Figure 1.
Detecting two types of brain dynamics by assessing the ability of multivariate pattern classifiers to generalize across time.
The temporal generalization method can characterize the dynamics of neural activity. (left) When the stimulus evokes a serial chain of brain activations, “diagonal classifiers”, trained and tested at each time point can extract stimulus information throughout the activation period. However, as each classifier is specific to the time point at which it has been trained, they cannot generalize across other time samples. The generalization time analysis thus reveals a diagonal generalization matrix. (right) By contrast, if the underlying activity is sustained over time, then all classifiers would capture the same pattern. These classifiers would thus generalize to one another and lead to a square generalization matrix.
Figure 2.
Violating two types of auditory regularities.
The Local-Global experimental design [7] is a variation of the auditory oddball task. It consists in presenting series of 5-sound sequences which are composed of five identical sounds (local standard) or four identical sounds followed by a deviant one (local deviant). The global regularity is established across trials by making 80% of the trials identical (global standard). The design thus dissociates the violation of local predictions (change of sound in a given trial) and global predictions (change of sequence across trials).
Figure 3.
Generalization across time of the local and global responses to auditory novelty.
At each time point, a classifier was trained to extract the pattern of MEG activity that distinguishes local-standard from local-deviant trials (mismatch effect, top) or to contrast global-standard from global-deviant trials (bottom). Each classifier was subsequently tested on its ability to generalize this discrimination to all other time samples. (left) Differential patterns (standard – deviant) of brain activity across subjects as well as in a single representative subject using classic univariate analyses. For simplicity purposes, only the magnetometers are plotted (n = 102/306 channels). Note that, unlike subject-specific decoding, classic event related fields (ERF) analyses are tested across subjects, and are thus insensitive to inter-individual variability of subjects' topographies. (middle) Generalization of six different classifiers trained at regularly spaced times between 100 ms and 600 ms (purple), compared to the traditional “diagonal” decoding method where a different classifier is trained and tested at the same time point (black). The thick lines indicate significant decoding scores. The yellow areas indicate when the diagonal performance was significantly different from the generalization across time. Error bars indicate the standard error of the mean (SEM) across subjects. (right) Generalization matrices. Decoding performance is plotted as a function of training time (vertical axis) and testing time (horizontal axis) for all classifiers. Decoding of the local-violation effect leads to a diagonal-shaped decoding performance from 82 ms to 508 ms (AUC over 50% in red), demonstrating that each classifier was only able to predict trials' classes for a short amount of time. Decoding of the global-violation effect leads to a square generalization matrix, suggesting that the underlying brain activity is essentially stable during this time period. Early classifiers of the global violation (<350 ms) are slightly lower than the traditional “diagonal” decoding performance, thus suggesting only a small change in the underlying pattern of activity.