Why do police officers face challenges when questioning accident witnesses?
Humans possess a remarkable ability to remember significant life events. They can recall what they did during their last vacation, their college major, and who delivered the funniest joke at an office party. This type of memory is known as episodic memory.
What’s fascinating about it is that it involves simulation. We recreate past experiences by simulating an approximation of them while engaging specific neural circuits. The same networks are activated when we envision future scenarios.
During the simulation of past events, our sensory neocortex fills in gaps in our memories, much as we can complete visual shapes in an incomplete drawing. However, unlike recognisable visual shapes, the line between an imagined scene and an actual episodic memory is very fine within the neocortex. There are indeed accident witnesses who describe events that never took place because they are convinced they did. Their certainty can hinder the accurate reconstruction of what truly happened.
In mammals, memories are accessed by providing fragments of the original experience, which reactivate the original pattern. For example, if you hear the start of a story you’ve previously encountered, you can continue it. If you see half of a cat picture, you can visualise the rest. However, since these memories are merely simulations (rather than the actual events), inaccuracies may arise. On the other hand, mammalian brains do not experience catastrophic forgetting. Auto-association indicates that their brains use content-addressable memory, which is stored across a population of neurons.
Computers operate differently. They use register-addressable memory, which can only be accessed if the user knows its specific ‘memory address.’ If you happen to lose this address, the memory becomes irretrievable. This type of memory allows computers to organise where data is stored, preventing new information from overwriting old data.
In the brains of mammals, episodic memory arises from a collaboration between the older hippocampus and the more recent neocortex. While the hippocampus is adept at recognising patterns, it cannot simulate the environment. Conversely, the neocortex can simulate intricate details of the world but struggles to learn new patterns effectively. The hippocampus specialises in quickly identifying locations. Research with rats shows that these animals can mentally re-activate hippocampal cells to navigate different paths in a maze. They employ ‘memory codes’ in the hippocampus to simulate past experiences, provided they have previously encountered the maze.
In artificial intelligence, neural networks tend to lose old patterns as they acquire new ones. The collaboration between the hippocampus and the neocortex is not entirely mirrored in intelligent machines. The hippocampus supports the neocortex by retrieving and replaying recent and older memories, allowing new memories to be integrated without interfering with existing ones. In the realm of AI, this is referred to as ‘generative replay’ or ‘experience replay.’ These methods help mitigate or smooth out catastrophic forgetting, though they do not entirely eliminate it.
Catastrophic forgetting poses a significant challenge for models built on deep learning and transformer architectures. This issue is a major reason why AI systems do not communicate with users; instead, they remain static after their initial training. Updates occur only when AI labs and companies decide to retrain the model with new data. Once these updates are integrated, the model reverts to its inactive state before being delivered to us, AI Muggles.