The hippocampal CA3 area is widely held to perform a core memory function: the retrieval of previously stored, distributed patterns of neural activity via its recurrent synaptic circuitry. While network-level models have emphasized the role of recurrent connections in pattern completion, it has remained unclear how the intrinsic physiological properties of individual CA3 neurons influence the stability and reliability of memory recall.
The present work addresses this gap by combining mathematical analysis of a canonical recurrent neural network model with direct intracellular recordings from CA3 pyramidal cells in behaving mice. The study focuses on the neuronal activation function—the relationship between input current and output firing rate—and how its shape affects the network’s ability to stably recall stored patterns.
The authors analyzed a standard recurrent network model of CA3 to derive conditions that enable stable memory retrieval. From this analysis they formulated three experimentally testable predictions about the physiological state of recurrent circuits that perform reliable recall. Two of these predictions match prior empirical observations: that pyramidal neurons should exhibit elevated intrinsic excitability and that recurrent connectivity should be inhibition-dominated. The third prediction, which the authors tested experimentally in this study, concerns the exponent of the neuronal activation function: theory predicts an exponent slightly above 1 for neurons engaged in optimal, stable recall.
To evaluate the third prediction, the investigators performed in vivo intracellular recordings from CA3 pyramidal neurons in behaving mice. A total of 49 pyramidal cells were recorded and their subthreshold and spiking voltage traces were collected. These recordings provide direct measurements of the input–output relationships of CA3 neurons in an intact, behaving preparation.
The recorded voltage traces were fitted with a range of parametric activation models to estimate the shape and exponent of each neuron's activation function. The authors performed statistical model comparison across the candidate parametric forms to identify which model best described the data on each cell. The primary quantitative outcome was the estimated activation exponent across cells.
Across the 49 recorded CA3 pyramidal neurons, the average estimated activation exponent was slightly above 1, in agreement with the theoretical prediction that an exponent just above linearity supports stable recall in the recurrent network model. The analysis also revealed substantial heterogeneity in exponent values between cells; the origin of this variability is not fully accounted for by the presented theory and is noted as an open question.
To test the generality of their findings, the authors analyzed activation exponent estimates derived from 133 additional cells reported in two earlier studies. The exponent estimates from these prior datasets were highly consistent with those obtained from the new CA3 recordings, reinforcing the conclusion that CA3 pyramidal neurons commonly exhibit activation exponents slightly greater than 1.
Together, the theoretical and empirical results suggest that the physiological properties of CA3 pyramidal neurons are tuned to support reliable pattern recall in recurrent hippocampal circuits. The finding that the activation exponent is slightly above 1 provides a candidate single-neuron mechanism contributing to network-level stability during memory retrieval.
However, the study also highlights heterogeneity across cells in activation exponents. The sources and functional consequences of this heterogeneity remain to be explained within the theoretical framework. Additionally, while two of the model’s predictions align with prior work (elevated intrinsic excitability and inhibition-dominated recurrence), testing how these features interact with activation exponent variability in vivo will require further experimental and modeling work.
The authors declare no competing interests. The work acknowledges funding from the Wellcome Trust and the Human Frontiers Science Programme. The preprint is posted on bioRxiv (doi reported in the source). Specific methodological details, parameter values for the mathematical model, and full statistical procedures were reported in the source article; readers should refer to the original preprint for those experimental and analytical specifics.