Data folding, i.e., division of data into training and testing sets, ensured that generalization testing was done on data that were not used for hyperalignment or classifier training (Kriegeskorte et al., 2009). Palbociclib nmr BSC of the face and object categories reached a maximal level with the top 12 PCs from the PCA of the face and object data (67.7% ± 2.1%). BSC of the animal species
reached a maximal level with the top nine PCs from the PCA of the animal species data (73.9% ± 3.0%). The top PCs from the face and object data, however, did not afford good classification of the animal species (55.0% ± 3.4%) or the movie time segments (50.1% ± 2.7%), nor did the top PCs from the animal species data afford good classification of the face and object categories (54.2% ± 2.6%) or the movie time segments (49.5% ± 2.6%; Figure 3B). Thus, the lower-dimensional representational spaces for the limited number of stimulus categories in the face and object experiment and in the animal species experiment
are different from each other and are of less general validity than the higher-dimensional movie-based common model space. We next asked whether a complex, natural stimulus, such as the movie, is necessary to derive hyperalignment parameters that generate a common space with general validity across a wide range of complex visual stimuli. find more In principle, a common space and hyperalignment parameters can be derived from any fMRI time series. We investigated whether hyperalignment
of the face and object data and hyperalignment of the animal species data would afford high levels of BSC accuracy using only the data from those experiments. In each experiment, we derived a common space based on all runs but one. We transformed the data from all runs, including the left-out run, into this common space. We trained the classifier on those runs used for hyperalignment in all subjects but one and tested the classifier on the data from the left-out run in the left-out subject. Thus, the test data for determining classifier accuracy played no role either in hyperalignment or in classifier isothipendyl training (Kriegeskorte et al., 2009). BSC of face and object categories after hyperalignment based on data from that experiment was equivalent to BSC after movie-based hyperalignment (62.9% ± 2.9% versus 63.9% ± 2.2%, respectively; Figure 4). Surprisingly, BSC of the animal species after hyperalignment based on data from that experiment was significantly better than BSC after movie-based hyperalignment (76.2% ± 3.7% versus 68.0% ± 2.8%, respectively; p < 0.05; Figure 4). This result suggests that the validity for a model of a specific subspace may be enhanced by designing a stimulus paradigm that samples the brain states in that subspace more extensively. We next asked whether hyperalignment based on these simpler stimulus sets was sufficient to derive a common space with general validity across a wider array of complex stimuli.