Neurophysiology of Active Perception

====== A Brief History ====== I have been active in a number of areas that, at first sight, appear to be unrelated: psychometrics, visual word recognition, statistical testing of electrophysiological data, and cognitive neurophysiology. On a closer look, one will see that there are clear links between these areas. My PhD project, supervised by Paul De Boeck, was in psychometrics. My thesis was about probabilistic modelling of behavioral responses that come in the form of a subjects-by-stimuli/items matrix. I developed methods for uncovering the cognitive processes behind the observed responses and provided illustrative applications. After my PhD, I decided to contribute to a better understanding of cognitive processes, and not only develop methods to do so. I investigated visual word recognition, which is about the reading of single words. In this field, it is assumed that two processes are involved, a phonological process that performs the mapping of letter strings to some phonological code, and an orthographic process that performs the mapping of letter strings to some semantic code. I attempted to identify these two processes by means of multinomial processing tree models, which are probabilistic models of the same type as the ones that I developed during my PhD project. I have always been convinced that some form of modelling is essential for our understanding of behavior. Personally, I strongly prefer probabilistic modelling over the computer models that are popular in several fields, such as visual word recognition. However, I became convinced that neither of the two modelling approaches will be able to uncover the mechanisms via which behavior is produced, at least not when they are applied to behavioral data only. The reason is that these behavioral data (typically, one accuracy score and/or one response time per trial) do not provide enough information to uncover these mechanisms. In principle, one may already be satisfied if a quantitative model can be constructed that produces the required behavior, as this provides us with at least one way to understand this behavior. However, this level of understanding cannot compete with the achievements in the natural and the life sciences. Here, the models take the form of mechanistic explanations of the observed phenomena (e.g., light absorption and reflection, cell division, protein synthesis, fluctuations in cell membrane potentials, etc.). Successful mechanistic explanations are consistent with the very rich observations that are obtained in organic and anorganic preparations, including cell and tissue cultures. The availability of a rich platform of observations (different types of spectra, counts based on microscope observations, counts of different types of metabolic residues, etc.) explains the main difference between models based on behavioral data and models (mechanistic explanations) based on biological data: the latter are subjected to much more empirical constraints. I am convinced that, to increase our understanding of behavior and cognition, we must link it to the underlying neurophysiological processes. By recording brain activity while the organism is engaged in a particular task, we obtain the rich information that is required to formulate and validate mechanistic explanations of behavior and cognition. This type of research is typically performed using electrophysiological measurements, which are electrical potentials recorded on the scalp (EEG), up or directly under the dura mater (ECoG/iEEG), or inside the neuropil (using wire recordings). My first project in electrophysiology was about the statistical testing of differences between experimental conditions. This is a serious statistical problem that originates from the fact that electrophysiological data are observed as spatiotemporal matrices, with the channels forming the spatial dimension, and the discrete time points (as determined by the sampling rate) forming the temporal dimension. This creates a huge multiple comparisons problem that is even further exacerbated after a time-resolved spectral analysis, which adds a frequency dimension to the spatiotemporal matrix. Together with my co-authors, I have shown that cluster-based permutation tests are an excellent solution for this statistical problem.