Reshaping Neural Representation via Associative, Presynaptic Short-Term Plasticity

TL;DR

This work develops a normative, information-theoretic account of associative short-term plasticity by extending Fisher-information learning to Tsodyks–Markram synapses. It derives analytic update rules for baseline strength $w_{ij}^0$ and release probability $U_{ij}$ that maximize stimulus information under resource constraints, revealing a postsynaptic term plus a phase-advanced presynaptic term that detects onset. The presynaptic component induces anti-causal connectivity and ramp-like representations, with the strength and direction of temporal asymmetry modulated by a global release-budget constraint, potentially explaining state-dependent replay in the hippocampus. Linear-response analysis shows frequency-dependent phase selectivity in presynaptic drive, suggesting presynaptic plasticity as a substrate for rapidly reconfigurable temporal coding with tangible implications for memory encoding and retrieval.

Abstract

Short-term synaptic plasticity (STP) is traditionally viewed as a purely presynaptic filter of incoming spike trains, independent of postsynaptic activity. Recent experiments, however, reveal an associative form of STP in which presynaptic release probability changes alongside long-term potentiation, implying a richer computational role for presynaptic plasticity. Here we develop a normative theory of associative STP using an information-theoretic framework. Extending Fisher-information-based learning to Tsodyks-Markram synapses, we derive analytic update rules for baseline synaptic strength and release probability that maximize encoded stimulus information under resource constraints. The learning rules separate into a conventional postsynaptic term tracking local firing and a distinct presynaptic term with a phase-advanced structure that selectively detects stimulus onset; critically, differences between plasticity of baseline strength and release probability arise within this presynaptic component. For stimulus variations slower than the EPSP time constant, onset sensitivity biases optimal connectivity toward anti-causal associations, strengthening synapses from neurons activated later to those activated earlier. In recurrent circuits, these rules yield ramp-like sustained representations and reverse replay after drive removal. Linear-response analysis further shows that STP confers frequency-dependent phase selectivity on presynaptic drive and that constraints on total release probability systematically tune temporal asymmetry. Together, our results provide a principled account of associative STP and identify presynaptic plasticity of release probability as a substrate for rapidly reconfigurable temporal coding.

Figures (16)

• Figure 1: Response of STP-derived sensitivity functions to step / sinusoidal inputs.A-C. Response to step input. A. Presynaptic firing rate $\nu_j(t)$. B. Dynamics of sensitivity functions $f^Z$. Note that $f^U$ (orange) decays more strongly than $f^{w_0}$ (blue). C. Dynamics of the effective presynaptic contribution $C(t) = f(t)\nu_j(t)$. D-F. Response to sinusoidal input. The effective presynaptic term (F) peaks during the rising phase of the input (D), illustrating the onset-detection property of the learning rule. Parameters: $\tau_d = 0.5$ s, $U = 0.15$.
• Figure 2: Frequency-domain response of the effective presynaptic term $C^{Z}(t)$. Response to a modulated rate $\nu(t) = \nu_0 + \delta\nu \cos(\omega t)$. Upper. Amplitude gain $|H_{C_{pre}}^{Z}(\omega)|$. Lower. Phase lag $\phi_{C_{pre}}^{Z}(\omega) = \arg H_{C_{pre}}^{Z} (\omega)$. Note the positive phase shift (lead) in the intermediate frequency range, indicating sensitivity to the input onset. The lead is more pronounced for $U$ (orange) than for $w_0$ (blue). Lines show the predictions by linear response theory, and dots show the results by numerical simulations. Parameters: $\nu_0 = 10$ Hz, $\tau_d = 0.5$ s, $U = 0.15$.
• Figure 3: Analysis of pre- and postsynaptic factors (for representative anti-causal pair). Pre- and postsynaptic neurons receive the same traveling-wave external input $h(z, t) = A[\cos(\omega t - z) - \cos\theta_c]_+$ at different phases. The plots illustrate a representative anti-causal offset ($\zeta>0$; postsynaptic activation precedes presynaptic activation), where the pre/post overlap is large. The STP factors $f^{w_0}$ and $f^U$ decay rapidly with increasing firing rate, so the presynaptic contributions $C_\mathrm{pre}^{w_0}$ and $C_\mathrm{pre}^U$ are dominated by stimulus onset. The contribution to the Fisher-information gradient is determined by the temporal overlap between $C_\mathrm{pre}$ and the postsynaptic term $C_{post}$. Parameters: $\tau_d = 0.5~\mathrm{s}$, $\beta = 2.0$, $g_M = 10.0$, $u_c = 1.0$, $\tau_s = 0.01~\mathrm{s}$, $A = 3.0$, $\theta_c = 0$.
• Figure 4: Gradient of Fisher information with respect to $U(\zeta)$. Heat map of $\frac{\delta J}{\delta U(\zeta)}$ as a function of phase difference $\zeta$ and $U$. White contour lines indicate optimal solutions under the constant-sum constraint. The red contour marks the boundary where $\frac{\delta J}{\delta U(\zeta)} = 0$. Parameters are identical to Figure \ref{['fig:pre-post_factors']}.
• Figure 5: Optimal synaptic profiles.A. Optimal release probability $U(\zeta)$. B. Optimal baseline weight $w_0(\zeta)$. Green, solid: optimization with plastic $U(\zeta)$. Yellow, dashed: optimization with non-plastic $U$. C. Optimal effective weight $w(\zeta) = w_0(\zeta) U(\zeta)$. Purple, solid: optimization with plastic $U(\zeta)$. Red, dashed: optimization with non-plastic $U$. Optimal $U(\zeta)$ is derived under the constraint $\frac{1}{2\pi} \int _{-\pi}^\pi d\zeta\, U(\zeta) = 0.15$. In the non-plastic condition, $U$ is fixed to 0.15. Other parameters are identical to Figure \ref{['fig:grad-U']}.