Peter Cochrane: Is forgetting essential for learning?

Wouldn't it be great if you could remember everything you ever learnt, saw or heard? Actually, it wouldn't, writes Professor Peter Cochrane

For decades ‘brain decay' due to dying synapses and reduced connectivity was seen as an inevitable consequence of old age. But more recently we have realised that healthy brains subject to continual stimulation remain almost as good as new until we die. So, why do we become absent minded and forget stuff?

First, our brains have a finite storage capacity (around one petabute) and are far more analogue than digital. This sees them subject to various information decay and distortion mechanisms. A useful demonstration of this is to recall some memorable event and then search your photograph collection to compare. The difference is generally stark!

Forgetting is an essential component of learning, and people who exhibit a lack of ‘forgetting' tend to be very limited in their flexibility, adaptability and creativity

Second, without a continual chemical, electrical and mechanical refresh, ‘Hebbian decay' sees an exponential fall in the strength of the detail recorded by each neuron. So; walking, talking, and writing are generally reinforced day-on-day, while the ability to surf, roller-skate, play golf (to an acceptable handicap) rapidly decay if we do not continually practice.

It's the same with mathematics, history, languages and our recall of events and images. But forgetting provides us with two key benefits:

• Mental agility/flexibility is improved by removing outdated memories and information cluttering our decision-making processes;
• Historically induced preconceptions and over-strong biases are minimised to reduce erroneous generalisations.

Ideally, we need to selectively and strategically remember in order to be sufficiently flexible, adaptive and creative.

This might seem counter-intuitive, and contrasts to my ‘notionally ideal' state: If only I could remember every detail and experience in my life, and never lose a skill or a fact - how powerful would I be?

Well, er, not quite! It turns out that forgetting is an essential component of learning, and people who exhibit a lack of ‘forgetting' tend to be very limited in their flexibility, adaptability and creativity.

Forgetting is not only necessary to make space for the new, it is a healthy mechanism for keeping relevant and up to date. But, I have to admit to using mainframes and laptops to help me remember and recall far more than any previous generation. And more recently, AI has entered the frame to expand my abilities and make me increasingly ‘super-human'.

All this raises the spectre of artificial intelligence (AI) systems that can remember everything to perfection, only to make decision errors based on old data, concepts and rules. Legal, medical and governance systems now spring to mind. We really need AI to be on the ball - up-to-date and continually adapting, and that means relegating knowledge that is old or irrelevant to its rightful place.

Fortunately; we have a number of techniques that ensures our AI systems ‘forget appropriately' or not at all. For example; deep neural networks continually reinforce the relevant and discount the old, the outdated and/or irrelevant/spurious.

Recurrent neural networks can be engineered to do the same or even better, while elastic synaptic consolidation has recently been proven and polished by Google Deep Mind.

So, you might ask, what's the big deal? Well, AI professionals remain bugged by two key features:

• We have a general inability to figure out how and why AI systems come to their decisions/conclusions… and, often, it remains a deep mystery;
• Another annoyance is that AI systems are different; they just don't think like we do.

The reality is that even now AI systems can solve problems way beyond any human ability and, provided the answers we get are valid, we probably ought to worry less about how those decisions are arrived at and celebrate more.

Forgetting might just be the key to learning for biological entities like us. But I am far more interested in exploiting a perfect (memory-loss free) AI-human partnership than I am in trying to make technologies that ape our frailties. In other words, we should seek to build AI that complements our abilities rather than mimick them.

"To me this is all like owning a hammer and refusing to use an electric drill"

Professor Peter Cochrane OBE is the former CTO of BT, who now works as a consultant focusing on solving problems and improving the world through the application of technology. He is also a professor at the University of Suffolk's School of Science, Technology and Engineering