AI needs to be able to Forget

When we discuss AI, almost everyone talks about the ability to learn - Supervised learning, Unsupervised learning, Reinforcement learning, Deep learning, and so on. Although one of the key traits that makes humans intelligent is not learning but the ability to forget.

• Forgetting reduces the influence of outdated information
• Forgetting enables us to make a bet in spite of specific past events
• Forgetting is inherent for creative thinking and imagination
• And, forgetting is a key strategy for adaptation as well

Sometimes, we forget naturally, and other times we make a strategic decision to forget (and yes, this is a request to those of you who have not pardoned my past mistakes yet). They both count.? However, most AI techniques today focus on learning and not on forgetting.?

It is a remarkable oversight that the field of AI has actually never taken forgetting seriously.

It is a remarkable oversight that the field of AI has actually never taken forgetting seriously. Most AI techniques do not address:

• Reducing the influence of outdated information (that was used for building the model)
• Strategically adapting prediction based on the new information
• A creative solution that is not learnt from data

Before we discuss how we can enable AI forget, let's take a look at forgetting happening inside our brain.

How do People Forget?

Ever since the discovery that the brain has a vast network of billions of nerve cells that communicate with each other via junctions called ‘synapses’, the role that synapses play in learning and memory has been a matter of extensive research. ? We continually build, lose, alter, and rebuild these synaptic connections.? A memory is nothing but a group of synaptic connections and the physical representation of the memory like this is referred to as engrams. Building or altering the engrams is the foundation of learning and memory.?

Forgetting happens when these engrams break down or become harder to access:

• New synaptic connections overriding old connections
• Decay of the synaptic connections between neurons
• Random death of some neurons

Most past theories about forgetting have emphasized on forgetting happening as a passive process, however, new research in the past decade, suggests that the loss of memories is not a passive process. Rather, the brain is built to forget.? Forgetting is an active mechanism that is constantly at work in the brain.

Forgetting is an active mechanism that is constantly at work in the brain.

• The brain actively works to forget information of lower value (i.e not useful, valuable, conflicting or relevant)?
• There is a group of neurons in the brain that are typically active when we are sleeping in R.E.M. stage are responsible for active forgetting in the brain.

There is a lot that we still don't know about how the forgetting process works and perhaps thats the reason that AI is significantly less intelligent than humans.

How to make AI Forget?

Well the “traditional” BIG data AI learning is all about learning.? There is no concept of forgetting.? In most cases the technique to forget is by learning again with the whole new data set.

For example, let's assume you have built a model to learn handwritten digits from 0 to 9.? You can do it with any handwritten digit data set.? Let's assume this dataset had all 7s written without a line through it.? And, now you want to update this model to relearn how 7s are written.? You want to make it forget that 7 is written without a line through it and only remember 7 with a line through it.? You have a new data set of 7s with a line through them available but you can’t access the old dataset.

We need the ability to incrementally forget or acquire new knowledge.? This makes it interesting.? You want to change the weights of your model but do it in such a way that your model can predict new knowledge, forget selective old knowledge, and be stable for existing knowledge.?

To have the ability to learn in continuous environments, we need to see beyond the current AI techniques.? How you will do weight regularization, define acceptable forgetting levels, current model protection, or even rehearsal to verify the impact of new learning in absence of old data to learn.? Current AI techniques result in models that are “stable” (i.e. perform well based on the data they are built upon) but not “plastic” (i.e. able to learn/forget incremental without disrupting what is learnt previously).? We need to find a technique that provides the right balance between stability and plasticity.

Future of AI running on less data or sequential data is not far. If you are struggling with lack of incremental learning in your AI solutions, or would like to brainstorm on how SECOMIND does this, please feel free to reach out to me.