Building and validating models of skill acquisition that explain speedup effects has been limited by difficulty dis-tinguishing quickly executed cognitive processes (e.g. Anderson, 1982; Logan, 1988; Rickard, 1997). In this experiment,magnetoencephalography (MEG) data are collected from participants solving a repeated math problem set. We use MEG signalto test the three-phase model of skill acquisition that describes the transition from problem-solving strategies of computation,to retrieval, to an automatic stimulus-response process (Fitts & Posner, 1967). We hypothesize that the processes of familiarityand recollection are early features that distinguish the three phases of skill acquisition. Analyzing event-related fields, we testtwo predictions. First, early frontal activation (akin to the FN400 old-new effect of ERP studies) should diminish in strengthwith each successive phase transition. Second, parietal activation (corresponding to the ERP P600 old-new effect) should bepresent in the second phase, but not in the first or last phase.

The current study manipulated how frequently different prob-lems were practiced during a first day of practice, with themore frequent items being more closely spaced. Fitting thedata to a skill acquisition model, we find that greater spac-ing between items is associated with an increased probabilityof transitioning to more efficient phases of performance, butwith a shallower speedup within each phase. Three days aftertraining, we find that performance is predicted not by the prac-tice frequency during training, but rather by the phase of skillacquisition attained during training. Thus, it is type of pro-cessing achieved not the amount and spacing of practice, thatdetermines retention. Spacing, however, promotes learning bydriving changes in cognitive processing.

A key challenge for cognitive neuroscience is to decipher therepresentational schemes of the brain. A recent class of decodingalgorithms for fMRI data, stimulus-feature-based encodingmodels, is becoming increasingly popular for inferring thedimensions of neural representational spaces from stimulus-feature spaces. We argue that such inferences are not always valid,because decoding can occur even if the neural representationalspace and the stimulus-feature space use different representationalschemes. This can happen when there is a systematic mappingbetween them. In a simulation, we successfully decoded the binaryrepresentation of numbers from their decimal features. Sincebinary and decimal number systems use different representations,we cannot conclude that the binary representation encodes decimalfeatures. The same argument applies to the decoding of neuralpatterns from stimulus-feature spaces and we urge caution ininferring the nature of the neural code from such methods. Wediscuss ways to overcome these inferential limitations.