. They say the resulting network has “memory aware synapses.”

They’ve put this idea through its paces with a set of tests in which a neural network trained to do one thing is then given data that trains it to do something else. For example, a network trained to recognize flowers is then shown birds. The researchers then show it flowers again to see how much of this skill is preserved.

Neural networks with memory aware synapses turn out to perform better in these tests than other networks. In other words, they preserve more of the original skill than networks without this ability, although the results certainly allow room for improvement

The key point, though, is that the team has found a way for neural networks to employ Hebbian learning. “We show that a local version of our method is a direct application of Hebb’s rule in identifying the important connections between neurons,” say Aljundi and co.

That has implications for the future of machine learning. If these scientists can make their version of Hebbian learning better, it should make machines more flexible in their learning. And that will allow them to better adapt to the real world.

Ref: : Memory Aware Synapses: Learning What (Not) To Forget

Deep Dive

Artificial intelligence

chasm concept
chasm concept

Artificial intelligence is creating a new colonial world order

An MIT Technology Review series investigates how AI is enriching a powerful few by dispossessing communities that have been dispossessed before.

open sourcing language models concept
open sourcing language models concept

Meta has built a massive new language AI—and it’s giving it away for free

Facebook’s parent company is inviting researchers to pore over and pick apart the flaws in its version of GPT-3

spaceman on a horse generated by DALL-E
spaceman on a horse generated by DALL-E

This horse-riding astronaut is a milestone in AI’s journey to make sense of the world

OpenAI’s latest picture-making AI is amazing—but raises questions about what we mean by intelligence.

labor exploitation concept
labor exploitation concept

How the AI industry profits from catastrophe

As the demand for data labeling exploded, an economic catastrophe turned Venezuela into ground zero for a new model of labor exploitation.

Stay connected

Illustration by Rose WongIllustration by Rose Wong

Get the latest updates from
MIT Technology Review

Discover special offers, top stories, upcoming events, and more.

Thank you for submitting your email!

Explore more newsletters

It looks like something went wrong.

We’re having trouble saving your preferences. Try refreshing this page and updating them one more time. If you continue to get this message, reach out to us at with a list of newsletters you’d like to receive.