New IDP model rethinks how our brains actually retrieve memories

Ever wondered how hearing just the first few notes of a familiar song can instantly bring the entire melody to mind? This marvel of cognition is known as associative memory, a fundamental neural mechanism that allows us to link pieces of information, retrieve complete patterns from partial cues, and navigate the complexities of our world.

Now, researchers are proposing a new model that offers deeper insights into how our brains achieve this, particularly how external stimuli actively guide the memory retrieval process – an aspect they believe has been largely overlooked in traditional AI models of memory.

The power and puzzle of associative memory

Associative memory is not just about recalling songs; it’s a cornerstone of learning, problem-solving, and our general ability to make sense of reality. When one piece of information triggers the memory of a larger, related pattern – like a scent evoking a childhood memory, or a single word bringing forth a complex concept – associative memory is at work. This isn’t a simple one-to-one storage system.

“It’s a network effect,” explained UC Santa Barbara mechanical engineering professor Francesco Bullo. “Memory storage and memory retrieval are dynamic processes that occur over entire networks of neurons.” These intricate neural networks allow for the robust and flexible recall that characterizes human memory.

In 1982, physicist John Hopfield, who was awarded the Nobel Prize for his work in 2024, famously translated this neuroscience concept into the realm of artificial intelligence with the Hopfield network. This was a landmark achievement, providing a mathematical framework to understand memory storage and retrieval. The Hopfield network, one of the first recurrent artificial neural networks, became renowned for its ability to retrieve complete patterns even when presented with noisy or incomplete inputs, much like our own brains.

The missing role of external inputs

Despite the power of the traditional Hopfield network, Professor Bullo and his collaborators—Simone Betteti, Giacomo Baggio, and Sandro Zampieri at the University of Padua in Italy—argue that it doesn’t fully capture the nuances of how new, incoming information steers the memory retrieval process. In their paper published in the journal Science Advances, they state, “Notably, the role of external inputs has largely been unexplored, from their effects on neural dynamics to how they facilitate effective memory retrieval.”

The researchers suggest that current AI models, including the very sophisticated Large Language Models (LLMs), don’t truly replicate the way animal brains, including human ones, handle memories. “The modern version of machine learning systems, these large language models — they don’t really model memories,” Bullo explained. “You put in a prompt and you get an output. But it’s not the same way in which we understand and handle memories in the animal world.” While LLMs can generate impressively coherent and intelligent-sounding responses based on the vast patterns in their training data, they lack the continuous, experience-based reasoning grounded in the physical world that animals possess.

Simone Betteti, lead author of the paper, elaborated on this point: “The way in which we experience the world is something that is more continuous and less start-and-reset.” He noted that many previous treatments of the Hopfield model adopted a mechanistic, computer-like perspective of the brain. “Instead, since we are working on a memory model, we want to start with a human perspective.” The central question driving their research was: As we continuously perceive the world around us, how do these incoming signals enable us to retrieve relevant memories?

AI may soon detect dyslexia early from children’s handwriting

Hopfield’s original model conceptualizes memory retrieval using an “energy landscape” metaphor. In this landscape, memories are represented by valleys, or energy minima. The process of memory retrieval is akin to a ball rolling across this landscape until it settles into one of these valleys, signifying recognition or recall. Your starting point on this landscape is your initial condition, influenced by the cue you receive.

“Imagine you see a cat’s tail,” Bullo illustrated. “Not the entire cat, but just the tail. An associative memory system should be able to recover the memory of the entire cat.” According to the traditional Hopfield model, the stimulus of the cat’s tail acts as an initial condition, placing your “ball” closest to the “cat” valley. However, a critical question remained.

“The classic Hopfield model does not carefully explain how seeing the tail of the cat puts you in the right place to fall down the hill and reach the energy minimum,” Bullo pointed out. “How do you move around in the space of neural activity where you are storing these memories? It’s a little bit unclear.”

Introducing the Input-Driven Plasticity (IDP) model

To address this lack of clarity, the researchers propose their Input-Driven Plasticity (IDP) model. This new framework introduces a mechanism where past and new information are gradually integrated, actively guiding the memory retrieval process towards the correct memory. Unlike the somewhat static energy landscape of the original Hopfield network where retrieval is a two-step algorithmic process, the IDP model describes a dynamic, input-driven mechanism.

“We advocate for the idea that as the stimulus from the external world is received (e.g., the image of the cat tail), it changes the energy landscape at the same time,” Bullo stated. In essence, the external input – the cat’s tail – doesn’t just give you a starting point; it actively reshapes the terrain. “The stimulus simplifies the energy landscape so that no matter what your initial position, you will roll down to the correct memory of the cat.”

A significant advantage of the IDP model is its robustness to noise. When an input is vague, ambiguous, or partially obscured, the model doesn’t just falter. Instead, it can effectively use this “noise” to its advantage, filtering out less stable memories (the shallower valleys in the energy landscape) and favoring the more stable, deeply entrenched ones.

Betteti provided an analogy: “We start with the fact that when you’re gazing at a scene your gaze shifts in between the different components of the scene. So at every instant in time you choose what you want to focus on but you have a lot of noise around.” Once you “lock into” an input to focus on, he explained, the network dynamically adjusts itself to prioritize that input, effectively sculpting the memory landscape in real-time.

Transformers are the engines behind today’s leading LLMs like ChatGPT.

While the IDP model proposed by Bullo and his colleagues “starts from a very different initial point with a different aim,” there’s exciting potential for this new understanding of associative memory to inform the design of future machine learning systems. The way the IDP model handles continuous input and dynamically adjusts its internal state could offer new pathways for creating AI that learns and reasons in a more human-like, context-aware manner.

“We see a connection between the two, and the paper describes it,” Bullo said. “It is not the main focus of the paper, but there is this wonderful hope that these associative memory systems and large language models may be reconciled.”

Featured image credit