Neural computations are often anatomically localized and executed on sub-second time scales. Understanding the brain therefore requires methods that offer sufficient spatial and temporal resolution. This poses a particular challenge for the study of the human brain because non-invasive methods have either high temporal or spatial resolution, but not both. Here, we introduce a novel multivariate analysis method for conventional blood-oxygen-level dependent functional magnetic resonance imaging (BOLD fMRI) that allows to study sequentially activated neural patterns separated by less than 100 ms with anatomical precision. Human participants underwent fMRI and were presented with sequences of visual stimuli separated by 32 to 2048 ms. Probabilistic pattern classifiers were trained on fMRI data to detect the presence of image-specific activation patterns in early visual and ventral temporal cortex. The classifiers were then applied to data recorded during sequences of the same images presented at increasing speeds. Our results show that probabilistic classifier time courses allowed to detect neural representations and their order, even when images were separated by only 32 ms. Moreover, the frequency spectrum of the statistical sequentiality metric distinguished between sequence speeds on sub-second versus supra-second time scales. These results survived when data with high levels of noise and rare sequence events at unknown times were analyzed. Our method promises to lay the groundwork for novel investigations of fast neural computations in the human brain, such as hippocampal replay.