Abstract
We investigate the performance of sparsely-connected networks of integrate-and-fire neurons for ultra-short term information processing. We exploit the fact that the population activity of networks with balanced excitation and inhibition can switch from an oscillatory firing regime to a state of asynchronous irregular firing or quiescence depending on the rate of external background spikes. We find that in terms of information buffering the network performs best for a moderate, non-zero, amount of noise. Analogous to the phenomenon of stochastic resonance the performance decreases for higher and lower noise levels. The optimal amount of noise corresponds to the transition zone between a quiescent state and a regime of stochastic dynamics. This provides a potential explanation on the role of non-oscillatory population activity in a simplified model of cortical micro-circuits.
Introduction
The mammalian brain processes sensory information that changes on extremely rapid timescales, such as saccadic eye movements or phonemic transitions in speech. While traditional neural network models excel at explaining long-term memory through stable attractors, they often struggle to account for "ultra-short-term" memory—the ability to maintain a temporal trace of a signal over 10–100 milliseconds. This study explores how cortical micro-circuits function as "liquid" buffers, specifically examining how background noise and signal autocorrelation influence the network's capacity to retain past information.
Methods
The study utilizes a simulated network of 200 Leaky Integrate-and-Fire (IF) neurons. The network architecture consists of 80% excitatory and 20% inhibitory neurons with a sparse connectivity ratio of 0.2.
The individual neuron dynamics are governed by standard parameters, including a membrane time constant of 20 ms and a 2 ms absolute refractory period.
Two noise models were evaluated:
- Model A: Driven by purely excitatory Poisson spikes.
- Model B: Driven by a balanced input of excitatory and inhibitory spikes, allowing for independent control of input variance (noise) and mean.
Information buffering was quantified by attempting to reconstruct an input signal from the past (time $t - T$) using the current membrane potentials at time $t$. Reconstructions were performed using both simple linear readouts and non-linear Support Vector Machines (SVM) with Gaussian kernels. Synthetic input signals with triangular autocorrelation profiles were used to isolate the network's buffering capacity from the signal's inherent predictability.
Results
The Role of Noise and Stochastic Resonance
A primary finding is that network performance is maximized in a "Goldilocks zone" of moderate noise. This optimal state occurs near the transition between a quiescent state and a regime of asynchronous irregular firing. Insufficient noise leaves the network unresponsive, while excessive noise or rhythmic oscillations destroy the temporal trace of the signal. This behavior is a network-level manifestation of stochastic resonance.
Dimensionality and Readout
The recurrent network acts as a high-dimensional filter, projecting a low-dimensional input into a 200-dimensional state space. This "dimensional blow-up" facilitates information extraction; results showed that a simple linear readout performed nearly as well as a complex SVM, indicating that the network's intrinsic dynamics perform the necessary computational "heavy lifting."
Performance Metrics
In Model A, reconstruction error was minimized at an external background spiking rate of approximately 2 Hz. The network demonstrated the ability to buffer information for tens of milliseconds and could process up to 16 independent input signals simultaneously with high accuracy. Furthermore, a linear relationship was observed between the input signal's autocorrelation and the effective buffer depth, suggesting the network extends the natural persistence of incoming data.
Discussion
While the 200-neuron model provides insights into cortical micro-columns, future work should address potential finite-size effects in larger networks. The current model uses static synapses; incorporating short-term synaptic plasticity could significantly enhance buffering capacity. Additionally, more complex neuron models (e.g., Hodgkin-Huxley) might reveal further nuances in how ion channel dynamics contribute to information flow. These findings suggest that the brain may actively tune its internal noise levels to optimize information processing for specific tasks.
Summary updated