\documentclass[a4paper, 12pt]{article}

\usepackage[left=2cm,right=2cm,top=2cm,bottom=2cm,includeheadfoot]{geometry}
\usepackage[onehalfspacing]{setspace}
\usepackage{graphicx}
\usepackage{svg}
\usepackage{import}
\usepackage{float}
\usepackage{placeins}
\usepackage{parskip}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{subcaption}
\usepackage[labelfont=bf, textfont=small]{caption}
\usepackage[german,english]{babel}
\addto\captionsenglish{\renewcommand{\figurename}{Fig.}}
\addto\captionsenglish{\renewcommand{\tablename}{Tab.}}
\usepackage[separate-uncertainty=true, locale=DE]{siunitx}
\sisetup{output-exponent-marker=\ensuremath{\mathrm{e}}}
% \usepackage[capitalize]{cleveref}
% \crefname{figure}{Fig.}{Figs.}
% \crefname{equation}{Eq.}{Eqs.}
% \creflabelformat{equation}{#2#1#3}
\usepackage[
    backend=biber,
    style=authoryear,
    pluralothers=true,
    maxcitenames=1,
    mincitenames=1
    ]{biblatex}
\addbibresource{cite.bib}
%\bibdata
%\bibstyle
%\citation

\title{Emergent intensity invariance vs. signal-to-noise ratio at three consecutive processing stages along the grasshopper song recognition pathway}
\author{Jona Hartling, Jan Benda}
\date{}

\begin{document}
\maketitle{}

% Text references and citations:
\newcommand{\bcite}[1]{\mbox{\cite{#1}}}
% \newcommand{\fref}[1]{\mbox{\cref{#1}}}
% \newcommand{\fref}[1]{\mbox{Fig.\,\ref{#1}}}
% \newcommand{\eref}[1]{\mbox{\cref{#1}}}
% \newcommand{\eref}[1]{\mbox{Eq.\,\ref{#1}}}

% Subplot lettering:
\newcommand{\figa}{\textbf{a}}
\newcommand{\figb}{\textbf{b}}
\newcommand{\figc}{\textbf{c}}
\newcommand{\figd}{\textbf{d}}
\newcommand{\fige}{\textbf{e}}

% Math shorthands - Standard symbols:
\newcommand{\dec}{\log_{10}} % Logarithm base 10
\newcommand{\infint}{\int_{-\infty}^{+\infty}} % Indefinite integral

% Math shorthands - Spectral filtering:
\newcommand{\bp}{h_{\text{BP}}(t)} % Bandpass filter function
\newcommand{\lp}{h_{\text{LP}}(t)} % Lowpass filter function
\newcommand{\hp}{h_{\text{HP}}(t)} % Highpass filter function
\newcommand{\fc}{f_{\text{cut}}} % Filter cutoff frequency
\newcommand{\tlp}{T_{\text{LP}}} % Lowpass filter averaging interval
\newcommand{\thp}{T_{\text{HP}}} % Highpass filter adaptation interval

% Math shorthands - Early representations:
\newcommand{\raw}{x_{\text{raw}}} % Placeholder input signal
\newcommand{\filt}{x_{\text{filt}}} % Bandpass filtered signal
\newcommand{\env}{x_{\text{env}}} % Signal envelope
\newcommand{\db}{x_{\text{log}}} % Logarithmically scaled signal
\newcommand{\dbref}{x_{\text{ref}}} % Decibel reference intensity
\newcommand{\adapt}{x_{\text{adapt}}} % Adapted signal

% Math shorthands - Kernel parameters:
\newcommand{\kw}{\sigma} % Unspecific Gabor kernel width
\newcommand{\kf}{\omega} % Unspecific Gabor kernel frequency
\newcommand{\kp}{\phi} % Unspecific Gabor kernel phase
\newcommand{\kn}{n} % Unspecific Gabor kernel lobe number
% \newcommand{\ks}{s} % Unspecific Gabor kernel sign
\newcommand{\kwi}{\kw_i} % Specific Gabor kernel width
\newcommand{\kfi}{\kf_i} % Specific Gabor kernel frequency
\newcommand{\kpi}{\kp_i} % Specific Gabor kernel phase
\newcommand{\kni}{\kn_i} % Specific Gabor kernel lobe number
% \newcommand{\ksi}{\ks_i} % Specific Gabor kernel sign

% Math shorthands - Auxiliary kernel parameters:
\newcommand{\fsin}{f_{\text{sin}}} % Carrier frequency
\newcommand{\rh}{h_{\text{rel}}} % Relative Gaussian height for FWRH
\newcommand{\fwrh}{\text{FWRH}} % Gaussian full-width at relative height
\newcommand{\off}{\beta_0} % Offset for linear frequency approximation

% Math shorthands - Thresholding nonlinearity:
\newcommand{\thr}{\Theta_i} % Step function threshold value
\newcommand{\nl}{H(c_i\,-\,\thr)} % Shifted Heaviside step function

% Math shorthands - Intensity invariance analysis:
\newcommand{\soc}{s} % Song component of synthetic mixture
\newcommand{\noc}{\eta} % Noise component of synthetic mixture
\newcommand{\sca}{\alpha} % Multiplicative scale of song component
\newcommand{\xvar}{\sigma_{x}^{2}} % Variance of synthetic mixture
\newcommand{\svar}{\sigma_{\text{s}}^{2}} % Song component variance
\newcommand{\nvar}{\sigma_{\eta}^{2}} % Noise component variance
\newcommand{\pc}{p(c,\,T)} % Probability density (general interval)
\newcommand{\pclp}{p(c,\,\tlp)} % Probability density (lowpass interval)

\section{Exploring a grasshopper's sensory world}

% Why functional models of sensory systems?
Our scientific understanding of sensory processing systems results from the
distributed accumulation of anatomical, physiological and ethological evidence.
This process is undoubtedly without alternative; however, it leaves us with the
challenge of integrating the available fragments into a coherent whole in order
to address issues such as the interaction between individual system components,
the functional limitations of the system overall, or taxonomic comparisons
between systems that process the same sensory modality. Any unified framework
that captures the essential functional aspects of a given sensory system thus
has the potential to deepen our current understanding and fasciliate systematic
investigations. However, building such a framework is a challenging task. It
requires a wealth of existing knowledge of the system and the signals it
operates on, a clearly defined scope, and careful reduction, abstraction, and
formalization of the underlying structures and mechanisms.

% Why the grasshopper auditory system?
% Why focus on song recognition among other auditory functions?
One sensory system about which extensive information has been gathered over the
years is the auditory system of grasshoppers~(\textit{Acrididae}). Grasshoppers
rely on their sense of hearing primarily for intraspecific communication, which
includes mate attraction~(\bcite{helversen1972gesang}) and
evaluation~(\bcite{stange2012grasshopper}), sender
localization~(\bcite{helversen1988interaural}), courtship
display~(\bcite{elsner1968neuromuskularen}), rival
deterrence~(\bcite{greenfield1993acoustic}), and loss-of-signal predator
alarm~(SOURCE). In accordance with this rich behavioral repertoire,
grasshoppers have evolved a variety of sound production mechanisms to generate
acoustic communication signals for different contexts and ranges using their
wings, hindlegs, or mandibles~(\bcite{otte1970comparative}). Among the most
conspicuous acoustic signals of grasshoppers are their species-specific calling
songs, which broadcast the presence of the singing individual --- mostly the
males of the species --- to potential mates within range. These songs are
usually more characteristic of a species than morphological
traits~(\bcite{tishechkin2016acoustic}; \bcite{tarasova2021eurasius}), which
can vary greatly within species~(\bcite{rowell1972variable};
\bcite{kohler2017morphological}). The reliance on songs to mediate reproduction
represents a strong evolutionary driving force, that resulted in a massive
species diversification~(\bcite{vedenina2011speciation};
\bcite{sevastianov2023evolution}), with over 6800 recognized grasshopper
species in the \textit{Acrididae} family~(\bcite{cigliano2024orthoptera}). It
is this diversity of species, and the crucial role of acoustic communication in
its emergence, that makes the grasshopper auditory system an intriguing
candidate for attempting to construct a functional model framework. As a
necessary reduction, the model we propose here focuses on the pathway
responsible for the recognition of species-specific calling songs, disregarding
other essential auditory functions such as directional
hearing~(\bcite{helversen1984parallel}; \bcite{ronacher1986routes};
\bcite{helversen1988interaural}).

% What are the signals the auditory system is supposed to recognize?
% Why is intensity invariance important for song recognition?
% (Obviously, split this paragraph)
To understand the functional challenges faced by the grasshopper auditory
system, one has to understand the properties of the songs it is designed to
recognize. Grasshopper songs are amplitude-modulated broad-band acoustic
signals. Most songs are produced by stridulation, during which the animal pulls
the serrated stridulatory file on its hindlegs across a resonating vein on the
forewings~(\bcite{helversen1977stridulatory}; \bcite{stumpner1994song};
\bcite{helversen1997recognition}). Every tooth that strikes the vein generates
a brief pulse of sound. Multiple pulses make up a syllable; and the alternation
of syllables and relatively quiet pauses forms a characteristic, through noisy,
waveform pattern. Song recognition depends on certain temporal and structural
parameters of this pattern, such as the duration of syllables and
pauses~(\bcite{helversen1972gesang}), the slope of pulse
onsets~(\bcite{helversen1993absolute}), and the accentuation of syllable onsets
relative to the preceeding pause~(\bcite{balakrishnan2001song};
\bcite{helversen2004acoustic}). The amplitude modulation of the song is
sufficient for recognition~(\bcite{helversen1997recognition}). However, the
essential recognition cues can vary considerably with external physical
factors, which requires the auditory system to be invariant to such variations
in order to reliably recognize songs under different conditions. For instance,
the temporal structure of grasshopper songs warps with
temperature~(\bcite{skovmand1983song}). The auditory system can compensate for
this variability by reading out relative temporal relationships rather than
absolute time intervals~(\bcite{creutzig2009timescale};
\bcite{creutzig2010timescale}), as those remain relatively constant across
different temperatures~(\bcite{helversen1972gesang}). Another, perhaps even
more fundamental external source of song variability lays in the attenuation of
sound intensity with increasing distance to the sender. Sound attenuation
depends on both the frequency content of the signal and the vegetation of the
habitat~(\bcite{michelsen1978sound}). For the receiving auditory system, this
has two major implications. First, the amplitude dynamics of the song pattern
are steadily degraded over distance, which limits the effective communication
range of grasshoppers to~\mbox{1\,-\,2\,m} in their typical grassland
habitats~(\bcite{lang2000acoustic}). Second, the overall intensity level of
songs at the receiver's position varies depending on the location of the
sender, which should ideally not affect the recognition of the song pattern.
This neccessitates that the auditory system achieves a certain degree of
intensity invariance --- a time scale-selective sensitivity to faster amplitude
dynamics and simultaneous insensitivity to slower, more sustained amplitude
dynamics. Intensity invariance in different auditory systems is often
associated with neuronal adaptation~(\bcite{benda2008spike};
\bcite{barbour2011intensity}; \bcite{ozeri2018fast}; more
general:~\bcite{benda2021neural}). In the grasshopper auditory system, a number
of neuron types along the processing chain exhibit spike-frequency adaptation
in response to sustained stimulus
intensities~(\bcite{romer1976informationsverarbeitung};
\bcite{gollisch2004input}; \bcite{hildebrandt2009origin};
\bcite{clemens2010intensity}; \bcite{fisch2012channel}) and thus likely
contribute to the emergence of intensity-invariant song representations. This
means that intensity invariance is not the result of a single processing step
but rather a gradual process, in which different neuronal populations
contribute to varying degrees~(\bcite{clemens2010intensity}) and by different
mechanisms~(\bcite{hildebrandt2009origin}). Approximating this process within a
functional model framework thus requires a considerable amount of
simplification. In this work, we demonstrate that even a small number of basic
physiologically inspired signal transformations --- specifically, pairs of
nonlinear and linear operations --- is sufficient to achieve a meaningful
degree of intensity invariance.

% How can song recognition be modelled functionally (feat. Jan Clemens & Co.)?
% How did we expand on the previous framework?
% (Still can't stand some of this paragraph's structure and wording...)
Invariance to non-informative song variations is crucial for reliable song
recognition; however, it is not sufficient to this end. In order to recognize a
conspecific song as such, the auditory system needs to extract sufficiently
informative features of the song pattern and then integrate the gathered
information into a final categorical percept. Previous authors have proposed a
functional model framework that describes this process --- feature extraction,
evidence accumulation, and categorical decision making --- in both
crickets~(\bcite{clemens2013computational}; \bcite{hennig2014time}) and
grasshoppers~(\bcite{clemens2013feature}; review on
both:~\bcite{ronacher2015computational}). Their framework provides a
comprehensible and biologically plausible account of the computational
mechanisms required for species-specific song recognition, which has served as
the inspiration for the development of the model pathway we propose here. The
existing framework relies on pulse trains as input signals, which were designed
to capture the essential structural properties of natural song
envelopes~(\bcite{clemens2013feature}). In the first step, a bank of parallel
linear-nonlinear feature detectors is applied to the input signal. Each feature
detector consists of a convolutional filter and a subsequent sigmoidal
nonlinearity. The outputs of these feature detectors are temporally averaged to
obtain a single feature value per detector, which is then assigned a specific
weight. The linear combination of weighted feature values results in a single
preference value, that serves as predictor for the behavioral response of the
animal to the presented input signal. Our model pathway adopts the general
structure of the existing framework but modifies it in several key aspects. The
convolutional filters, which have previously been fitted to behavioral data for
each individual species~(\bcite{clemens2013computational}), are replaced by a
larger, generic set of unfitted Gabor basis functions in order to cover a wide
range of possible song features across different species. Gabor functions
approximate the general structure of the filters used in the existing framework
as well as the filter functions found in various auditory
neurons~(\bcite{rokem2006spike}; \bcite{clemens2011efficient};
\bcite{clemens2012nonlinear}). The fitted sigmoidal nonlinearities in the
existing framework consistently exhibited very steep slopes and are therefore
replaced by shifted Heaviside step-functions, which results in a binarization
of the feature detector outputs. Another, more substantial modification is that
the feature detector outputs are temporally averaged in a way that does not
condense them into single feature values but retains their time-varying
structure. This is in line with the fact that songs are no discrete units but
part of a continuous acoustic stream that the auditory system has to process in
real time. Moreover, a time-varying feature representation only stabilizes
after a certain delay following the onset of a song, which emphasizes the
temporal dynamics of evidence accumulation towards a final categorical
decision. The most notable difference between our model pathway and the
existing framework, however, lays in the addition of a physiologically inspired
preprocessing stage, whose starting point corresponds to the initial reception
of airborne sound waves. This allows the model to operate on unmodified
recordings of natural grasshopper songs instead of condensed pulse train
approximations, which widens its scope towards more realistic, ecologically
relevant scenarios. For instance, we were able to investigate the contribution
of different processing stages to the emergence of intensity-invariant song
representations based on actual field recordings of songs at different
distances from the sender.
% Forgive me, it's friday.
In the following, we outline the structure of the proposed model of the
grasshopper auditory pathway, from the initial reception of sound waves up to
the generation of a high-dimensional, time-varying feature representation that
is suitable for species-specific song recognition. We provide a side-by-side
account of the known physiological processing steps and their functional
approximation by basic mathematical operations. We then elaborate on two key
mechanisms that drive the emergence of intensity-invariant song representations
within the auditory pathway.

% SCRAPPED UNTIL FURTHER NOTICE:
% Multi-species, multi-individual communally inhabited environments\\
% - Temporal overlap: Simultaneous singing across individuals/species common\\
% - Frequency overlap: Little speciation into frequency bands (likely unused)\\
% - "Biotic noise": Hetero-/conspecifics ("Another one's songs are my noise")\\
% - "Abiotic noise": Wind, water, vegetation, anthropogenic\\
% - Effects of habitat structure on sound propagation (landscape - soundscape)\\
% $\rightarrow$ Sensory constraints imposed by the (acoustic) environment

% Cluster of auditory challenges (interlocking constraints $\rightarrow$ tight coupling):\\
% From continuous acoustic input, generate neuronal representations that...\\
% 1)...allow for the separation of relevant (song) events from ambient noise floor\\
% 2)...compensate for behaviorally non-informative song variability (invariances)\\
% 3)...carry sufficient information to characterize different song patterns,
% recognize the ones produced by conspecifics, and make appropriate behavioral
% decisions based on context (sender identity, song type, mate/rival quality)

% How can the auditory system of grasshoppers meet these challenges?\\
% - What are the minimum functional processing steps required?\\
% - Which known neuronal mechanisms can implement these steps?\\
% - Which and how many stages along the auditory pathway contribute?\\
% $\rightarrow$ What are the limitations of the system as a whole?

% How can a human observer conceive a grasshopper's auditory percepts?\\
% - How to investigate the workings of the auditory pathway as a whole?\\
% - How to systematically test effects and interactions of processing parameters?\\
% - How to integrate the available knowledge on anatomy, physiology, ethology?\\
% $\rightarrow$ Abstract, simplify, formalize $\rightarrow$ Functional model framework

\section{Developing a functional model of the\\grasshopper song recognition pathway}

% Too long (no splitting, only pruning).
The essence of constructing a functional model of a given system is to gain a
sufficient understanding of the system's essential structural components and
their presumed functional roles; and to then build a formal framework of
manageable complexity around these two aspects. Anatomically, the organization
of the grasshopper song recognition pathway can be outlined as a feed-forward
network of three consecutive neuronal
populations~(Fig.\,\mbox{\ref{fig:pathway}a-c}): Peripheral auditory receptor
neurons, whose axons enter the ventral nerve cord at the level of the
metathoracic ganglion; local interneurons that remain exclusively within the
thoracic region of the ventral nerve cord; and ascending neurons projecting
from the thoracic region towards the supraesophageal
ganglion~(\bcite{rehbein1974structure}; \bcite{rehbein1976auditory};
\bcite{eichendorf1980projections}). The input to the network originates at the
tympanal membrane, which acts as acoustic receiver and is coupled to the
dendritic endings of the receptor neurons~(\bcite{gray1960fine}). The outputs
from the network converge in the supraesophageal ganglion, which is presumed to
harbor the neuronal substrate for conspecific song recognition and response
initiation~(\bcite{ronacher1986routes}; \bcite{bauer1987separate};
\bcite{bhavsar2017brain}). Functionally, the ascending neurons are the most
diverse of the three populations along the pathway. Individual ascending
neurons possess highly specific response properties that contrast with the
rather homogeneous response properties of the preceding receptor neurons and
local interneurons~(\bcite{clemens2011efficient}), indicating a transition from
a uniform population-wide processing stream into several parallel branches.
Based on these anatomical and physiological considerations, the overall
structure of the model pathway is divided into two distinct
stages~(Fig.\,\ref{fig:pathway}d). The preprocessing stage incorporates the
known physiological processing steps at the levels of the tympanal membrane,
the receptor neurons, and the local interneurons; and operates on
one-dimensional signal representations. The feature extraction stage
corresponds to the processing within the ascending neurons and further
downstream towards the supraesophageal ganglion; and operates on
high-dimensional signal representations. The details of each physiological
processing step and its functional approximation within the two stages are
outlined in the following sections.

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_auditory_pathway.pdf}
    \caption{\textbf{Schematic organisation of the grasshopper song recognition
                     pathway and structure of the functional model pathway.}
                     \textbf{a}:~Simplified course of the pathway in the
                     grasshopper, from the tympanal membrane over receptor
                     neurons, local interneurons, and ascending neurons further
                     towards the supraesophageal ganglion.
                     \textbf{b}:~Schematic of synaptic connections between
                     the three neuronal populations within the metathoracic
                     ganglion.
                     \textbf{c}:~Network representation of neuronal connectivity.
                     \textbf{d}:~Flow diagram of consecutive signal
                     representations~(boxes) and transformations~(arrows) along
                     the model pathway. All representations are time-varying.
                     1st half: Preprocessing stage~(one-dimensional
                     representation). 2nd half: Feature extraction
                     stage~(high-dimensional representation). }
    \label{fig:pathway}
\end{figure}

\subsection{Population-driven signal preprocessing}

Grasshoppers receive airborne sound waves by a tympanal organ at either side of
the body. The tympanal membrane acts as a mechanical resonance filter for
sound-induced vibrations~(\bcite{windmill2008time}; \bcite{malkin2014energy}).
Vibrations that fall within specific frequency bands are focused on different
membrane areas, while others are attenuated. This processing step can be
approximated by an initial bandpass filter
\begin{equation}
    \filt(t)\,=\,\raw(t)\,*\,\bp, \qquad \fc\,=\,5\,\text{kHz},\,30\,\text{kHz}
    \label{eq:bandpass}
\end{equation}
applied to the acoustic input signal $\raw(t)$. The auditory receptor neurons
transduce the vibrations of the tympanal membrane into sequences of action
potentials. Thereby, they encode the amplitude modulation, or envelope, of the
signal~(\bcite{machens2001discrimination}), which likely involves a rectifying
nonlinearity~(\bcite{machens2001representation}). This can be modelled as
full-wave rectification followed by lowpass filtering
\begin{equation}
    \env(t)\,=\,|\filt(t)|\,*\,\lp, \qquad \fc\,=\,250\,\text{Hz}
    \label{eq:env}
\end{equation}
of the tympanal signal $\filt(t)$. Furthermore, the receptors exhibit a
sigmoidal response curve over logarithmically compressed intensity
levels~(\bcite{suga1960peripheral}; \bcite{gollisch2002energy}). In the model
pathway, logarithmic compression is achieved by conversion to decibel scale
\begin{equation}
    \db(t)\,=\,20\,\cdot\,\dec \frac{\env(t)}{\dbref}, \qquad \dbref\,=\,1
    \label{eq:log}
\end{equation}
relative to the common reference intensity $\dbref$.
Both the receptor neurons~(\bcite{romer1976informationsverarbeitung};
\bcite{gollisch2004input}; \bcite{fisch2012channel}) and, on a larger scale,
the subsequent local interneurons~(\bcite{hildebrandt2009origin};
\bcite{clemens2010intensity}) adapt their firing rates in response to sustained
stimulus intensity levels, which allows for the robust encoding of faster
amplitude modulations against a slowly changing overall baseline intensity.
Functionally, the adaptation mechanism resembles a highpass filter
\begin{equation}
    \adapt(t)\,=\,\db(t)\,*\,\hp, \qquad \fc\,=\,10\,\text{Hz}
    \label{eq:highpass}
\end{equation}
over the logarithmically scaled envelope $\db(t)$. This processing step
concludes the preprocessing stage of the model pathway. The resulting
intensity-adapted envelope $\adapt(t)$ is then passed on from the local
interneurons to the ascending neurons, where it serves as the basis for the
following feature extraction stage.

% Cite somewhere:
\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_pre_stages.pdf}
    \caption{\textbf{Representations of a song of \textit{O. rufipes} during
                     the preprocessing stage.}
                     \textbf{a}:~Bandpass filtered tympanal signal $\filt(t)$.
                     \textbf{b}:~Signal envelope $\env(t)$.
                     \textbf{c}:~Logarithmically compressed envelope $\db(t)$.
                     \textbf{d}:~Intensity-adapted envelope $\adapt(t)$.
                     }
    \label{fig:stages_pre}
\end{figure}
\FloatBarrier

\subsection{Feature extraction by individual neurons}

The ascending neurons extract and encode a number of different features of the
preprocessed signal. As a population, they hence represent the signal in a
higher-dimensional space than the preceding receptor neurons and local
interneurons. Each ascending neuron is assumed to scan the signal for a
specific template pattern, which can be thought of as a kernel of a particular
structure and on a particular time scale. This process, known as template
matching, can be modelled as a convolution
\begin{equation}
    c_i(t)\,=\,\adapt(t)\,*\,k_i(t)
    = \infint \adapt(\tau)\,\cdot\,k_i(t\,-\,\tau)\,d\tau
    \label{eq:conv}
\end{equation}
of the intensity-adapted envelope $\adapt(t)$ with a kernel $k_i(t)$ per
ascending neuron. We use Gabor kernels as basis functions for creating
different template patterns. An arbitrary one-dimensional, real Gabor kernel is
generated by multiplication of a Gaussian envelope and a sinusoidal carrier
\begin{equation}
    k_i(t,\,\kwi,\,\kfi,\,\kpi)\,=\,e^{-\frac{t^{2}}{2{\kwi}^{2}}}\,\cdot\,\sin(\kfi\,t\,+\,\kpi), \qquad \kfi\,=\,2\pi\fsin
    \label{eq:gabor}
\end{equation}
with Gaussian standard deviation or kernel width $\kwi$, carrier frequency
$\kfi$, and carrier phase $\kpi$. Different combinations of $\kw$ and $\kf$
result in Gabor kernels with different lobe number $\kn$, which is the number
of half-periods of the carrier that fit under the Gaussian envelope within
reasonable limits of attenuation. The interval under the Gaussian envelope that
contains the relevant lobes of the kernel can be defined as Gaussian full-width
measured at relative peak height $\rh$
\begin{equation}
    \fwrh(\kw,\,\rh)\,=\,2\,\cdot\,\sqrt{-2\,\cdot\,\ln \rh}\cdot\,\kw, \qquad \rh\,\in\,(0,\,1]
\end{equation}
With this, an appropriate carrier frequency $\kf$ for obtaining a Gabor kernel
with width $\kw$ and desired lobe number $\kn$ can be approximated as
% \begin{equation}
%     \kf(\kn,\,\fwrh)\,=\,\frac{0.5\,\cdot\,\kn\,+\,\off}{\fwrh}, \qquad \kn\,\geq\,2\enspace\forall\enspace \kn\,\in\,\mathbb{Z}
% \end{equation}
\begin{equation}
    \kf(\kn,\,\kw,\,\rh)\,=\,\frac{\kn\,+\,\off}{4\,\cdot\,\sqrt{-2\,\cdot\,\ln \rh}}, \qquad \kn\,\geq\,2\enspace\forall\enspace \kn\,\in\,\mathbb{Z}
\end{equation}
where $\off$ is a small positive offset to the near-linear relationship between
$\kf$ and $\kn$ to balance the amplitude of the $\kn$ desired lobes of the
kernel --- which should be maximized --- against the amplitude of the
next-outer lobes, which should not exceed the threshold value determined by
$\rh$. For $\kn=1$, carrier frequency $\kf$ is set to zero, which results in a
simple Gaussian kernel. Carrier phase $\kp$ determines the position of the
kernel lobes relative to the kernel center. By setting $\kp$ to one of only
four specific phase values~(Tab.\,\ref{tab:gabor_phases}), we restrict the
Gabor kernels to be either even functions~(mirror-symmetric, uneven $\kn$) or
odd functions~(point-symmetric, even $\kn$) with either positive or negative
sign, which refers to the sign of the kernel's central lobe (even kernels) or
the left of the two central lobes (odd kernels).
\FloatBarrier
\begin{table}[!ht]
    \centering
    \captionsetup{width=.46\textwidth}
    \caption{Values of phase $\kp$ that are specific for the four major groups
             of Gabor kernels.}
    \begin{tabular}{|ccc|}
        \hline
        sign & even kernels & odd kernels\\
        \hline
        $+$ & $+\pi\,/\,2$ & $\pi$\\
        $-$ & $-\pi\,/\,2$ & $0$\\
        \hline
    \end{tabular}
    \label{tab:gabor_phases}
\end{table}
\FloatBarrier
These four major groups of Gabor kernels allow for the extraction of different
types of signal features, such as the presence of peaks (even, $+$), troughs
(even, $-$), onsets (odd, $+$), and offsets (odd, $-$) at various time scales.
% Add kernel normalization here.
Following the convolutional template matching, each kernel-specific response
$c_i(t)$ is passed through a shifted Heaviside step-function $\nl$ with
threshold value $\thr$ to obtain a binary response
\begin{equation}
    b_i(t,\,\thr)\,=\,\begin{cases}
        \;1, \quad c_i(t)\,>\,\thr\\
        \;0, \quad c_i(t)\,\leq\,\thr
    \end{cases}
    \label{eq:binary}
\end{equation}
which can be thought of as a categorization into "relevant" and "irrelevant"
response values. In the grasshopper, these thresholding nonlinearities might
either be part of the processing within the ascending neurons or take place
further downstream~(SOURCE). Finally, the responses of the ascending neurons
are assumed to be integrated somewhere in the supraesophageal
ganglion~(\bcite{ronacher1986routes}; \bcite{bauer1987separate};
\bcite{bhavsar2017brain}). This processing step can be approximated as temporal
averaging of the binary responses $b_i(t)$ by a lowpass filter
\begin{equation}
    f_i(t)\,=\,b_i(t)\,*\,\lp, \qquad \fc\,=\,1\,\text{Hz}
    \label{eq:lowpass}
\end{equation}
to obtain a final set of slowly changing kernel-specific features $f_i(t)$. In
the resulting high-dimensional feature space, different species-specific song
patterns are characterized by a distinct combination of feature values, which
can be read out by a simple linear classifier.

% Cite somewhere:
\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_feat_stages.pdf}
    \caption{\textbf{Representations of a song of \textit{O. rufipes} during
                     the feature extraction stage.}
                     Different color shades indicate different types of Gabor
                     kernels with specific lobe number $\kn$ and either $+$ or
                     $-$ sign, sorted (dark to light) first by increasing $\kn$
                     and then by sign~($1\,\leq\,\kn\,\leq\,4$; first $+$, then
                     $-$ for each $\kn$; two kernel widths $\kw$ of $4\,$ms and
                     $32\,$ms per type; 8 types, 16 kernels in total).
                     \textbf{a}:~Kernel-specific filter responses $c_i(t)$.
                     \textbf{b}:~Binary responses $b_i(t)$.
                     \textbf{c}:~Finalized features $f_i(t)$.}
    \label{fig:stages_feat}
\end{figure}
\FloatBarrier

\section{Mechanisms driving the emergence of\\intensity-invariant song representation}

% Still missing the SNR analysis. Should be able to write around it for now.
The robustness of song recognition is tied to the degree of intensity
invariance of the finalized feature representation. Ideally, the values of each
feature should depend only on the relative amplitude dynamics of the song
pattern but not on the overall intensity of the song. In the grasshopper, the
emergence of intensity-invariant representations along the song recognition
pathway likely is a distributed process that involves different neuronal
populations, which raises the question of what the essential computational
mechanisms are that drive this process. Within the model pathway, we identified
two key mechanisms that render the song representation more invariant to
intensity variations. The two mechanisms each comprise a nonlinear signal
transformation followed by a linear signal transformation but differ in the
specific operations involved, as outlined in the following sections.

\subsection{Full-wave rectification \& lowpass filtering}

The first nonlinear transformation along the model pathway is the full-wave
rectification of the tympanal signal $\filt(t)$ during the extraction of the
signal envelope (Eq.\,\ref{eq:env}). Rectification transforms the distribution
of $\filt(t)$ from an approximately zero-centered distribution with both
positive and negative values into a strictly non-negative distribution. Signal
envelope $\env(t)$ is then obtained by lowpass filtering the rectified
$\filt(t)$. The effects of this transformation pair on SNR and potential
intensity invariance were analyzed by rescaling and processing the input signal
$\raw(t)$ and comparing standard deviations between the resulting $\filt(t)$
and $\env(t)$, once for the noiseless case~(Fig.\,\ref{fig:rect-lp}a) and once
for the noisy case~(Fig.\,\ref{fig:rect-lp}b). In addition, the cutoff
frequency $\fc$ of the lowpass filter was varied to investigate the influence
of different filter bandwidths. In the noiseless case, the standard deviations
of $\filt(t)$ and $\env(t)$ are each reduced compared to the input $\raw(t)$ by
a multiplicative factor. These factors are constant across all $\sca$, which
results in a downward shift of the respective curve on a double-logarithmic
scale, away from the diagonal~(Fig.\,\ref{fig:rect-lp}c). For $\filt(t)$, the
reduction is a consequence of the bandpass filtering~(Eq.\,\ref{eq:bandpass})
of $\raw(t)$. For $\env(t)$, the standard deviation is further reduced compared
to $\filt(t)$. Rectification contributes much less to this reduction than
lowpass filtering. The degree of reduction by lowpass filtering depends on the
cutoff frequency $\fc$, with lower $\fc$ (narrow bandwidth) resulting in a
stronger reduction. In the noisy case, the standard deviations of $\filt(t)$
and $\env(t)$ can be related to the respective pure-noise reference standard
deviation~(Fig.\,\ref{fig:rect-lp}d). This causes each curve to start with a
constant regime of SNR values near 1 for smaller $\sca$, which reflects the
dominance of the noise component $\noc(t)$ over the song component $\soc(t)$ in
the input $\raw(t)$. For larger $\sca$, all curves transition into a regime of
linearly increasing SNR on a double-logarithmic scale. For $\filt(t)$, the
linear part of the curve deviates only slightly from the diagonal. For
$\env(t)$, however, the transition occurs at lower $\sca$ compared to
$\filt(t)$, and the linear part of the curve is shifted leftward away from the
diagonal, which means that higher SNR values are achieved for the same $\sca$.
This effect is more pronounced for lower $\fc$ of the lowpass filter and is
presumably caused by the attenuation of high-frequency components in the
signal, which are more prominent in the noise component $\noc(t)$ than in the
song component $\soc(t)$. The effect also appears relatively consistent across
different species, although small variations exist~(Fig.\,\ref{fig:rect-lp}e)
that are presumably based on different song structures and frequency spectra.
In summary, the standard deviation of $\env(t)$ has never been observed to
transition into a saturation regime for larger $\sca$ but rather continues to
increase proportionally to $\sca$ for all tested $\fc$, in both the noiseless
and the noisy case and across different species. Consequently, the combination
of rectification and lowpass filtering does not contribute to intensity
invariance. However, this transformation pair does improve the SNR of $\env(t)$
relative to $\filt(t)$ and thus provides subsequent processing stages with a
more robust input representation and higher input SNR.

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_rect_lp.pdf}
\caption{\textbf{Rectification and lowpass filtering improves SNR
                 but does not contribute to intensity invariance.}
                 Input $\raw(t)$ consists of song component $\soc(t)$ scaled by
                 $\sca$ with optional noise component $\noc(t)$ and is
                 successively transformed into tympanal signal $\filt(t)$ and
                 envelope $\env(t)$. Different line styles indicate different
                 cutoff frequencies $\fc$ of the lowpass filter extracting
                 $\env(t)$.
                 \textbf{Top}:~Example representations of $\filt(t)$ and
                 $\env(t)$ for different $\sca$.
                 \textbf{a}:~Noiseless case.
                 \textbf{b}:~Noisy case.
                 \textbf{Bottom}:~Intensity metrics over a range of $\sca$.
                 \textbf{c}:~Noiseless case: Standard deviations $\sigma_x$ of
                 $\filt(t)$ and $\env(t)$.
                 \textbf{d}:~Noisy case: Ratios of $\sigma_x$ of $\filt(t)$ and
                 $\env(t)$ to the respective reference standard deviation
                 $\sigma_{\eta}$ for input $\raw(t)=\noc(t)$.
                 \textbf{e}:~Ratios of $\sigma_x$ to $\sigma_{\eta}$ of
                 $\env(t)$ as in \textbf{d} for different species (averaged
                 over songs and recordings, see appendix
                 Fig.\,\ref{fig:app_rect-lp}).
                }
    \label{fig:rect-lp}
\end{figure}
\FloatBarrier

\subsection{Logarithmic compression \& spike-frequency adaptation}

The second nonlinear transformation along the model pathway is the logarithmic
compression of the signal envelope $\env(t)$ into $\db(t)$, Eq.\,\ref{eq:log},
which is then followed by the highpass filtering of $\db(t)$,
Eq.\,\ref{eq:highpass}, to obtain the intensity-adapted envelope $\adapt(t)$.
The interplay of this transformation pair was analyzed by rescaling and
processing the input signal $\filt(t)$ and comparing standard deviations
between the resulting $\env(t)$, $\db(t)$, and $\adapt(t)$. It is necessary to
use $\filt(t)$ as input for this analysis instead of $\env(t)$, because
$\env(t)$ results from a nonlinear transformation and hence cannot be
synthesized as an additive mixture of song component $\soc(t)$ and noise
component $\noc(t)$. % <-- Sentence may be methods section material.
However, it is much easier to conceive a mathematical description of the
effects of logarithmic compression and adaptation if $\env(t)$ itself is
assumed to be composed of $\soc(t)$ and $\noc(t)$. In the noiseless
case~(Fig.\,\ref{fig:log-hp}a), $\env(t)$ takes the form of
\begin{equation}
    \env(t)\,=\,\sca\,\cdot\,\soc(t), \qquad \env(t)\,>\,0\enspace\forall\enspace t\,\in\,\mathbb{R}
    \label{eq:toy_env_pure}
\end{equation}
The standard deviation of $\env(t)$ increases linearly with $\sca$ on a
double-logarithmic scale and is slightly reduced~(Fig.\,\ref{fig:log-hp}c)
compared to the input $\filt(t)$, which is consistent with the results of the
previous analysis~(Fig.\,\ref{fig:rect-lp}c). By conversion of $\env(t)$ to
decibel scale, $\sca$ turns from a multiplicative scale in linear space into an
additive term, or offset, in logarithmic space:
\begin{equation}
    \db(t)\,=\,20\,\cdot\,\dec \left[\,\sca\,\cdot\,s(t)\,\right]\,=\,20\,\cdot\,\left[\dec \sca\,+\,\dec s(t)\right], \qquad \sca\,>\,0
    \label{eq:toy_log_pure}
\end{equation}
The highpass filtering of $\db(t)$ can be approximated as a subtraction of the
local signal offset within a suitable time interval $0 \ll \thp <
\frac{1}{\fc}$:
\begin{equation}
    \begin{split}
    \adapt(t)\,\approx\,\db(t)\,-\,20\,\cdot\,\dec \sca\,=\,20\,\cdot\,\dec s(t)
    \end{split}
    \label{eq:toy_highpass_pure}
\end{equation}
This eliminates $\sca$ from $\adapt(t)$ and thus renders it perfectly
intensity-invariant, with a constant standard deviation of around 10\,dB across
all $\sca>0$~(Fig.\,\ref{fig:log-hp}c). In contrast, in the noisy
case~(Fig.\,\ref{fig:log-hp}b), $\env(t)$ takes the form of
\begin{equation}
    \env(t)\,=\,\sca\,\cdot\,\soc(t)\,+\,\noc(t), \qquad \env(t)\,>\,0\enspace\forall\enspace t\,\in\,\mathbb{R}
    \label{eq:toy_env_noise}
\end{equation}
Similar to the previous analysis~(Fig.\,\ref{fig:rect-lp}d), the ratio of the
standard deviation of $\env(t)$ to its pure-noise reference standard deviation
on a double-logarithmic scale follows a constant regime for small $\sca$ and a
linearly increasing regime for larger $\sca$~(Fig.\,\ref{fig:log-hp}d). Decibel
conversion of $\env(t)$
% \begin{equation}
%     \begin{split}
%         \db(t)\,&=\,20\,\cdot\,\dec \left[\,\sca\,\cdot\,s(t)\,+\,\eta(t)\,\right]\\
%         &=\,20\,\cdot\,\left(\dec \sca\,+\,\dec \left[s(t)\,+\,\frac{\eta(t)}{\sca}\right]\right), \qquad \sca\,>\,0
%     \end{split}
%     \label{eq:toy_log_noise}
% \end{equation}
\begin{equation}
    \db(t)\,=\,20\,\cdot\,\left(\dec \sca\,+\,\dec \left[s(t)\,+\,\frac{\eta(t)}{\sca}\right]\right), \qquad \sca\,>\,0
    \label{eq:toy_log_noise}
\end{equation}
allows for the separation of $\sca$ from $\soc(t)$ but introduces a scaling of
$\noc(t)$ by the inverse of $\sca$, which remains present even after the offset
subtraction:
\begin{equation}
    \begin{split}
    \adapt(t)\,\approx\,20\,\cdot\,\dec\left[s(t)\,+\,\frac{\eta(t)}{\sca}\right]
    \end{split}
    \label{eq:toy_highpass_noise}
\end{equation}
% \begin{equation}
%     \begin{split}
%     \adapt(t)\,\approx\,\db(t)\,-\,20\,\cdot\,\dec \sca\,=\,20\,\cdot\,\dec\left[s(t)\,+\,\frac{\eta(t)}{\sca}\right]
%     \end{split}
%     \label{eq:toy_highpass_noise}
% \end{equation}
This means that, in the noisy case, $\sca$ cannot be entirely eliminated from
$\adapt(t)$, only redistributed between $\soc(t)$ and $\noc(t)$. If $\sca$ is
sufficiently large ($\sca\gg1$, saturation regime), $\noc(t)$ is attenuated to
the point of being negligible, so that $\adapt(t)$ is a scale-free
representation of $\soc(t)$. If $\sca$ and $\noc(t)$ are at similar scales
($\sca\approx1$, transient regime), $\adapt(t)$ largely resembles $\db(t)$.
Finally, if $\sca$ is sufficiently small ($0<\sca\ll1$, noise regime),
$\noc(t)$ masks $\soc(t)$ even after the intensity adaptation. Accordingly, the
effective intensity invariance of $\adapt(t)$ through logarithmic compression
and adaptation is limited by the SNR of $\env(t)$: Songs that have already
sunken into the noise floor at the level of $\env(t)$ cannot be recovered by
subsequent processing steps, which emphasizes the importance of the SNR
improvement by rectification and lowpass filtering during the previous
processing step~(Fig.\,\ref{fig:rect-lp}d). The general pattern of noise
regime, transient regime, and saturation regime remains consistent across
different species~(Fig.\,\ref{fig:log-hp}e). However, the specific value of
$\sca$ at which the saturation regime is reached (see appendix
Fig.\,\ref{fig:app_log-hp_saturation}) and the maximum SNR value of $\adapt(t)$
within the saturation regime vary considerably between and within species. For
example, \textit{C. biguttulus} and \textit{C. mollis} display a noticably
lower maximum SNR of $\adapt(t)$ compared to other species. These differences
are not to be underestimated, since the SNR of $\adapt(t)$ within the
saturation regime determines the maximum input SNR for subsequent processing
steps. In other words, the fact that $\adapt(t)$ eventually reaches a
saturation regime is, of course, desirable in the context of intensity
invariance, but it also means to pass up on the higher SNR values that are
achieved by $\env(t)$ for the same $\sca$ (up to several orders of magnitude,
Fig.\,\ref{fig:log-hp}d). This trade-off between intensity invariance and SNR
is a recurring phenomenon that is further addressed in the following sections.

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_log_hp.pdf}
    \caption{\textbf{Intensity invariance through logarithmic compression and
                     adaptation is restricted by the noise floor and decreases
                     SNR.}
                     Input $\filt(t)$ consists of song component $\soc(t)$
                     scaled by $\sca$ with optional noise component $\noc(t)$
                     and is successively transformed into envelope $\env(t)$,
                     logarithmically compressed envelope $\db(t)$, and
                     intensity-adapted envelope $\adapt(t)$.
                     \textbf{Top}:~Example representations of $\env(t)$,
                     $\db(t)$, and $\adapt(t)$ for different $\sca$.
                     \textbf{a}:~Noiseless case.
                     \textbf{b}:~Noisy case.
                     \textbf{Bottom}:~Intensity metrics over a range of $\sca$.
                     \textbf{c}:~Noiseless case: Standard deviations $\sigma_x$
                     of $\env(t)$, $\db(t)$, and $\adapt(t)$.
                     \textbf{d}:~Noisy case: Ratios of $\sigma_x$ of $\env(t)$,
                     $\db(t)$, and $\adapt(t)$ to the respective reference
                     standard deviation $\sigma_{\eta}$ for input
                     $\filt(t)=\noc(t)$. Shaded areas indicate $5\,\%$ (dark
                     grey) and $95\,\%$ (light grey) curve span for
                     $\adapt(t)$.
                     \textbf{e}:~Ratios of $\sigma_x$ to $\sigma_{\eta}$ of
                     $\adapt(t)$ as in \textbf{d} for different species
                     (averaged over songs and recordings, see appendix
                     Fig\,\ref{fig:app_log-hp_curves}). Dots indicate $95\,\%$
                     curve span per species.
                     }
    \label{fig:log-hp}
\end{figure}
\FloatBarrier

\subsection{Thresholding nonlinearity \& temporal averaging}

The third nonlinear transformation along the model pathway is the thresholding
nonlinearity $\nl$ that transforms each kernel response $c_i(t)$ into a binary
binary response $b_i(t)$, Eq.\,\ref{eq:binary}. This transformation takes place
after the convolutional filtering of $\adapt(t)$ with kernel $k_i(t)$,
Eq.\,\ref{eq:conv}, and is followed by the temporal averaging of $b_i(t)$ into
the feature set $f_i(t)$ by a lowpass filter, Eq.\,\ref{eq:lowpass}. The
effects of thresholding and temporal averaging are best illustrated based on a
single kernel~(Fig.\,\ref{fig:thresh-lp_single}) instead of the full set. For
this analysis, input $\adapt(t)$ was
rescaled~(Fig.\,\ref{fig:thresh-lp_single}a) and convolved with kernel $k(t)$.
The resulting kernel response $c(t)$ was passed through $H(c\,-\,\Theta)$ with
three different threshold values
$\Theta$~(Fig.\,\ref{fig:thresh-lp_single}b-d). Each resulting binary response
$b(t)$ was transformed into $f(t)$, whose average feature value serves as a
measure of intensity~(Fig.\,\ref{fig:thresh-lp_single}ef). The thresholding
nonlinearity $H(c\,-\,\Theta)$ categorizes the values of $c(t)$ into "relevant"
($c(t)>\Theta$, $b(t)=1$) and "irrelevant" ($c(t)\leq\Theta$, $b(t)=0$)
response values. It thereby splits the probability density $\pc$ of $c(t)$
within some observed time interval $T$ into two complementary parts around
$\Theta$:
\begin{equation}
    \int_{\Theta}^{+\infty} \pc\,dc\,=\,1\,-\,\int_{-\infty}^{\Theta} \pc\,dc\,=\,\frac{T_1}{T}, \qquad \infint \pc\,dc\,=\,1
    \label{eq:pdf_split}
\end{equation}
The right-sided part of the split $\pc$ corresponds to time $T_1$ where
$c(t)>\Theta$, while the left-sided part corresponds to time $T_0=T-T_1$ where
$c(t)\leq\Theta$. The semi-definite integral over the right-sided part of $\pc$
represents the ratio of time $T_1$ to total time $T$ because the indefinite
integral of a probability density is normalized to 1. The lowpass filtering of
$b(t)$ can be approximated as temporal averaging over a suitable time interval
$\tlp>\frac{1}{\fc}$ in order to express $f(t)$ as a similar temporal ratio
\begin{equation}
    f(t)\,\approx\,\frac{1}{\tlp} \int_{t}^{t\,+\,\tlp} b(\tau)\,d\tau\,=\,\frac{T_1}{\tlp}, \qquad b(t)\,\in\,\{0,\,1\}
    \label{eq:feat_avg}
\end{equation}
of time $T_1$ during which $b(t)$ is 1 within the averaging interval $\tlp$.
Therefore, the value of $f(t)$ at every time point $t$ approximately signifies
the cumulative probability that $c(t)$ exceeds $\Theta$ during the
corresponding averaging interval $\tlp$:
\begin{equation}
    f(t)\,\approx\,\int_{\Theta}^{+\infty} \pclp\,dc\,=\,P(c\,>\,\Theta,\,\tlp)
    \label{eq:feat_prop}
\end{equation}
In a sense, $f(t)$ can be interpreted as some sort of duty cycle with respect
to $\Theta$. For example, a feature value of $f(t)=0.4$ means that $c(t)$
exceeds $\Theta$ for approximately 40\,\% of the time within $\tlp$ around $t$.
In the most extreme cases, $\Theta$ lays either above the maximum of $c(t)$ or
below the minimum of $c(t)$, which results in a minimum or maximum possible
feature value of $f(t)=0$~(Fig.\,\ref{fig:thresh-lp_single}d, left column) or
$f(t)=1$, respectively.

Importantly, $f(t)$ neither retains information about the timing of individual
threshold crossings nor the precise values of $c(t)$ apart from their relation
to $\Theta$. Accordingly, for a given $\Theta$, different $\sca$ can still
result in similar $T_1$ segments (and hence similar feature values) depending
on the magnitude of the derivative of $c(t)$ in temporal proximity to time
points at which $c(t)$ crosses $\Theta$: The steeper the slope of $c(t)$, the
less $T_1$ changes with variations in $\sca$. The most reliable way of
exploiting this invariant porperty of $f(t)$ is to set $\Theta$ to a value near
0, because these values are least affected by different scales of $c(t)$. For
sufficiently large $\sca$, $f(t)$ then approaches the same constant value in
both the noiseless and the noisy case~(Fig.\,\ref{fig:thresh-lp_single}e,
saturation regime).

The value of $f(t)$ in the saturation regime is independent of the precise
value of $\Theta$, but the value of $\sca$ at which the saturation regime is
reached decreses with $\Theta$~(Fig.\,\ref{fig:thresh-lp_single}e). Therefore,
a threshold value of $\Theta=0$ would be the optimal choice for achieving
intensity invariance at the lowest possible $\sca$. In stark contrast, the
closer $\Theta$ is to 0, the higher the pure-noise response of $f(t)$ and the
lower the resulting SNR of $f(t)$ between noise regime and saturation
regime~(Fig.\,\ref{fig:thresh-lp_single}b-d, left column, and
Fig.\,\ref{fig:thresh-lp_single}e). It is even possible to achieve an
"unlimited" SNR of $f(t)$ by setting $\Theta$ above the maximum of the
pure-noise $c(t)$, so that any value of $f(t)$ greater than 0 indicates the
presence of the song component $\soc(t)$ in input $\adapt(t)$ at the cost of
requiring a higher $\sca$ to reach the saturation regime. This trade-off
between intensity invariance and SNR has already been observed during the
previous analysis on logarithmic compression and
adaptation~(Fig.\,\ref{fig:log-hp}d). However, the parameters that determine
the SNR of $\adapt(t)$ are much less understood and likely relate to properties
of the signal, whereas the SNR of $f(t)$ depends on the choice of $\Theta$ and
can be more directly manipulated by the system.

Finally, the effects of thresholding and temporal averaging must be seen in the
context of the previous transformation pair of logarithmic compression and
adaptation. 

Finally, the question remains whether the intensity-invariant output $\adapt(t)$
of the previous transformation pair allows feature 

Finally, the output $\adapt(t)$ of the previous transformation
pair~(Fig.\,\ref{fig:log-hp}cd) can be related to the input $\adapt(t)$ of the
current transformation pair by plotting the values of $f(t)$ over the standard
deviation of input $\adapt(t)$ instead of
$\sca$~(Fig.\,\ref{fig:thresh-lp_single}f). This is relevant because, unlike
$\sca$, the standard deviation of $\adapt(t)$ is capped to a maximum value of
around 10\,dB by the previous transformation pair~(Fig.\,\ref{fig:log-hp}cd)


\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_thresh_lp_single.pdf}
    \caption{\textbf{Intensity invariance through thresholding and temporal
                     averaging is mediated by the interaction of threshold
                     value and noise floor.}
                     Input $\adapt(t)$ consists of song component $\soc(t)$
                     scaled by $\sca$ with optional noise component $\noc(t)$
                     and is transformed into single kernel response $c(t)$,
                     binary response $b(t)$, and feature $f(t)$. Different
                     color shades indicate different threshold values $\Theta$
                     (multiples of reference standard deviation $\sigma_{\eta}$
                     of $c(t)$ for input $\adapt(t)=\noc(t)$, with darker
                     colors for higher $\Theta$).
                     \textbf{Left}:~Noisy case: Example representations of
                     $\adapt(t)$ as well as $c(t)$, $b(t)$, and $f(t)$ for
                     different $\sca$.
                     \textbf{a}:~$\adapt(t)$ with kernel $k(t)$ in black.
                     \textbf{b\,-\,d}: $c(t)$, $b(t)$, and $f(t)$ based on the
                     same $\adapt(t)$ from \textbf{a} but with different
                     $\Theta$.
                     \textbf{Right}:~Average value $\mu_f$ of $f(t)$ for each
                     $\Theta$ from \textbf{b\,-\,d}. Dots indicate $95\,\%$
                     curve span (noisy case).
                     \textbf{e}:~$\mu_f$ over a range of $\sca$, once for the
                     noisy case (solid lines) and once for the noiseless case
                     (dotted lines).
                     \textbf{f}:~Noisy case: $\mu_f$ over the standard
                     deviation of input $\adapt$ corresponding to the values of
                     $\sca$ shown in \textbf{e}. Shaded area indicates standard
                     deviations that would be capped in the output $\adapt(t)$
                     of the previous transformation pair (see
                     Fig.\,\ref{fig:log-hp}cd).
                     }
    \label{fig:thresh-lp_single}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_thresh_lp_species.pdf}
    \caption{\textbf{Feature representation of different species-specific songs
                     saturates at different points in feature space.}
                     Same input and processing as in
                     Fig.\,\ref{fig:thresh-lp_single} but with three different
                     kernels $k_i$, each with a single kernel-specific
                     threshold value $\thr=0.5\cdot\sigma_{\eta_i}$.
                     \textbf{a}:~Examples of species-specific grasshopper
                     songs.
                     \textbf{Middle}:~Average value $\mu_{f_i}$ of each feature
                     $f_i(t)$ over $\sca$ per species (averaged over songs and
                     recordings, see appendix
                     Figs.\,\ref{fig:app_thresh-lp_pure} and
                     \ref{fig:app_thresh-lp_noise}). Different color shades
                     indicate different kernels $k_i$. Dots indicate $95\,\%$
                     curve span per $k_i$.
                     \textbf{b}:~Noiseless case.
                     \textbf{c}:~Noisy case.
                     \textbf{Bottom}:~2D feature spaces spanned by each pair of
                     $f_i(t)$. Each trajectory corresponds to a
                     species-specific combination of $\mu_{f_i}$ that develops
                     with $\sca$ (colorbars). Horizontal dashes in the colorbar
                     indicate $5\,\%$ (dark grey) and $95\,\%$ (light grey)
                     curve span of the norm across all three $\mu_{f_i}$ per
                     species.
                     \textbf{d}:~Noiseless case.
                     \textbf{e}:~Noisy case. Shaded areas indicate the average
                     minimum $\mu_{f_i}$ across all species-specific trajectories.
                     }
    \label{fig:thresh-lp_species}
\end{figure}
\FloatBarrier
% \caption{\textbf{Rectification and lowpass filtering improves SNR
%                  but does not contribute to intensity invariance.}
%                  Input $\raw(t)$ consists of song component $\soc(t)$ scaled by
%                  $\sca$ with optional noise component $\noc(t)$ and is
%                  successively transformed into tympanal signal $\filt(t)$ and
%                  envelope $\env(t)$. Different line styles indicate different
%                  cutoff frequencies $\fc$ of the lowpass filter extracting
%                  $\env(t)$.
%                  \textbf{Top}:~Example representations of $\filt(t)$ and
%                  $\env(t)$ for different $\sca$.
%                  \textbf{a}:~Noiseless case.
%                  \textbf{b}:~Noisy case.
%                  \textbf{Bottom}:~Intensity metrics over a range of $\sca$.
%                  \textbf{c}:~Noiseless case: Standard deviations of $\filt(t)$
%                  and $\env(t)$.
%                  \textbf{d}:~Noisy case: Ratios of standard deviations of
%                  $\filt(t)$ and $\env(t)$ to the respective reference standard
%                  deviation for input $\raw(t)=\noc(t)$.
%                  \textbf{e}:~Ratios of standard deviations of $\env(t)$ as in
%                  \textbf{b} for different species (averaged over songs and
%                  recordings, see appendix Fig.\,\ref{fig:app_rect-lp}).
%                 }
\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_full_Omocestus_rufipes.pdf}
    \caption{\textbf{Step-wise emergence of intensity-invariant song
                     representation along the full model pathway.}
                     Input $\raw(t)$ consists of song component $\soc(t)$
                     scaled by $\sca$ with added noise component $\noc(t)$ and
                     is processed up to the feature set $f_i(t)$. Different
                     color shades indicate different types of Gabor kernels
                     with specific lobe number $\kn$ and either $+$ or $-$
                     sign, sorted (dark to light) first by increasing $\kn$ and
                     then by sign~($1\,\leq\,\kn\,\leq\,4$; first $+$, then $-$
                     for each $\kn$; five kernel widths $\kw$ of 1, 2, 4, 8,
                     and $16\,$ms per type; 8 types, 40 kernels in total).
                     \textbf{a}:~Example representations of $\filt(t)$,
                     $\env(t)$, $\db(t)$, $\adapt(t)$, $c_i(t)$, and $f_i(t)$
                     for different $\sca$.
                     \textbf{b}:~Intensity metrics over $\sca$. For $c_i(t)$
                     and $f_i(t)$, the median over kernels is shown. Dots
                     indicate $95\,\%$ curve span for $\db(t)$, $\adapt(t)$,
                     $c_i(t)$, and $f_i(t)$.
                     \textbf{c}:~Average value $\mu_{f_i}$ of each feature
                     $f_i(t)$ over $\sca$.
                     \textbf{d}:~Ratios of intensity metrics to the respective
                     reference value for input $\raw(t)=\noc(t)$. For $c_i(t)$
                     and $f_i(t)$, the median over kernel-specific ratios is
                     shown.
                     \textbf{e}:~Ratios of standard deviation $\sigma_{c_i}$ of
                     each $c_i(t)$.
                     \textbf{f}:~Ratios of $\mu_{f_i}$.
                     \textbf{g}:~Distributions of kernel-specific $\sca$ that
                     correspond to $95\,\%$ curve span for $c_i(t)$ and
                     $f_i(t)$. Dots indicate the values from \textbf{b}.
                     }
    \label{fig:pipeline_full}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_short_Omocestus_rufipes.pdf}
    \caption{\textbf{Step-wise emergence of intensity invariant song
                     representation along the model pathway without logarithmic
                     compression.}
                     Input $\raw(t)$ consists of song component $\soc(t)$
                     scaled by $\sca$ with added noise component $\noc(t)$ and
                     is processed up to the feature set $f_i(t)$, skipping
                     $\db(t)$. Different color shades indicate different types
                     of Gabor kernels with specific lobe number $\kn$ and
                     either $+$ or $-$ sign, sorted (dark to light) first by
                     increasing $\kn$ and then by
                     sign~($1\,\leq\,\kn\,\leq\,4$; first $+$, then $-$ for
                     each $\kn$; five kernel widths $\kw$ of 1, 2, 4, 8, and
                     $16\,$ms per type; 8 types, 40 kernels in total).
                     \textbf{a}:~Example representations of $\filt(t)$,
                     $\env(t)$, $\adapt(t)$, $c_i(t)$, and $f_i(t)$ for
                     different $\sca$.
                     \textbf{b}:~Intensity metrics over $\sca$. For $c_i(t)$
                     and $f_i(t)$, the median over kernels is shown. Dots
                     indicate $95\,\%$ curve span for $f_i(t)$.
                     \textbf{c}:~Average value $\mu_{f_i}$ of each feature
                     $f_i(t)$ over $\sca$.
                     \textbf{d}:~Ratios of intensity metrics to the respective
                     reference value for input $\raw(t)=\noc(t)$. For $c_i(t)$
                     and $f_i(t)$, the median over kernel-specific ratios is
                     shown.
                     \textbf{e}:~Ratios of $\mu_{f_i}$.
                     \textbf{f}:~Distribution of kernel-specific $\sca$ that
                     correspond to $95\,\%$ curve span for $f_i(t)$. Dots
                     indicate the value from \textbf{b}.
                     }
    \label{fig:pipeline_short}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_features_cross_species.pdf}
    \caption{\textbf{Interspecific and intraspecific feature variability.}
                     Average value $\mu_{f_i}$ of each feature $f_i(t)$ against
                     its counterpart from a 2nd feature set based on a
                     different input $\raw(t)$. Each dot within a subplot
                     represents a single feature $f_i(t)$. Different color
                     shades indicate different types of Gabor kernels with
                     specific lobe number $\kn$ and either $+$ or $-$ sign,
                     sorted (dark to light) first by increasing $\kn$ and then
                     by sign~($1\,\leq\,\kn\,\leq\,4$; first $+$, then $-$ for
                     each $\kn$; five kernel widths $\kw$ of 1, 2, 4, 8, and
                     $16\,$ms per type; 8 types, 40 kernels in total). Data is
                     based on the analysis underlying
                     Fig\,\ref{fig:pipeline_full}.
                     \textbf{Lower triangular}:~Interspecific comparisons
                     between single songs of different species.
                     \textbf{Upper triangular}:~Intraspecific comparisons
                     between different songs of a single species (\textit{O.
                     rufipes}).
                     \textbf{Lower left}:~Distribution of correlation
                     coefficients $\rho$ for each interspecific and
                     intraspecific comparison. Dots indicate single $\rho$
                     values.
                     }
    \label{fig:feat_cross_species}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_field.pdf}
    \caption{\textbf{Step-wise emergence of intensity invariant song
                     representation along the model pathway.}
                     }
    \label{fig:pipeline_field}
\end{figure}
\FloatBarrier

\section{Conclusions \& outlook}

\textbf{Song recognition pathway: Grasshopper vs. model:}\\
The model pathway includes a rather large number of Gabor kernels compared to
the 15 to 20 ascending neurons in the grasshopper auditory
system~(\bcite{stumpner1991auditory}). 


\textbf{Definition of invariance (general, systemic):}\\
Invariance = Property of a system to maintain a stable output with respect to a
set of relevant input parameters (variation to be represented) but irrespective
of one or more other parameters (variation to be discarded)
$\rightarrow$ Selective input-output decorrelation

\textbf{Definition of intensity invariance (context of neurons and songs):}\\
Intensity invariance = Time scale-selective sensitivity to certain faster
amplitude dynamics (song waveform, small-scale AM) and simultaneous
insensitivity to slower, more sustained amplitude dynamics (transient baseline,
large-scale AM, current overall intensity level)\\
$\rightarrow$ Without time scale selectivity, any fully intensity-invariant
output will be a flat line


\textbf{Log-HP: Implication for intensity invariance:}\\
- Logarithmic scaling is essential for equalizing different song intensities\\
$\rightarrow$ Intensity information can be manipulated more easily when in form
of a signal offset in log-space than a multiplicative scale in linear space

- Capability to compensate for intensity variations, i.e. selective amplification
of output $\adapt(t)$ relative to input $\env(t)$, is limited by input SNR (Eq.\,\ref{eq:toy_snr}):\\
$\rightarrow$ Ability to equalize between different sufficiently large scales of $s(t)$\\
$\rightarrow$ Inability to recover $s(t)$ when initially masked by noise floor $\eta(t)$

- Logarithmic scaling emphasizes small amplitudes (song onsets, noise floor) \\
$\rightarrow$ Recurring trade-off: Equalizing signal intensity vs preserving initial SNR

\textbf{Thresh-LP: Implication for intensity invariance:}\\
- Role of song periodicity for feature representation!

- Suggests a relatively simple rule for optimal choice of threshold value $\thr$:\\
$\rightarrow$ Find amplitude $c_i$ that maximizes absolute derivative of $c_i(t)$ over time\\
$\rightarrow$ Optimal with respect to intensity invariance of $f_i(t)$, not necessarily for
other criteria such as song-noise separation or diversity between features

- Nonlinear operations can be used to detach representations from graded physical
stimulus (to fasciliate categorical behavioral decision-making?):\\
1) Capture sufficiently precise amplitude information: $\env(t)$, $\adapt(t)$\\
$\rightarrow$ Closely following the AM of the acoustic stimulus\\
2) Quantify relevant stimulus properties on a graded scale: $c_i(t)$\\
$\rightarrow$ More decorrelated representation, compared to prior stages\\
3) Nonlinearity: Distinguish between "relevant vs irrelevant" values: $b_i(t)$\\
$\rightarrow$ Trading a graded scale for two or more categorical states\\
4) Represent stimulus properties under relevance constraint: $f_i(t)$\\
$\rightarrow$ Graded again but highly decorrelated from the acoustic stimulus\\
5) Categorical behavioral decision-making requires further nonlinearities\\
$\rightarrow$ Parameters of a behavioral response may be graded (e.g. approach speed),
initiation of one behavior over another is categorical (e.g. approach/stay)


\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_noise_env_sd_conversion_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_env-sd}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_rect-lp_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_rect-lp}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_log-hp_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_log-hp_curves}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
   \centering
   \includegraphics[width=\textwidth]{figures/fig_saturation_log-hp_appendix.pdf}
   \caption{\textbf{}
                    }
   \label{fig:app_log-hp_saturation}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_thresh-lp_pure_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_thresh-lp_pure}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_thresh-lp_noise_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_thresh-lp_noise}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_kernel_sd_perc_thresh_lp_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_thresh-lp_kern-sd}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_kernel_sd_perc_full_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_full_kern-sd}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_kernel_sd_perc_short_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_short_kern-sd}
\end{figure}
\FloatBarrier

\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_kernel_sd_perc_field_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_field_kern-sd}
\end{figure}
\FloatBarrier


\begin{figure}[!ht]
    \centering
    \includegraphics[width=\textwidth]{figures/fig_invariance_cross_species_thresh_appendix.pdf}
    \caption{\textbf{}
                     }
    \label{fig:app_cross_species_thresh}
\end{figure}
\FloatBarrier

\end{document}