Norris, McQueen & Cutler provide two possible explanations for neighborhood effects. The first suggests that nonwords that are more similar to words tend to activate those words more than do less similar nonwords, and the second is based on sequential probabilities between phonemes. Unfortunately, neither explanation is sufficient to explain all reported neighborhood effects.

The case for the independence of featural processing supports Merge and FLMP alike. The target article's criticisms of the latter model are founded on misunderstanding its application to natural language processing. In fact, the main difference in the functional properties of the models is the FLMP's ability to account for graded perceptual experience. @T3:It is startling to be informed that beliefs you have held for a quarter of a century are indefensible and implausible. So it comes as a relief to discover that it was all just a big misunderstanding.

Norris et al. present a sufficiency case for Merge, but not for autonomy. The simulations make clear that there is little reason to favor Merge over TRACE. The slanted presentation of the empirical evidence gives the illusion that the autonomous position is stronger than it really is.

We urge caution in generalising from content words to function words, in which lexical-to-phonemic feedback might be more likely. Speech perception involves more than word recognition; feedback might be outside the narrow logic of word identification but still be present for other purposes. Finally, we raise the issue of evidence from imaging studies of auditory hallucination.

Norris, McQueen & Cutler claim that Merge is an autonomous model, superior to the interactive TRACE model and the autonomous Race model. Merge is actually an interactive model, despite claims to the contrary. The presentation of the literature seriously distorts many findings, in order to advocate autonomy. It is Merge's interactivity that allows it to simulate findings in the literature.

Hesitations about accepting whole-heartedly Norris et al.'s suggestion to abandon feedback in speech processing models concern (1) whether accounting for all available data justifies additional layers of complexity in the model and (2) whether characterizing Merge as non- interactive is valid. Spoken word recognition studies support the nature of Merge's lexical level and suggest that phonemes should comprise the prelexical level.

The proposition that feedback is never necessary in speech recognition is examined for utterances consisting of sequences of words. In running speech the features near word boundaries are often modified according to language-dependent rules. Application of these rules during word recognition requires top-down processing. Because isolated words are not usually modified by rules, their recognition could be achieved by bottom-up processing only.

Norris et al.'s claim that feedback is unnecessary is compromised by (1) a questionable application of Occam's razor, given strong evidence for feedback in perception; (2) an idealization of the speech recognition problem that simplifies those aspects of the input that create conditions where feedback is useful; (3) Norris et al.'s use of decision nodes that incorporate feedback to model some important empirical results; and (4) problematic linking hypotheses between crucial simulations and behavioral data.

Norris, McQueen & Cutler present a detailed account of the decision stage of the phoneme monitoring task. However, we question whether this contributes to our understanding of the speech recognition process itself, and we fail to see why phonotactic knowledge is playing a role in phoneme recognition.

I agree with Norris et al. that feedback to a phonemic level is never necessary, but disagree strongly with their reason why this is true. I believe the available evidence indicates that there is no feedback because there is no phonemic level employed in the perceptual processing of speech.

Although Norris, McQueen & Cutler have provided convincing evidence that there is no need for contributions from the lexicon to phonetic processing, their simplification of the communication between levels comes at a cost to the processes themselves. Although their arrangement may ultimately prove correct, its validity is not due to a successful application of Occam's razor.

Models are not adequately evaluated simply by whether they capture the data, after the fact. Other criteria are needed. One criterion is parsimony; but utility and generality are at least as important. Even with respect to parsimony, however, the case against feedback is not as straightforward as Norris et al. present it. We use feedback consistency effects to illustrate these points.

The central thesis of our target article is that feedback is never necessary in spoken word recognition. In this response we begin by clarifying some terminological issues that have led to a number of misunderstandings. We provide some new arguments that the feedforward model Merge is indeed more parsimonious than the interactive alternatives, and that it provides a more convincing account of the data than alternative models. Finally, we extend the arguments to deal with new issues raised by the commentators such as infant speech perception and neural architecture.

A general conceptual framework for large-scale neocortical dynamics based on data from many laboratories is applied to a variety of experimental designs, spatial scales, and brain states. Partly distinct, but interacting local processes (e.g., neural networks) arise from functional segregation. Global processes arise from functional integration and can facilitate (top down) synchronous activity in remote cell groups that function simultaneously at several different spatial scales. Simultaneous local processes may help drive (bottom up) macroscopic global dynamics observed with electroencephalography (EEG) or magnetoencephalography (MEG).

A local/global dynamic theory that is consistent with EEG data and the proposed conceptual framework is outlined. This theory is neutral about properties of neural networks embedded in macroscopic fields, but its global component makes several qualitative and semiquantitative predictions about EEG measures of traveling and standing wave phenomena. A more general “metatheory” suggests what large-scale quantitative theories of neocortical dynamics may be like when more accurate treatment of local and nonlinear effects is achieved.

The theory describes the dynamics of excitatory and inhibitory synaptic action fields. EEG and MEG provide large-scale estimates of modulation of these synaptic fields around background levels. Brain states are determined by neuromodulatory control parameters. Purely local states are dominated by local feedback gains and rise and decay times of postsynaptic potentials. Dominant local frequencies vary with brain region. Other states are purely global, with moderate to high coherence over large distances. Multiple global mode frequencies arise from a combination of delays in corticocortical axons and neocortical boundary conditions. Global frequencies are identical in all cortical regions, but most states involve dynamic interactions between local networks and the global system. EEG frequencies may involve a “matching” of local resonant frequencies with one or more of the many, closely spaced global frequencies.

The EEG activity recorded from the human sensorimotor cortical area exhibits rhythmic activity covering a broad range of frequencies, including alpha, mu, beta, and gamma (40-Hz) rhythms. This commentary elaborates on connections between these sensorimotor rhythms and Nunez's neocortical dynamic theory.

A critical discussion is provided of three central assumptions underlying Nunez's approach to modeling cortical activity. A plea is made for neurophysiologically realistic models involving nonlinearities, multiple time scales, and stochasticity.

A unifing theory of spatiotemporal brain dynamics should incorporate multiple spatial and temporal scales. Between the microscopic (local) and macroscopic (global) components proposed by Nunez, mesoscopic (intermediate-range) elements should be integral parts of models. The corresponding mathematical formalism requires tools of nonlinear dynamics and the use of aperiodic (chaotic) attractors. Some relations between local-mesoscopic and mesoscopic-global components are outlined.

The spatial time-frequency distribution matrix and associated Rényi entropy is proposed as the basis for a method that may be useful for estimating the significance of nonlocal neocortical interactions in the analysis of scalp EEG data. Implications of nonlocal interactions for source estimation are also considered.

Current EEG research emphasizes gamma band coherence as a signature of functional integration, that is, the solution to the binding problem. We note that spatial patterns of coherent neural activity are also observed at other EEG frequencies. If these oscillations reflect Nunez's resonant modes, they offer a solution to the binding problem that emerges naturally from the architecture of cortical connections.

This commentary focuses on how bottom-up neocortical models can be developed into eigenfunction expansions of probability distributions appropriate to describe short-term memory in the context of scalp EEG. The mathematics of eigenfunctions are similar to the top-down eigenfunctions developed by Nunez, despite different physical manifestations. The bottom-up eigenfunctions are at the local mesocolumnar scale, whereas the top-down eigenfunctions are at the global regional scale. Our respective approaches have regions of substantial overlap, and future studies may expand top-down eigenfunctions into the bottom-up eigenfunctions, yielding a model of scalp EEG expressed in terms of columnar states of neocortical processing of attention and short-term memory.