benda/nonlinearbaseline2025
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Jan Gs great comments on manuscript and rebuttal

Browse Source
This commit is contained in:
Jan Benda
2026-02-11 11:00:31 +01:00
parent c96d318931
commit 086264121f
2 changed files with 51 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
nonlinearbaseline.tex
View File

@@ -275,7 +275,7 @@ Supported by SPP 2205 ``Evolutionary optimisation of neuronal processing'' by th
% 250 words
\section{Abstract}
Spiking thresholds in neurons or rectification at synapses are essential for neuronal computations rendering neuronal processing inherently nonlinear. Nevertheless, linear response theory has been instrumental for understanding, for example, the impact of noise or neuronal synchrony on signal transmission, or the emergence of oscillatory activity, but is valid only at low stimulus amplitudes or large levels of intrinsic noise. At higher signal-to-noise ratios, however, nonlinear response components become relevant. Theoretical results for leaky integrate-and-fire neurons in the weakly nonlinear regime suggest strong responses at the sum of two input frequencies if one of these frequencies or their sum match the neuron's baseline firing rate.
We here analyze nonlinear responses in two types of primary electroreceptor afferents, the P-units of the active and the ampullary cells of the passive electrosensory system of the wave-type electric fish \textit{Apteronotus leptorhynchus}. In our combined experimental and modeling approach we identify these predicted nonlinear responses primarily in a few low-noise P-units and in more than every second ampullary cell. Our results provide experimental evidence for nonlinear responses of spike generators in the weakly nonlinear regime. We conclude that such nonlinear responses occur in any sensory neuron that operates in similar regimes particularly at near-threshold stimulus conditions.
We here analyze nonlinear responses in two types of primary electroreceptor afferents, the P-units of the active and the ampullary cells of the passive electrosensory system of the wave-type electric fish \textit{Apteronotus leptorhynchus}. In our combined experimental and modeling approach we identify these predicted nonlinear responses only in individual low-noise P-units, but in more than half of the ampullary cells. Our results provide experimental evidence for nonlinear responses of spike generators in the weakly nonlinear regime. We conclude that such nonlinear responses occur in any sensory neuron that operates in similar regimes particularly at near-threshold stimulus conditions.
% max 120 words
\section{Significance statement}
@@ -286,7 +286,7 @@ The generation of action potentials involves a strong threshold nonlinearity. Ne
\begin{figure*}[t]
\includegraphics[width=\columnwidth]{lifsuscept}
\caption{\label{fig:lifsuscept} First- and second-order response functions of the leaky integrate-and-fire model. \figitem{A} Magnitude of the first-order (linear) response function $|\chi_1(f)|$, also known as the ``gain'' function, quantifies the response amplitude relative to the stimulus amplitude, both measured at the same stimulus frequency. $r$ marks the neuron's baseline firing rate. \figitem{B} Magnitude of the second-order (nonlinear) response function $|\chi_2(f_1, f_2)|$ quantifies the response at the sum of two stimulus frequencies. Because frequencies can also be negative, the sum is actually a difference in the upper-left and lower-right quadrants, since $f_2 + f_1 = f_2 - |f_1|$ for $f_1 < 0$. For linear systems, the second-order response function is zero, because linear systems do not create new frequencies and thus there is no response at the sum of the two frequencies. The plots show the analytical solutions from \citet{Lindner2001} and \citet{Voronenko2017} with $\mu = 1.1$ and $D = 0.001$. Note that the leaky integrate-and-fire model is formulated without dimensions, frequencies are given in multiples of the inverse membrane time constant.}
\caption{\label{fig:lifsuscept} First- and second-order response functions of the leaky integrate-and-fire model. \figitem{A} Magnitude of the first-order (linear) response function $|\chi_1(f)|$, also known as the ``gain'' function, quantifies the response amplitude relative to the stimulus amplitude, both measured at the same stimulus frequency. $r$ and $2r$ mark the neuron's baseline firing rate and its harmonic (dashed vertical lines). \figitem{B} Magnitude of the second-order (nonlinear) response function $|\chi_2(f_1, f_2)|$ quantifies the response at the sum of two stimulus frequencies. Because Fourier spectra have positive and negative frequencies, the sum is actually a difference in the upper-left and lower-right quadrants, since $f_2 + f_1 = f_2 - |f_1|$ for $f_1 < 0$. For linear systems, the second-order response function is zero, because linear systems do not create new frequencies and thus there is no response at the sum of the two frequencies. The plots show the analytical solutions from \citet{Lindner2001} and \citet{Voronenko2017} with $\mu = 1.1$ and $D = 0.001$. Note that the leaky integrate-and-fire model is formulated without dimensions, frequencies are given in multiples of the inverse membrane time constant.}
\end{figure*}
We like to think about signal encoding in terms of linear relations with unique mapping of a given input value to a certain output of the system under consideration. Indeed, such linear methods, for example the transfer function or first-order susceptibility shown in \subfigref{fig:lifsuscept}{A}, have been widely and successfully applied to describe and predict neuronal responses and are an invaluable tools to characterize neural systems \citep{Eggermont1983,Borst1999}. Nonlinear mechanisms, on the other hand, are key on different levels of neural processing. Deciding for one action over another is a nonlinear process on the systemic level. On the cellular level, spiking neurons are inherently nonlinear. Whether an action potential is elicited depends on the membrane potential to exceed a threshold \citep{Hodgkin1952, Koch1995}. Because of such nonlinearities, understanding and predicting neuronal responses to sensory stimuli is in general a difficult task.
@@ -319,7 +319,7 @@ For monitoring the EOD without the stimulus, two vertical carbon rods ($11\,\cen
\subsection{Stimulation}\label{rammethods}
Electric stimuli were attenuated (ATN-01M, npi-electronics, Tamm, Germany), isolated from ground (ISO-02V, npi-electronics, Tamm, Germany) and delivered via two horizontal carbon rods (30 cm length, 8 mm diameter) located $15\,\centi\meter$ parallel to each side of the fish. The fish were stimulated with band-limited white noise stimuli with a cut-off frequency of 150\,Hz (ampullary), 300 or 400\,Hz (P-units). The stimuli were generated by drawing normally distributed real and imaginary numbers for all frequencies up to the desired cut-off frequency in the Fourier domain and then applying an inverse Fourier transform \citep{Skorjanc2023}. The stimulus intensity is given as a contrast, i.e. the standard deviation of the resulting amplitude modulation relative to the fish's EOD amplitude. The contrast varied between 1 and 20\,\% (median 5\,\%) for P-units and 2.5 and 20\,\% (median 5\,\%) for ampullary cells. Only recordings with noise stimuli with a duration of at least 2\,s (maximum of 50\,s, median 10\,s) and enough repetitions to results in at least 100 FFT segments (see below, P-units: 100--1520, median 313, ampullary cells: 105 -- 3648, median 722) were included into the analysis. When ampullary cells were recorded, the noise stimuli $s(t)$ were directly applied as the stimulus and thus were simply added to the fish's own EOD: $s(t) + EOD(t)$. To create random amplitude modulations (RAM) for P-unit recordings, the noise stimulus was first multiplied with the EOD of the fish (MXS-01M; npi electronics) and then played back through the stimulation electrodes: $(1 + s(t))EOD(t)$.
Electric stimuli were attenuated (ATN-01M, npi-electronics, Tamm, Germany), isolated from ground (ISO-02V, npi-electronics, Tamm, Germany) and delivered via two horizontal carbon rods (30 cm length, 8 mm diameter) located $15\,\centi\meter$ parallel to each side of the fish. The fish were stimulated with band-limited white noise stimuli with a cut-off frequency of 150\,Hz (ampullary), 300 or 400\,Hz (P-units). The stimuli were generated by drawing normally distributed real and imaginary numbers for all frequencies up to the desired cut-off frequency in the Fourier domain and then applying an inverse Fourier transform \citep{Skorjanc2023}. The stimulus intensity is given as a contrast, i.e. the standard deviation of the resulting amplitude modulation relative to the fish's EOD amplitude. The contrast varied between 1 and 20\,\% (median 5\,\%) for P-units and 2.5 and 20\,\% (median 5\,\%) for ampullary cells. Only recordings with noise stimuli with a duration of at least 2\,s (maximum of 50\,s, median 10\,s) and enough repetitions to results in at least 100 FFT segments (see below, P-units: 100--1520, median 313, ampullary cells: 105 -- 3648, median 722) were included into the analysis. When ampullary cells were recorded, the noise stimuli $s(t)$ were directly applied as the stimulus and thus were simply added to the fish's own EOD: $s(t) + EOD(t)$. To create random amplitude modulations (RAM) for P-unit recordings, the noise stimulus was first multiplied with the EOD of the fish (MXS-01M; npi electronics) and then played back through the stimulation electrodes: $EOD(t) + s(t)EOD(t) = (1 + s(t))EOD(t)$.
\subsection{Data analysis} Data analysis was done in Python 3 using the packages matplotlib \citep{Hunter2007}, numpy \citep{Walt2011}, scipy \citep{scipy2020}, nixio \citep{Stoewer2014}, and thunderlab (\url{https://github.com/bendalab/thunderlab}).
@@ -338,7 +338,7 @@ To characterize the relation between the spiking response evoked by white-noise
\label{fourier}
X(\omega) = \sum_{k=0}^{N-1} \, x_k e^{- i \omega k}
\end{equation}
for $N=512$ long segments of $T=N \Delta t = 256$\,ms duration with no overlap, resulting in a spectral resolution of about 4\,Hz. In \figref{fig:regimes} we used $N=2^{18}$ and $\Delta t=0.1$\,ms. Note, that for a real Fourier integral a factor $\Delta t$ is missing. For simplicity we use angular frequencies $\omega=2\pi f$ instead of frequencies $f$. Also note, that the Fourier transform of a signal results in both positive as well as negative frequencies. This is because the signal is decomposed into complex exponentials, not sine or cosine waves. To get back a real-valued sine or cosine wave, the components at positive and negative frequencies need to be combined. For example, the sum $e^{-\omega t} + e^{+\omega t}$ of two complex exponentials at frequencies $\pm\omega$ is $2\cos(\omega t)$.
for $N=512$ long segments of $T=N \Delta t = 256$\,ms duration with no overlap, resulting in a spectral resolution of about 4\,Hz. In \figref{fig:regimes} we used $N=2^{18}$ and $\Delta t=0.1$\,ms (26.2\,s). Note, that for a real Fourier integral a factor $\Delta t$ is missing. For simplicity we use angular frequencies $\omega=2\pi f$ instead of frequencies $f$. Also note, that the Fourier transform of a signal results in both positive as well as negative frequencies. This is because the signal is decomposed into complex exponentials, not sine or cosine waves. To get back a real-valued sine or cosine wave, the components at positive and negative frequencies need to be combined. For example, the sum $e^{-\omega t} + e^{+\omega t}$ of two complex exponentials at frequencies $\pm\omega$ is $2\cos(\omega t)$.
In the experimental data the duration of the noise stimuli varied and they were presented once or repeatedly (frozen noise). For the analysis we discarded the responses within the initial 200\,ms of stimulation in each trial. To make the recordings comparable we always used the first 100 segments from as many trials as needed for the following analysis.
@@ -436,7 +436,7 @@ First, the input $y(t)$ is thresholded by setting negative values to zero:
\label{eq:dendrite}
\tau_{d} \frac{d V_{d}}{d t} = -V_{d}+ \lfloor y(t) \rfloor_{0}
\end{equation}
the threshold operation is required for extracting the amplitude modulation from the input \citep{Barayeu2024}. As detailed in the discussion, the details of the threshold nonlinearity \eqnref{eq:threshold2} and the low-pass filter \eqnref{eq:dendrite} do not matter here, since we compute all cross-spectra between the spiking response and the amplitude-modulation $s(t)$ --- and not the full input signal $y(t)$, \eqnref{eq:ram_equation}.
the threshold operation is required for extracting the amplitude modulation from the input \citep{Barayeu2024}. As detailed in the discussion, the intricacies of the threshold nonlinearity \eqnref{eq:threshold2} and the low-pass filter \eqnref{eq:dendrite} do not matter here, as all cross-spectra were computed between the spiking response and the amplitude-modulation $s(t)$ --- and not the full input signal $y(t)$, \eqnref{eq:ram_equation}.
The low-pass filter models passive signal conduction in the afferent's dendrite (\subfigrefb{flowchart}{B}) from the synapse to the spike initiation zone. $\tau_{d}$ is the effective membrane time constant of this part of the dendrite. Dendritic low-pass filtering was also necessary to reproduce the loose coupling of P-unit spikes to the EOD while maintaining high sensitivity at small amplitude modulations.
@@ -498,10 +498,10 @@ When stimulating with both foreign signals simultaneously, additional peaks appe
\begin{figure*}[p]
\includegraphics[width=\columnwidth]{regimes}
\caption{\label{fig:regimes} Linear and nonlinear responses of a model P-unit in a three-fish setting in dependence on stimulus amplitudes. The model P-unit (identifier ``2018-05-08-ad'') was stimulated with two sine waves of equal amplitude (contrast) at difference frequencies $\Delta f_1=40$\,Hz and $\Delta f_2=228$\,Hz relative the receiver's EOD frequency $f_{EOD}=656$\,Hz. $\Delta f_2$ was set to match the baseline firing rate $r$ of the P-unit. \figitem{A--D} Top row: the stimulus, an amplitude modulation of the receiver's EOD resulting from the stimulation with the two sine waves. The contrasts of both beats increase from \panel{A} to \panel{D} as indicated. Second row: Spike raster of the model P-unit response. Third row: average firing rate and, bottom row: power spectrum estimated from 1000 trials of spike trains in response to a 26.2\,s long stimulus. \figitem{A} At low stimulus contrasts the response to the two beat frequencies is linear. These frequencies are present in the response spectrum (green and purple circles), the peak at $\Delta f_2$ (purple) enhances the peak at baseline firing rate (blue). The largest peak, however, is always the one at the EOD frequency of the receiver (black), reflecting the locking of P-unit spikes to the receiver's own EOD. The peak at $f_{EOD}$ is also flanked by harmonics of $\Delta f_1$ (gray triangles), where $f_{EOD} + \Delta f_1 = f_1$ is the EOD frequency of the first fish. \figitem{B} At moderately higher stimulus contrast, the peaks in the response spectrum at the two beat frequencies become larger. In addition, two peaks at $f_{EOD} - \Delta f_2$ and $f_{EOD} + \Delta f_2$ appear (blue diamonds), the latter is the EOD frequency of the second fish, $f_2$. These peaks indicate nonlinear processes acting on the three EOD frequencies from which the beat frequencies are generated. \figitem{C} At intermediate stimulus contrasts, nonlinear responses start to appear at the sum and the difference of the beat frequencies (orange and red circles). Similarly, side peaks appear at $\pm \Delta f_1$ around $f_{EOD} \pm \Delta f_2$ (purple squares) as well as harmonics of $\Delta f_1$ (small green circles). \figitem{D} At higher stimulus contrasts many more peaks appear in the power spectrum. Most of them are interactions of $f_{EOD}$ and $\Delta f_2$ with harmonics of $\Delta f_1$ ($k=1, 2, ...$), since the cell responds strongest to $\Delta f_1$. \figitem{E} Amplitude of the linear (at $\Delta f_1$ and $\Delta f_2$) and nonlinear (at $\Delta f_2 - \Delta f_1$ and $\Delta f_1 + \Delta f_2$) responses of the model P-unit as a function of beat contrast (thick lines). Thin lines indicate the initial linear and quadratic dependence on stimulus amplitude for the linear and nonlinear responses, respectively. In the linear regime, below a stimulus contrast of about 1.2\,\% (left vertical line), the only peaks in the response spectrum are at the stimulus frequencies. In the weakly nonlinear regime up to a contrast of about 3.5\,\% peaks arise at the sum and the difference of the two stimulus frequencies. At stronger stimulation the amplitudes of these nonlinear responses deviate from the quadratic dependency on stimulus contrast.}
\caption{\label{fig:regimes} Linear and nonlinear responses of a model P-unit in a three-fish setting in dependence on stimulus amplitudes. The model P-unit (identifier ``2018-05-08-ad'') was stimulated with two sine waves of equal amplitude (contrast) at difference frequencies $\Delta f_1=40$\,Hz and $\Delta f_2=228$\,Hz relative the receiver's EOD frequency $f_{EOD}=656$\,Hz. $\Delta f_2$ was set to match the baseline firing rate $r$ of the P-unit. \figitem{A--D} Top row: the stimulus, an amplitude modulation of the receiver's EOD resulting from the stimulation with the two sine waves. The contrasts of both beats increase from \panel{A} to \panel{D} as indicated. Second row: Spike raster of the model P-unit response. Third row: average firing rate and, bottom row: power spectrum estimated from 1000 trials of spike train responses. \figitem{A} At low stimulus contrasts the response to the two beat frequencies is linear. These frequencies are present in the response spectrum (green and purple circles), the peak at $\Delta f_2$ (purple) enhances the peak at baseline firing rate (blue). The largest peak, however, is always the one at the EOD frequency of the receiver (black), reflecting the locking of P-unit spikes to the receiver's own EOD. The peak at $f_{EOD}$ is also flanked by harmonics of $\Delta f_1$ (gray triangles), where $f_{EOD} + \Delta f_1 = f_1$ is the EOD frequency of the first fish. \figitem{B} At moderately higher stimulus contrast, the peaks in the response spectrum at the two beat frequencies become larger. In addition, two peaks at $f_{EOD} - \Delta f_2$ and $f_{EOD} + \Delta f_2$ appear (blue diamonds), the latter is the EOD frequency of the second fish, $f_2$. These peaks indicate nonlinear processes acting on the three EOD frequencies from which the beat frequencies are generated. \figitem{C} At intermediate stimulus contrasts, nonlinear responses start to appear at the sum and the difference of the beat frequencies (orange and red circles). Similarly, side peaks appear at $\pm \Delta f_1$ around $f_{EOD} \pm \Delta f_2$ (purple squares) as well as harmonics of $\Delta f_1$ (small green circles). \figitem{D} At higher stimulus contrasts many more peaks appear in the power spectrum. Most of them are interactions of $f_{EOD}$ and $\Delta f_2$ with harmonics of $\Delta f_1$ ($k=1, 2, ...$), since the cell responds strongest to $\Delta f_1$. \figitem{E} Amplitude of the linear (at $\Delta f_1$ and $\Delta f_2$) and nonlinear (at $\Delta f_2 - \Delta f_1$ and $\Delta f_1 + \Delta f_2$) responses of the model P-unit as a function of beat contrast (thick lines). Thin lines indicate the initial linear and quadratic dependence on stimulus amplitude for the linear and nonlinear responses, respectively. In the linear regime, below a stimulus contrast of about 1.2\,\% (left vertical line), the only peaks in the response spectrum are at the stimulus frequencies. In the weakly nonlinear regime up to a contrast of about 3.5\,\% peaks arise at the sum and the difference of the two stimulus frequencies. At stronger stimulation the amplitudes of these nonlinear responses deviate from the quadratic dependency on stimulus contrast.}
\end{figure*}
The stimuli used in \figref{fig:twobeats} had the same not-small amplitude resulting in AM contrasts of 10\,\% that evoke strong modulations in a P-unit's firing rate response. Whether this stimulus condition falls into the weakly nonlinear regime as in \citet{Voronenko2017} is not clear. In order to illustrate how the responses to two beat frequencies develop over a range of amplitudes we use a stochastic leaky-integrate-and-fire (LIF) based P-unit model fitted to a specific electrophysiologically measured cell \citep{Barayeu2023}.
The stimuli used in \figref{fig:twobeats} had the same AM contrasts of 10\,\% that evoke rather strong modulations in a P-unit's firing rate response. Whether these definitely not-small stimulus amplitudes fall into the weakly nonlinear regime as in \citet{Voronenko2017} is not clear. In order to illustrate how the responses to two beat frequencies develop over a range of amplitudes we use a stochastic leaky-integrate-and-fire (LIF) based P-unit model fitted to a specific electrophysiologically measured cell \citep{Barayeu2023}.
At very low stimulus contrasts (in the example cell less than approximately 1.2\,\% relative to the receiver's EOD amplitude) the spectrum has small peaks at the beat frequencies (\subfigref{fig:regimes}{A,B}, green and purple). The amplitudes of these peaks initially increase linearly with stimulus amplitude (\subfigref{fig:regimes}{E}, thin lines), an indication of the linear response at lowest stimulus amplitudes. The largest peak at the receiver's EOD frequency indicates the stochastic locking of the P-unit spikes to the EOD, and its side-peaks are remnants of the two stimulating EOD frequencies \citep{Sinz2020}.
@@ -574,7 +574,7 @@ The SI($r$) correlates with the CVs of the cell's baseline interspike intervals
\subsection{Weakly nonlinear interactions vanish for higher stimulus contrasts}
As pointed out above, strong noise simuli act as a linearizing background noise and thus shift the neural signal transmission into a regime free of nonlinearities. Sine-wave stimuli, as for example used in the introductory figure~\ref{fig:regimes}, do not have such a linearizing effect. What exactly constitutes a 'strong' stimulus causing an effective increase of the background noise, also depends on all the other cellular properties of the neuron, or on the parameters in our leaky integrate-and-fire neuron. This is what we explore now by varying the contrast of RAM stimuli driving our model cells.
As pointed out above, noise stimuli act as a linearizing background noise and thus may shift the neural signal transmission into a regime free of nonlinearities (see discussion). Sine-wave stimuli, as for example used in the introductory figure~\ref{fig:regimes}, do not have such a linearizing effect. At which amplitudes a noise stimulus effectively linearizes the system, also depends on cellular properties of the neuron, or on the parameters in our leaky integrate-and-fire neuron. This is what we explore now by varying the contrast of RAM stimuli driving our model cells.
In the model cells we estimated second-order susceptibilities for RAM stimuli with a contrast of 1, 3, and 10\,\%. The estimates for 1\,\% contrast (\figrefb{fig:modelsusceptcontrasts}\,\panel[ii]{E}) were quite similar to the estimates from the noise-split method, corresponding to a stimulus contrast of 0\,\% ($r=0.97$, $p\ll 0.001$). Thus, RAM stimuli with 1\,\% contrast are sufficiently small to not destroy weakly nonlinear interactions by their linearizing effect. At this low contrast, 51\,\% of the model cells have an SI($r$) value greater than 1.2.