Four common conceptual fallacies in mapping the time course of recognition

Abstract

Determining the moment at which a visual recognition process is completed, or the order in which various processes come into play, are fundamental steps in any attempt to understand human recognition abilities, or to replicate the corresponding hierarchy of neuronal mechanisms within artificial systems. Common experimental paradigms for addressing these questions involve the measurement and/or comparison of backward-masking (or rapid serial visual presentation) psychometric functions and of physiological EEG/MEG/LFP signals (peak latencies, differential activities, single-trial decoding techniques). I review and illustrate four common mistakes that scientists tend to make when using these paradigms, and explain the conceptual fallacies that motivate their reasoning. First, contrary to collective intuition, presentation times, or stimulus-onset asynchrony masking thresholds cannot be taken to reflect, directly or indirectly, the timing of relevant brain processes. Second, psychophysical or electrophysiological measurements should not be compared without assessing potential physical differences between experimental stimulus sets. Third, such comparisons should not be performed in any manner contingent on subjective responses, so as to avoid response biases. Last, the filtering of electrophysiological signals alters their temporal structure, and thus precludes their interpretation in terms of time course. Practical solutions are proposed to overcome these common mistakes.

Keywords: EEG; backward-masking; low-level image properties; methodological and conceptual mistakes; rapid serial visual presentation; response bias; signal filtering; visual timing.

Figure 1

Psychometric functions (hypothetical) for a typical RSVP or masking experiment. The x-axis would represent presentation time of each image in the sequence, or stimulus-mask SOA (respectively). The natural tendency to read from this graph that processes A and B have similar speeds whereas process C is 50 ms slower is inaccurate. See text for explanation.

Figure 2

Low-level differences exist between image classes in the pixel and in the Fourier domains. These confounds are systematic enough to allow distinguishing between images containing animals or cars, for example. As can be seen in the two columns on the right, the average of 124 pictures of animals (including mammals, birds, reptiles, fish) itself resembles (a “hazy” view of) an animal, whereas the average of 124 cars is much more similar to a car picture. The same can be said of averages in the Fourier domain (bottom). A simple “detector” or “classifier” using the two patterns on the right as templates would easily distinguish between the two examples on the left, without any need for feature, object, or category representations.

Figure 3

Taking into account subjective responses when contrasting brain signals can generate spurious differences. 1000 trials were simulated, half of them with a stimulus from category A and the other half from category B. For the purpose of demonstrating the existence of response biases, each EEG waveform was drawn randomly with a 1/f power spectrum, thus approximating the statistics of natural EEG but without any selective response evoked by category A or B. Hence, we should normally expect to obtain flat event-related potentials (ERPs, computed by averaging all trials of each category, with each trial baseline-corrected using the period [−50, +50 ms] around stimulus onset). We assume that the observer correctly categorizes 75% of all trials, but on the remaining 25%, the observer decides to respond not on the basis of the stimulus category (A or B), but based on EEG activity during the 200-ms pre-stimulus period: when the average pre-stimulus activity is positive, the observer presses response button A, and button B when the activity is negative. Of course, negative or positive pre-stimulus activity is equally likely to occur regardless of stimulus type, and therefore the ERPs obtained for stimulus categories A (in blue) and B (in red) will be statistically indistinguishable (bottom panel). However, when only the correct trials are included in the ERPs, many trials with negative pre-stimulus activity and response A and many trials with positive pre-stimulus activity and response B will be discarded from the ERPs (because they correspond to “incorrect” categorization). As a result, the ERPs will show a purely artifactual but significant difference during the pre-stimulus period (top panel).

Figure 4

The dangers of filtering. 50 trials of a “fake” EEG signal were simulated. Activity is null until the “onset” of a neural process, occurring at a random time between 150 and 180 ms (uniform distribution), after which activity is set to 1. Gaussian-distributed noise (mean 0 and SD of 0.05) is added to all signals. In the top panel, the original trials are stacked vertically and the EEG amplitude is color-coded. The middle panel represents the same trials after low-pass filtering with a 30-Hz cut-off (using the function eegfilt from the EEGLAB software, and its default parameters). The bottom panel illustrates the corresponding ERPs. The red * symbols on the horizontal axis indicate the moments at which the filtered ERPs depart from zero. Even though, by design, the process under study never started before 150 ms, its EEG correlates are detected with latencies as early as 100 ms!