CA3 sparsity stabilises high-connectivity recurrent autoassociation: complementary binary and spiking computational modes in a DG->CA3 model

The brain's ability to remember is built on a delicate balance between two opposing tasks. To store a new memory, the brain must first separate it from similar past experiences, ensuring that a new event is not confused with an old one. This process, known as pattern separation, acts like a filing system that keeps similar documents in different folders. Once a memory is stored, the brain must be able to retrieve it even when given only a vague or incomplete hint. This retrieval process, called pattern completion, allows a single scent or a fragment of a sound to bring a whole scene back to mind. These two operations happen in neighboring regions of the hippocampus, a seahorse-shaped structure deep inside the brain. The first region, the dentate gyrus, performs the separation, while the second, area CA3, performs the completion.

For decades, scientists have debated how the CA3 region manages to retrieve memories so reliably. The key to this debate lies in how densely the neurons in this area are connected to one another. Some measurements suggest these connections are rare, while others, using more advanced imaging, suggest they are ten times more common. This disagreement has left researchers wondering: does having more connections help the brain recall memories, or does it cause chaos? A new study by Tadanobu Chuyo Kamijo and colleagues uses computer simulations to resolve this conflict. They found that the answer depends entirely on how the neurons behave. If the neurons are kept very quiet and sparse, a highly connected network works beautifully. But if the neurons become too active, that same high connectivity causes the system to collapse, making it impossible to distinguish one memory from another.

The researchers built a detailed computer model of the pathway from the dentate gyrus to CA3 to test how these systems handle memory under different conditions. They created two versions of the CA3 region to represent different ways scientists have modeled it. One version used a simplified, binary approach where neurons are either strictly "on" or "off," representing a system with hard limits on activity. The other version used a more realistic spiking model, where neurons fire electrical signals and regulate their own activity through a balance of excitation and inhibition, much like a real biological network. They then fed both versions a stream of inputs and tested their ability to complete partial memories while varying the number of connections between neurons.

The results revealed a striking difference between the two models. In the simplified binary version, increasing the number of connections always helped. The more links the neurons had, the better the system became at retrieving full memories from partial cues. This suggests that if the brain operates with strict limits on activity, having a dense web of connections is a clear advantage. However, the realistic spiking model told a different story. In this version, increasing connections only helped up to a point. Once the number of connections grew too large relative to the number of active neurons, the system failed. The memories did not just get slightly fuzzy; they collapsed entirely. The network became unable to distinguish between different stored patterns, effectively losing the memory.

The study identified a specific rule that governs this failure. The system remains stable only when the product of the number of active neurons and the number of connections per neuron stays below a certain threshold. In simple terms, if the neurons fire too often, the high number of connections causes them to all light up together, blurring the distinct patterns of memory. The researchers found that this collapse happens even if the network tries to compensate by strengthening its internal brakes, or inhibition. The failure is not a runaway explosion of activity, but a loss of specificity where the brain can no longer tell one memory from another. This explains why the brain likely keeps the activity of CA3 neurons very low, with only about two to five percent of them active at any given time. This extreme sparsity is the condition that allows a highly connected network to function without falling apart.

The team also explored how adult neurogenesis, the birth of new neurons in the adult brain, fits into this picture. New neurons are known to be more excitable than mature ones, which could potentially disrupt the delicate balance of the network. The simulations showed that if these new neurons simply became more active without changing their connections, they would push the system over the edge, causing the memory network to collapse. However, if these new neurons also recruited inhibitory cells to help quiet the network, they actually helped stabilize the system. This suggests that the brain's method of integrating new cells is critical: if the new cells help keep the overall activity low, they protect memory; if they just add noise, they destroy it.

This work reframes the long-standing debate about how many connections exist in the CA3 region. The disagreement between low and high estimates may not be a contradiction in measurement, but rather a reflection of different operating modes. If the brain relies on a sparse, controlled code, it can support a very high density of connections, which would align with the higher estimates. If the code were less sparse, the brain would need far fewer connections to avoid chaos. The study suggests that the brain's ability to recall memories is not just about having the right number of wires, but about keeping the traffic on those wires light enough to prevent a gridlock. By keeping activity sparse, the brain can maintain a rich, highly connected network capable of stable and precise memory retrieval.