Memory's Geometry: A New Lens on Forgetting and False Memories

Why does our memory sometimes betray us? It's a question that's puzzled scientists for decades. The conventional wisdom points to the limitations of our biological hardware. But what if the real answer lies not in our biology but in the geometry of information?

The Geometry of Memory

Recent findings suggest that high-dimensional embedding spaces, rather than biological decay, may hold the key to understanding memory phenomena. These spaces, prone to noise and interference, mimic the quirks of human memory without any specific engineering geared toward simulating these quirks. The study found that power-law forgetting, with a coefficient near 0.46, aligns closely with the human experience of memory decay, traditionally pegged at around 0.5. This suggests interference plays a significant role over time-dependant decay. Without competitors, the decay function drops dramatically to about 0.009, indicating that competition, not time, drives forgetting in this context.

False Memories Uncovered

Even more intriguing is the revelation about false memories. The researchers discovered that false memories require no specialized programming. Basic cosine similarity on pre-trained embeddings generates false alarm rates akin to human experiences, matching closely with the Deese, Roediger, McDermott false alarm rate of approximately 0.55. It's a stark reminder that the creation of false memories could be an intrinsic property of any system that organizes and retrieves information based on meaning and proximity.

The Big Picture

So, what does this mean for our understanding of memory? At its core, these findings challenge the notion that memory quirks are bugs in our biological code. Instead, they appear to be features of any system that handles information similarly. If both human and machine memory can be explained by the same geometric principles, what does this say about the future of cognitive AI?

In a world where AI and human cognition increasingly converge, it's essential to ask: Are we ready to embrace the complexity and potential pitfalls of an AI that mirrors our own cognitive biases and strengths? The AI-AI Venn diagram is getting thicker, with implications not just for how we understand memory, but for the very nature of intelligence.