A balanced memory network

A fundamental problem in neuroscience is understanding how working memory – the ability to store information at intermediate timescales, like 10s of seconds – is implemented in realistic neuronal networks. The most likely candidate mechanism is the attractor network, and a great deal of effort has gone toward investigating it theoretically. Yet, despite almost a quarter century of intense work, attractor networks are not fully understood. In particular, there are still two unanswered questions. First, how is it that attractor networks exhibit irregular firing, as is observed experimentally during working memory tasks? And second, how many memories can be stored under biologically realistic conditions? Here we answer both questions by studying an attractor neural network in which inhibition and excitation balance each other. Using mean field analysis, we derive a three-variable description of attractor networks. From this description it follows that irregular firing can exist only if the number of neurons involved in a memory is large. The same mean field analysis also shows that the number of memories that can be stored in a network scales with the number of excitatory connections, a result that has been suggested for simple models but never shown for realistic ones. Both of these predictions are verified using simulations with large networks of spiking neurons.

💡 Research Summary

This paper tackles two long‑standing puzzles in the neuroscience of working memory: (1) why attractor networks, which are the leading theoretical framework for persistent activity, display the irregular, near‑Poisson spiking observed experimentally during memory tasks, and (2) how many distinct memory patterns can be stored under biologically realistic constraints. The authors propose a “balanced memory network” in which excitatory and inhibitory currents are tightly matched, and they analyze it using a mean‑field reduction that collapses the high‑dimensional spiking dynamics into three coupled order parameters: the average excitatory drive, the average inhibitory drive, and the memory‑specific activation variable.

The mean‑field equations reveal two key insights. First, irregular firing emerges only when the number of neurons that participate in a given memory (N_mem) is sufficiently large—specifically, when N_mem greatly exceeds the square root of the total number of excitatory connections per neuron. In this regime each neuron receives many weak, fluctuating inputs, and the global inhibitory feedback stabilizes the mean membrane potential near threshold, producing spike trains whose inter‑spike intervals follow an exponential distribution, as seen in vivo. If N_mem is small, the network either locks into a highly synchronized state or exhibits overly regular firing, contradicting experimental data.

Second, the storage capacity scales linearly with the total number of excitatory synapses (K_E) in the network. By examining the overlap terms that couple different stored patterns, the authors show that the inhibitory feedback suppresses cross‑talk proportionally to K_E, allowing the number of reliably retrievable memories C to satisfy C ≈ α K_E, where α is a modest constant determined by the balance of excitation and inhibition. This linear scaling had been demonstrated for highly simplified binary models, but here it is derived for a realistic network of leaky‑integrate‑and‑fire neurons with conductance‑based synapses and Hebbian learning.

To validate the theory, large‑scale simulations were performed with 10⁴–10⁵ spiking neurons. The simulations confirmed that, within the predicted parameter region (appropriate inhibitory gain, sufficient N_mem, and realistic connection density), the network settles into stable attractor states while each neuron’s spike train remains irregular and approximately Poisson. Moreover, when the number of stored patterns was increased, the recall performance degraded only when the total number of patterns approached the linear bound set by K_E, matching the analytical prediction. In contrast, removing the global inhibitory coupling eliminated the irregularity and caused a precipitous drop in capacity, underscoring the essential role of E/I balance.

The discussion connects these findings to experimental observations in prefrontal cortex and other areas implicated in working memory, suggesting that the brain may exploit a finely tuned excitation‑inhibition balance to achieve both robust persistent activity and the stochastic firing that supports flexible coding. The authors also speculate that disorders characterized by disrupted E/I balance, such as schizophrenia or ADHD, could impair working memory by violating the conditions identified here. Limitations include the absence of long‑term synaptic plasticity dynamics and the lack of spatially structured connectivity, which are earmarked for future work.

Overall, the paper provides a rigorous theoretical framework that reconciles irregular spiking with high storage capacity in attractor networks, and it demonstrates that these properties naturally arise when excitation and inhibition are balanced—a principle that may be fundamental to cortical working memory circuits.