genephys configuration hyperparameters.

Left and middle: single-trial example of the generated signal (in blue) and its constitutive components (in red): instantaneous phase, frequency and amplitude, as well as an additive non-oscillatory and oscillatory response components; the left panel reflects a phase resetting plus frequency entrainment effect, while the middle panel corresponds to an additive oscillatory response. Right: real (filtered) magnetoencephalography data collected during passive stimuli viewing; the red curves are the analytically computed phase, frequency and amplitude (via the Hilbert transform).

Example of an empirical TGM, where the different characteristic features have been highlighted.

Examples of two configurations based on phase-resetting and frequency entrainment (left) and additive oscillatory responses (middle), shown together with results obtained from real data (right). The top panels show temporal generalisation matrices (TGM) from standard decoding analysis. The bottom panels show average evoked responses (ERP/F) for the sampled signal (blue) and its components (red); the two stimuli are represented with different tonalities of blue or red.

Having more dimensions (channels) carrying stimulus-specific information, but with a larger degree of stochasticity, produces more realistic decoding patterns than having fewer dimensions with a lower degree of stochasticity. Here stochasticity referred to the channels’ probability of activation.

Oscillatory responses to stimulation occur in a diversity of frequencies and latencies across channels. Top row, single latency of response; bottom row, diverse latencies across channels. Left column, single frequency; right column, diverse frequencies across channels. Bottom-right is the most realistic TGM.

An additive non-oscillatory response is needed to produce realistic TGMs with a broadening of decoding accuracy at later stages of the trial. A: By increasing the strength of the non-oscillatory response, the broadening of accuracy becomes more prominent. B: Changing the nature of the phasic relationship between the slower and the faster (oscillatory component) greatly influences the TGM, from having all channels in-phase (left) towards having all channels anti-phase (right).

Enhancing the amplitude of the stimulus-relevant oscillation is not strictly necessary to produce realistic TGMs, but it can enlarge the effects. Two features of the TGM are highlighted: the diagonal, and a vertical slice at the time point of maximum accuracy.

Correlation between an empirical TGM and the sampled TGM (averaged across ten runs of the simulation) across a range of configurations of the response function, which defines the temporal dynamics of the effect. The maximum correlation with the empirical TGM is slightly over 0.7

TGMs from six real data example subjects performing a passive viewing paradigm

Phase reset plus frequency entrainment effect. Top row, single latency of response; bottom row, diverse latencies across channels; left column, single frequency; right column; diverse frequencies across channels.