There are virtually infinite manners by which a constant stimulus can impinge into the sensorium of an animal. For instance, our noses have many receptors that can sense a given odorant molecule, but only a small subset of those are excited each time the odour is perceived (Axel, 1995). Similarly, photons hit and excite photoreceptors in the retina randomly and sparsely for a given presented image (Dowling, 1987). An important aspect of sensation and perception is that never the exact same receptors are excited every time we perceive and, still, our perceptual experiences feel quite stable. So the brain must have a way to transit from lack of invariance at the microscopic sensory level towards invariance at the macroscopic level, which ultimately supports the invariant aspects of conscious perception and behaviour. However, we observe significant variability in the brain responses at the trial level (Stein et al., 2005; McIntosh et al., 2008; Garrett et al., 2013), including at the earliest layers of the perceptual hierarchy (Croner et al., 1993; Freeman, 1978)—that is, each perceptual experience is associated with a unique neural trajectory that does not repeat. How the gap between stability in subjective perception and the changing nature of brain responses is bridged is an important question in neuroscience. Here, I investigate the distributed aspects of brain activity that are most stable across experimental repetitions, and are therefore most likely to relate to stable perceptual experiences.

Decoding analysis uses multivariate machine-learning algorithms to predict the identity of the observed stimulus from brain data (Haxby et al., 2014; Stokes et al., 2015). This way, per time point, it estimates a function of the data that maximally discriminate between conditions, as well as a temporal measure of accuracy that reflects how much information the data convey about the stimuli per time point. The assessment, via decoding accuracy, of how the discriminative space changes throughout the trial offers a view of the properties of stimuli representation and processing. However, it is not straightforward to know what specific aspects of the signal cause the patterns of decoding accuracy that we observe in perceptual experiments. Without this capacity, it is hard to link these patterns to actual neural mechanisms.

To gain insight into what aspects of the signal underpin decoding accuracy, and therefore the most stable aspects of stimulus processing, I introduce a generative model of multichannel electrophysiological activity (e.g. electroencephalography[EEG] or magnetoencephalography [MEG]) that, under no stimulation, exhibits chaotic phasic and amplitude fluctuations; and that, when stimulated, responds by manipulating certain aspects of the data, such as ongoing phase or signal amplitude, in a stimulus-specific manner. Specifically, in every trial, each channel may or may not respond to stimulation, according to a certain probability. In the model, when a channel responds, it can do it in different ways: (1) by phase resetting the ongoing oscillation to a given target phase and then entraining to a given frequency, (2) by an additive oscillatory response independent of the ongoing oscillation, (3) by modulating the amplitude of the stimulus-relevant oscillations, or (4) by an additive non-oscillatory (slower) response. This (not exhaustive) list of effects was considered given previous literature (Shah et al., 2004; Mazaheri and Jensen, 2006; Makeig et al., 2002; Vidaurre et al., 2021), and each effect may be underpinned by distinct neural mechanisms. For example, it is not completely clear the extent to which stimulus processing is sustained by oscillations, and disentangling these effects can help resolving this question. I named this model genephys by generative model of empirical electrophysiological signals. Genephys is empirical in the sense that it purely accounts for features in the signal that are observable, without making assumptions about the underlying neurobiological causes; that is, it can generate signals that, with the right parametrisation, can share empirical properties with real data. In particular, when confronted with a decoding algorithm, the data generated by this model can show patterns of decoding accuracy with similar characteristics to what we observe in real data.

Given the effects that we observe in the stimulus processing literature, and using an example of visual perception as a reference, I observed that two different mechanisms can produce realistic decoding results as we see in real perception: either phase resetting to a stimulus-specific phase followed by frequency entrainment, or an additive oscillation (unrelated to the ongoing oscillations) with a stimulus-specific phase. Either way, a cross-frequency coupling effect is also necessary, where an additive slower response holds a specific phasic relation with the oscillatory (faster) response. Furthermore, the stimulus-related oscillation needs to span multiple channels, and have a diversity of frequencies and latencies of response across channels. Other experimental paradigms, including motor tasks and decision making, can be investigated with genephys, which is publicly available as a Python package in PyPI http://github.com/vidaurre/genephys; http://genephys-doc.readthedocs.io/en/latest/.

I confronted genephys to classification-based decoding analysis to characterise how brain activity carries stimulus-specific information. In each simulation, I generated 10 data sets per combination of parameters, each with N = 250 trials and T = 250 time points per trial (1 s for a sampling frequency of 250 Hz). The number of channels is 32. Only one endogenous oscillation was sampled, with (angular) frequencies spontaneously ranging between 0.01 and 0.25π (0.4–10 Hz). I computed a TGM per run and took the average across runs.

For reference, I considered MEG data recorded while participants viewed object images across 118 categories, as presented in Cichy et al., 2016. Each image category was presented 30 times. Presentation occurred during the first 500 ms of the trial, and trials were 1 s long, sampled at 250 Hz. The multi-channel sensor-space data, epoched around the presentation of each visual stimulus, can be used to train a decoder to predict which visual image is being presented (Cichy et al., 2016; Vidaurre et al., 2021). Here, I decoded whether the image corresponded to an animate category (a dog) or inanimate (a pencil). Figure 2A shows a TGM for an example subject, where some archetypal characteristics are highlighted. In the experiments below, specifically, I focus on the strong narrow diagonal at the beginning of the trial, the broadening of accuracy later in the trial, and the vertical/horizontal bars of higher-than-chance accuracy. Importantly, given the variability observed across subjects (as seen in Figure 2B, which shows TGMs for six subjects), this example is only meant as a reference; therefore I did not optimise the model hyperparameters to this TGM (except in the last subsection), or showed any quantitative metric of similarity.

Oscillatory components underlying the observed decoding patterns

In real data, we often see oscillatory patterns in the TGM, indicating that the subspace of brain activity that carries information about the stimulus must have oscillatory elements. At least two distinct mechanisms may be behind this phenomenon: first, an ongoing oscillation might reset its phase and then entrain to a given frequency in a stimulus-specific fashion; second, a stimulus-specific oscillatory response might by added to the signal after stimulus presentation on top of the existing ongoing oscillation. Essentially, the difference between the two is that, in the phase-resetting case, the ongoing oscillations are altered; while in the other case the additive oscillation coexists with the ongoing oscillations. Next, I use genephys together with decoding analysis to compare between these two alternatives, showing that both can produce decoding patterns similar to what we observe empirically in real experiments.

In the simulations, all 32 channels convey information but with a relatively low activation probability (${\displaystyle \theta =1/6}$). For phase resetting and frequency entrainment, I considered a diverse range of entrainment frequencies (between 0.1 and 0.2 in angular frequency) and latencies of response (${\displaystyle {\delta }_1=0-160\mathrm{m}\mathrm{s}}$) across channels. For the additive oscillatory response, I considered a similar range of frequencies and latencies of response across channels. (I will show below that channel stochasticity and frequency diversity are both important to produce realistic decoding patterns.) The difference between the two fictitious stimuli lied in their different target frequencies (i.e. $\overline{{\phi }^k}$ for phase resetting and ${\gamma }^k$ for the additive oscillation; see specification of the model in Methods). I also included an additive non-oscillatory response with a stimulus-specific amplitude, which (as I will also show later) is important to produce realistic decoding results. I did not optimise the parameters of the model to reflect the fine details from real-data TGMs, since TGMs vary across subjects (see Figure 2B) and experimental paradigms (King and Dehaene, 2014).

Figure 3 shows TGMs (top) together with one-channel ERPs (where each stimulus is represented by a different tonality of blue or red; bottom) for the sampled signal $x$ and its various constitutive elements: ongoing phase $\phi$, frequency $f$, amplitude $a$, additive non-oscillatory response $z$, and, when applicable, an additive oscillatory response $y$. The left panels show results for phase reset plus frequency entrainment, where we can see an effect on the ongoing phase. The middle panels show results for the additive oscillation; here, there is no effect on the ongoing phase, and, instead, the additive oscillatory response $y$ is shown at the bottom. The right panels show an example from real data, where phase, frequency, and amplitude were computed analytically using the Hilbert transform on the filtered data.

Figure 3

Download asset Open asset

Although the exact details differ from the right panels of the figure, both types of effects produce patterns reproducing the characteristic signatures of real data. These include the strong diagonal, the vertical/horizontal bands of high generalisation accuracy, and the broadening of accuracy in later stages of the trial. Note that phase resetting plus frequency entrainment is, everything else held constant, a stronger effect than the additive oscillatory response. This is because, for the latter, the ongoing oscillations (here, non-stimulus specific) can interfere with the phase of the additive response, impeding cross-trial phase locking throughout the trial. For phase resetting, on the other hand, the ongoing oscillation is the effect and there is no interference. In this particular example, anyway, the range for the additive oscillation (0.1–0.2) was much narrower than that of the ongoing oscillation (0.01–0.25π), making the interference more unlikely; that is, the ongoing oscillation phase is approximately averaged out, and treated as noise by the decoding algorithm.

The distribution of stimulus-specific information spans multiple channels and is stochastic

Next, I use genephys to show that the stimulus-specific information spans a large number of channels, and do so stochastically. I focus on the additive oscillatory response effect, which, as shown in the previous section, can produce comparable results to phase resetting plus frequency entrainment.

I varied the number of relevant channels $p$ as well as the probability of activation $\theta$ for the relevant channels, so that the subset of channels that activate is different for every trial. Figure 4 shows the TGMs for various combinations of $p$ and $\theta$, from a configuration where there are few relevant channels that always respond (${\displaystyle p=1,\theta =1}$) to another where there are many relevant channels that only respond sparsely (${\displaystyle p=32,\theta =1/8}$). As previously, channels have diverse frequencies and latencies of response. An additive non-oscillatory response is also included as an effect.

Figure 4

Download asset Open asset

Contrary to empirical TGMs (which have a relatively stylised diagonal), having only a few relevant channels (first three panels of Figure 4) produces unrealistically geometric patterns. This indicates that, in real data, the subspace of the data that contains information about the stimulus needs to be multi-dimensional; that is, that the amount of relevant channels must be relatively high (as in the fourth panel of Figure 4). However, the contribution of each channel must entail some sort of noise or instability (in this case expressed by having probabilistic activations), or else the decoding accuracy becomes unrealistically perfect. In summary, the subspace of brain activity that carries out information about the condition is highly stochastic at the single-trial level.

The oscillatory effect spans multiple frequencies and latencies

In the previous sections, I showed that a noisy, additive oscillation effect (or, alternatively, a phase reset plus frequency entrainment effect) across multiple channels can generate decoding patterns as we see in real data. Next, I demonstrate that the effect must span a diversity of frequencies across channels, as well as a diversity of latencies of response. In real data, frequency diversity can be expressed as, for example, a gradient of frequencies from primary towards more associative areas; while latency diversity could reflect phenomena such that primary areas responding earlier than associative areas.

For frequency diversity, each channel is endowed with a different effect-related frequency, such that frequencies are not multiples of each other (specifically, they have different values of ${\omega }^k$ between 0.1 and 0.2 in angular frequency; see Methods). For latency diversity, channels do not respond simultaneously but with different values of ${\delta }_1$ between 0 and 120 ms. Figure 5 shows a two-by-two design. On the left column, I set genephys to have a uniform frequency of response (i.e. all channels have the same frequency of response), while on the right column it uses a diversity of frequencies. On the top row, we have a uniform latency of response (i.e. all channels have the same latency of response), whereas on the bottom row we have a diversity of latencies of response. See Figure 5—figure supplement 1 for a similar result for phase resetting followed by frequency entrainment.

Figure 5 with 1 supplement see all

Download asset Open asset

As mentioned, real data normally yield a relatively tight band of high decoding accuracy along the diagonal, often accompanied from contiguous parallel bands of below-baseline accuracy. Critically, the fact that we do not typically observe a chequer pattern means that the trajectory of phase across channels does not repeat itself periodically. If it did, it would show patterns as in the top-left panel, where the uniformity of frequencies and latencies gives rise to an unrealistically regular pattern—such that a decoder trained at a certain time point will become equally accurate again after one cycle at the entrained frequency. Having a diversity of latencies but not of frequencies produces another regular pattern consisting of alternating, parallel bands of higher/lower than baseline accuracy. This, shown in the bottom-left panel, is not what we see in real data either. Having a diversity of frequencies but not of latencies gets us closer to a realistic pattern, as we see in the top-right panel. Finally, having diversity of both frequencies and latencies produces the most realistic pattern, as we can see by comparing to the examples in Figure 2, and many others throughout the literature (King and Dehaene, 2014). Similar conclusions can be drawn from Figure 5—figure supplement 1 for phase resetting plus frequency entrainment.

In summary, these results show that it is not only important for the stimulus-relevant subspace of activity to be spatially high dimensional, but also temporally high dimensional.

A slower additive component is coupled with the oscillatory response

In real data, we normally see a broadening of decoding accuracy in the TGM as time progresses throughout the trial (see Figure 2). This is often interpreted as neural representations becoming more stable at latter stages of the trial, which is putatively linked to memory encoding and higher cognitive processes. In Figure 6, I show that this effect can be reproduced on synthetic data through the addition of a slowly progressing, non-oscillatory response.

Figure 6

Download asset Open asset

Specifically, I set up a response function such that, after stimulus presentation, the non-oscillatory additive response ramps up to a stimulus-specific target value in about 100 s, and then slowly decays to finally vanish at around 800 ms. Also, I modulate the strength of the oscillatory response using values of $\overline{z^k}$ (see Methods) that differ between the two stimuli by a magnitude that ranges from 0.0 (no difference) to 1.0 (for reference, the examples in the previous figures had a difference of 0.5). As seen in Figure 6A, the strength of decoding accuracy grows as the difference in the slow response increases.

Another feature commonly seen on real data are the vertical and horizontal bars of high accuracy stemming from the time point of maximum accuracy (see Figure 2A), which is sometimes interpreted as evidence of stimulus representation recurrence in the brain. I show in Figure 6B that this feature emerges from a phase coupling between the oscillatory component and the slower component (with respect to the stimuli). For example, following the notation established in Methods, an in-phase relationship means that the sign of ${\displaystyle \left(\overline{{\phi }^1}-\overline{{\phi }^2}\right)}$ at $t_{\text{max}}$ is the same than the sign of ${\displaystyle \left(\overline{z^1}-\overline{z^2}\right)}$. Note that the sign of ${\displaystyle \left(\overline{z^1}-\overline{z^2}\right)}$ will likely be maintained for most of the trial since this component is slow, therefore creating the effect in the TGM. That was the case for all panels of Figure 6A. For Figure 6B, while keeping ${\displaystyle \left(\overline{z^1}-\overline{z^2}\right)}$ = 0.5 (as in the middle panel of Figure 6A), I varied the phase consistency between the oscillatory and the non-oscillatory components. In the leftmost panel, the oscillatory and the non-oscillatory components are in-phase for all channels, while in the rightmost panel, they are anti-phase for all channels; in between, 25%, 50%, and 75% of the channels are in-phase (and 75%, 50%, and 25% are anti-phase, respectively). As observed, the type of phase consistency between the oscillatory and the non-oscillatory component has a strong impact on the TGM. In particular, the in-phase relation bears the most consistent patterns with the considered real data.

Amplitude increases modulate the size of the effect

Are modulations in the amplitude of the stimulus-specific oscillation necessary for the effects we observe in real data? They are not, but they can enhance the already existing patterns.

I generated data sets with an additive oscillatory component effect. In each of them, I applied a different amount of amplitude enhancement, from no increase to a fivefold increase. These did not entail any effect on the signals’ phase. I considered two characteristic features of the TGM: (1) the diagonal, and (2) a vertical slice stemming from the time point of maximum accuracy. Figure 7 shows the results. As observed, the main features are present in all runs regardless of the size of the amplitude effect, although they were more prominent for higher amplitude modulations. An amplitude modulation on a phase-resetting effect causes a similar effect (not shown). Note that an amplitude effect without a phase effect would not result in any phase locking across trials, and therefore could not lead to any significant decoding accuracy.

Figure 7

Download asset Open asset

Fitting an empirical TGM

The previous analyses were descriptive in the sense that they did not quantify how much the generated TGMs resembled a specific empirical TGM. This was deliberate, because empirical TGMs vary across subjects and experiments, and I aimed at characterising them as generally as possible by looking at some characteristic features in broad terms. For example, while TGMs typically have a strong diagonal and horizontal/vertical bars of high accuracy, questions such as when these effects emerge and for how long are highly dependent on the experimental paradigm. For the same reason, I did not optimise the model hyperparameters, limiting myself to observing the behaviour of the model across some characteristic configurations. But, often, one’s interests are more specific. Then, it would be interesting to optimise some key hyperparameters of the model to maximise the correlation with a particular empirical TGM.

To illustrate how a data set could be more explicitly considered, I took an empirical TGM computed from the visual experiment in Cichy et al., 2016. Specifically, this is a cross-average TGM (over 10 subjects) obtained from decoding animate versus inanimate stimuli. Using an additive oscillatory response, and fixing the rest of the hyperparameters to a sensible configuration (the one that produced the middle panel in Figure 6A), I varied, within a reasonable range, two parameters of the response function that control the temporal shape of the effect: the rise slope ${\delta }_2$ and the fall slope ${\delta }_3$ ; see Methods. From each pair of parameters, I generated 10 data sets and computed 10 TGMs; I then correlated the average of these with the empirical TGM from real data. Figure 8 shows a heatmap of the resulting correlations. For a specific configuration (${\displaystyle {\delta }_2=0.06s,{\delta }_3=0.09s}$), the correlation peaks at r = 0.7.

Figure 8

Download asset Open asset

Brain’s responses to even the simplest stimuli are characteristically variable. This is not surprising, given the brain’s plasticity, and the fact that its endogenous activity is ever-changing. It also speaks to the brain’s degeneracy (Edelman and Gally, 2001)—that is that there are many possible neural trajectories that can achieve a single purpose. How this variability translates into behaviour and experience is however an open question. Anyhow, it seems reasonable that, at some level, brain responses must keep some invariant aspects so that our perceptual experiences remain stable. Here, I investigated, using a novel generative model, the most stable aspects of brain responses to stimuli as seen through the lens of decoding accuracy, which is, by definition, based on averaging. Previous work has analysed the nature of brain responses to perceptual stimulation using average evoked responses (Sauseng et al., 2007), arguing either for a predominant role of additive responses (Shah et al., 2004; Mazaheri and Jensen, 2006) or phase resetting (Makeig et al., 2002). These studies looked at the average response to a given stimulus, but did not investigate what aspects of the data carry information about the identity of the stimulus; this is the focus of this paper and the goal of genephys, the proposed model.

Genephys has different available types of effect, including phase resets, additive damped oscillations, amplitude modulations, and non-oscillatory responses. All of these elements, which may relate to distinct neurobiological mechanisms, are configurable and can be combined to generate a plethora of TGMs that, in turn, can be contrasted to specific empirical TGMs. This way, we can gain insight on what mechanisms might be at play in a given task.

The demonstrations here are not meant to be tailored to a specific data set, and are, for the most part, intentionally qualitative. TGMs do vary across experiments and subjects; and the hyperparameters of the model can be explicitly optimised to specific scientific questions, data sets, and even individuals. In order to explore the space of configurations effectively, an automatic optimisation of the hyperparameter space using, for instance, Bayesian optimisation (Lorenz et al., 2017) could be advantageous. This may lead to the identification of very specific (spatial, spectral, and temporal) features in the data that may be neurobiologically interpreted.

Importantly, the list of effects that I have explored here is not exhaustive. For example, I have considered additive oscillatory responses in the form of sinusoidal waves. Another possibility could be to have additive oscillatory responses that are non-linear, that is with a tendency to spend more time in certain phases (e.g. having wider peaks than throughs); in this case, even in the absence of phase locking between trials, we could potentially have a significant evoked response due to phase asymmetry (Nikulin et al., 2007). For simplicity, also, I have considered independent sources of activity that exhibit correlations only through the induced effect. In practice, brain areas are in constant communication even in the absence of stimulation. Alternatives where sources are modelled as coupled oscillators (Breakspear et al., 2010; Cabral et al., 2014), or where there is correlated noise, are also possible. Also for simplicity, I have considered that every channel has one endogenous fundamental frequency only; in practice, multiple ongoing frequencies coexist and interact, potentially affecting the TGM if they are not completely averaged out across trials. Finally, the inherent stochasticity of the stimulus-specific space of activity can take various forms; here, I have explored a probabilistic activation of the channels, but others, such as noise in the phase distribution, are also possible and can also be simulated using genephys.

Also importantly, I have shown that standard decoding analysis can differentiate between these explanations only to some extent. For example, the effects induced by phase resetting and the use of additive oscillatory components are not enormously different in terms of the resulting TGMs. In future work, alternatives to standard decoding analysis and TGMs might be used to disentangle these sources of variation (Vidaurre et al., 2019).

Overall, the results obtained from applying genephys suggest that the stable aspects of brain activity regarding stimulus processing comprise phasic modulations of an oscillatory component coupled with a slower component, with an important role played by the nature of this coupling. The subspace of brain activity induced by the effect is high dimensional in both the spatial domain (it must span many channels) and the frequency domain (it must involve a great diversity of frequencies and exhibit a diversity of latencies). This effect may be accompanied by an amplitude modulation. Above and beyond these average patterns, the stimulus-specific subspace of brain responses remains highly stochastic at the trial level.

The simulated data can be regenerated using the code included as an example in the genephys Github repository at http://github.com/vidaurre/genephys (copy archived at Vidaurre, 2024). The real visual paradigm is openly available from the original publication Cichy et al., 2016 at: http://userpage.fu-berlin.de/rmcichy/fusion_project_page/main.html.

1. 1. Cichy RM
   2. Pantazis D
   3. Oliva D

   (2016) Userpage

   ID rmcichy/fusion_project_page/main.html. Spatio-temporal dynamics of information flow in ventral & dorsal visual cortex - 118 images.

   http://userpage.fu-berlin.de/rmcichy/fusion_project_page/main.html
2. 1. Cichy RM
   2. Pantazis D
   3. Oliva D

   (2018) MIT CSAIL

   ID MEG2_MEG_Epoched_Raw_Data/. Index of /MEG2_MEG_Epoched_Raw_Data.

   http://wednesday.csail.mit.edu/MEG2_MEG_Epoched_Raw_Data/

1. 1. Axel R

   (1995) The molecular logic of smell

   Scientific American 273:154–159.

   https://doi.org/10.1038/scientificamerican1095-154

   • PubMed
   • Google Scholar
2. 1. Breakspear M
   2. Heitmann S
   3. Daffertshofer A

   (2010) Generative models of cortical oscillations: Neurobiological implications of the kuramoto model

   Frontiers in Human Neuroscience 4:190.

   https://doi.org/10.3389/fnhum.2010.00190

   • PubMed
   • Google Scholar
3. 1. Cabral J
   2. Luckhoo H
   3. Woolrich M
   4. Joensson M
   5. Mohseni H
   6. Baker A
   7. Kringelbach ML
   8. Deco G

   (2014) Exploring mechanisms of spontaneous functional connectivity in MEG: How delayed network interactions lead to structured amplitude envelopes of band-pass filtered oscillations

   NeuroImage 90:423–435.

   https://doi.org/10.1016/j.neuroimage.2013.11.047

   • PubMed
   • Google Scholar
4. 1. Cichy RM
   2. Pantazis D
   3. Oliva A

   (2016) Similarity-based fusion of MEG and fMRI reveals spatio-temporal dynamics in human cortex during visual object recognition

   Cerebral Cortex 26:3563–3579.

   https://doi.org/10.1093/cercor/bhw135

   • PubMed
   • Google Scholar
5. 1. Croner LJ
   2. Purpura K
   3. Kaplan E

   (1993) Response variability in retinal ganglion cells of primates

   PNAS 90:8128–8130.

   https://doi.org/10.1073/pnas.90.17.8128

   • PubMed
   • Google Scholar
6. 1. Dawson GD

   (1954) A summation technique for the detection of small evoked potentials

   Electroencephalography and Clinical Neurophysiology 6:65–84.

   https://doi.org/10.1016/0013-4694(54)90007-3

   • PubMed
   • Google Scholar
7. Book

   1. Dowling JE

   (1987) The Retina: An Approachable Part of the Brain

   Harvard University Press.

   https://doi.org/10.1097/OPX.0b013e3182805b2b

   • Google Scholar
8. 1. Edelman GM
   2. Gally JA

   (2001) Degeneracy and complexity in biological systems

   PNAS 98:13763–13768.

   https://doi.org/10.1073/pnas.231499798

   • PubMed
   • Google Scholar
9. 1. Freeman WJ

   (1978) Spatial properties of an EEG event in the olfactory bulb and cortex

   Electroencephalography and Clinical Neurophysiology 44:586–605.

   https://doi.org/10.1016/0013-4694(78)90126-8

   • PubMed
   • Google Scholar
10. 1. Garrett DD
    2. Samanez-Larkin GR
    3. MacDonald SWS
    4. Lindenberger U
    5. McIntosh AR
    6. Grady CL

    (2013) Moment-to-moment brain signal variability: a next frontier in human brain mapping?

    Neuroscience and Biobehavioral Reviews 37:610–624.

    https://doi.org/10.1016/j.neubiorev.2013.02.015

    • PubMed
    • Google Scholar
11. 1. Grootswagers T
    2. Wardle SG
    3. Carlson TA

    (2017) Decoding dynamic brain patterns from evoked responses: A tutorial on multivariate pattern analysis applied to time series neuroimaging data

    Journal of Cognitive Neuroscience 29:677–697.

    https://doi.org/10.1162/jocn_a_01068

    • PubMed
    • Google Scholar
12. 1. Haxby JV
    2. Connolly AC
    3. Guntupalli JS

    (2014) Decoding neural representational spaces using multivariate pattern analysis

    Annual Review of Neuroscience 37:435–456.

    https://doi.org/10.1146/annurev-neuro-062012-170325

    • PubMed
    • Google Scholar
13. 1. King JR
    2. Dehaene S

    (2014) Characterizing the dynamics of mental representations: the temporal generalization method

    Trends in Cognitive Sciences 18:203–210.

    https://doi.org/10.1016/j.tics.2014.01.002

    • PubMed
    • Google Scholar
14. 1. Kragel PA
    2. Koban L
    3. Barrett LF
    4. Wager TD

    (2018) Representation, pattern information, and brain signatures: From neurons to neuroimaging

    Neuron 99:257–273.

    https://doi.org/10.1016/j.neuron.2018.06.009

    • PubMed
    • Google Scholar
15. 1. Lorenz R
    2. Hampshire A
    3. Leech R

    (2017) Neuroadaptive bayesian optimization and hypothesis testing

    Trends in Cognitive Sciences 21:155–167.

    https://doi.org/10.1016/j.tics.2017.01.006

    • PubMed
    • Google Scholar
16. Book

    1. Luck SJ
    2. Kappenman ES

    (2011) The Oxford Handbook of Event-Related Potential Components

    Oxford Academic.

    https://doi.org/10.1093/oxfordhb/9780195374148.001.0001

    • Google Scholar
17. 1. Makeig S
    2. Westerfield M
    3. Jung TP
    4. Enghoff S
    5. Townsend J
    6. Courchesne E
    7. Sejnowski TJ

    (2002) Dynamic brain sources of visual evoked responses

    Science 295:690–694.

    https://doi.org/10.1126/science.1066168

    • PubMed
    • Google Scholar
18. 1. Mazaheri A
    2. Jensen O

    (2006) Posterior alpha activity is not phase-reset by visual stimuli

    PNAS 103:2948–2952.

    https://doi.org/10.1073/pnas.0505785103

    • PubMed
    • Google Scholar
19. 1. McIntosh AR
    2. Kovacevic N
    3. Itier RJ

    (2008) Increased brain signal variability accompanies lower behavioral variability in development

    PLOS Computational Biology 4:e1000106.

    https://doi.org/10.1371/journal.pcbi.1000106

    • PubMed
    • Google Scholar
20. 1. Nikulin VV
    2. Linkenkaer-Hansen K
    3. Nolte G
    4. Lemm S
    5. Müller KR
    6. Ilmoniemi RJ
    7. Curio G

    (2007) A novel mechanism for evoked responses in the human brain

    The European Journal of Neuroscience 25:3146–3154.

    https://doi.org/10.1111/j.1460-9568.2007.05553.x

    • PubMed
    • Google Scholar
21. 1. Pfurtscheller G
    2. Lopes da Silva FH

    (1999) Event-related EEG/MEG synchronization and desynchronization: Basic principles

    Clinical Neurophysiology 110:1842–1857.

    https://doi.org/10.1016/s1388-2457(99)00141-8

    • PubMed
    • Google Scholar
22. 1. Sauseng P
    2. Klimesch W
    3. Gruber WR
    4. Hanslmayr S
    5. Freunberger R
    6. Doppelmayr M

    (2007) Are event-related potential components generated by phase resetting of brain oscillations? A critical discussion

    Neuroscience 146:1435–1444.

    https://doi.org/10.1016/j.neuroscience.2007.03.014

    • PubMed
    • Google Scholar
23. 1. Shah AS
    2. Bressler SL
    3. Knuth KH
    4. Ding M
    5. Mehta AD
    6. Ulbert I
    7. Schroeder CE

    (2004) Neural dynamics and the fundamental mechanisms of event-related brain potentials

    Cerebral Cortex 14:476–483.

    https://doi.org/10.1093/cercor/bhh009

    • PubMed
    • Google Scholar
24. 1. Stein RB
    2. Gossen ER
    3. Jones KE

    (2005) Neuronal variability: Noise or part of the signal?

    Nature Reviews. Neuroscience 6:389–397.

    https://doi.org/10.1038/nrn1668

    • PubMed
    • Google Scholar
25. 1. Stokes MG
    2. Wolff MJ
    3. Spaak E

    (2015) Decoding rich spatial information with high temporal resolution

    Trends in Cognitive Sciences 19:636–638.

    https://doi.org/10.1016/j.tics.2015.08.016

    • PubMed
    • Google Scholar
26. 1. Vidaurre D
    2. Myers NE
    3. Stokes M
    4. Nobre AC
    5. Woolrich MW

    (2019) Temporally unconstrained decoding reveals consistent but time-varying stages of stimulus processing

    Cerebral Cortex 29:863–874.

    https://doi.org/10.1093/cercor/bhy290

    • PubMed
    • Google Scholar
27. 1. Vidaurre D
    2. Cichy RM
    3. Woolrich MW

    (2021) Dissociable components of information encoding in human perception

    Cerebral Cortex 31:5664–5675.

    https://doi.org/10.1093/cercor/bhab189

    • PubMed
    • Google Scholar
28. Software

    1. Vidaurre D

    (2024) Genephys, version swh:1:rev:994179448e380155f997efbfaa3bd04aba4fecfb

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:b835c738cf4bf2057701c138df5287a1ffa2de5c;origin=https://github.com/vidaurre/genephys;visit=swh:1:snp:d3d2d5bd4926f13dd9fd7e5ce9d25175a2f778e2;anchor=swh:1:rev:994179448e380155f997efbfaa3bd04aba4fecfb

Author details

1. Diego Vidaurre

   1. Center for Functionally Integrative Neuroscience, Department of Clinical Medicine, Aarhus University, Aarhus, Denmark
   2. Department of Psychiatry, Oxford University, Oxford, United Kingdom

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing

   For correspondence

   dvidaurre@cfin.au.dk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-9650-2229

• 852

  views

• 59

  downloads

• 1

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.