Deep neural networks (DNNs) have had extraordinary successes in classifying photographic images of objects and are often described as the best models of biological vision. This conclusion is largely based on three sets of findings: (1) DNNs are more accurate than any other model in classifying images taken from various datasets, (2) DNNs do the best job in predicting the pattern of human errors in classifying objects taken from various behavioral datasets, and (3) DNNs do the best job in predicting (...) brain signals in response to images taken from various brain datasets (e.g., single cell responses or fMRI data). However, these behavioral and brain datasets do not test hypotheses regarding what features are contributing to good predictions and we show that the predictions may be mediated by DNNs that share little overlap with biological vision. More problematically, we show that DNNs account for almost no results from psychological research. This contradicts the common claim that DNNs are good, let alone the best, models of human object recognition. We argue that theorists interested in developing biologically plausible models of human vision need to direct their attention to explaining psychological findings. More generally, theorists need to build models that explain the results of experiments that manipulate independent variables designed to test hypotheses rather than compete on making the best predictions. We conclude by briefly summarizing various promising modeling approaches that focus on psychological data. (shrink)

In the last decade, reading research has seen a paradigmatic shift. A new wave of computational models of orthographic processing that offer various forms of noisy position or context-sensitive coding have revolutionized the field of visual word recognition. The influx of such models stems mainly from consistent findings, coming mostly from European languages, regarding an apparent insensitivity of skilled readers to letter order. Underlying the current revolution is the theoretical assumption that the insensitivity of readers to letter order reflects the (...) special way in which the human brain encodes the position of letters in printed words. The present article discusses the theoretical shortcomings and misconceptions of this approach to visual word recognition. A systematic review of data obtained from a variety of languages demonstrates that letter-order insensitivity is neither a general property of the cognitive system nor a property of the brain in encoding letters. Rather, it is a variant and idiosyncratic characteristic of some languages, mostly European, reflecting a strategy of optimizing encoding resources, given the specific structure of words. Since the main goal of reading research is to develop theories that describe the fundamental and invariant phenomena of reading across orthographies, an alternative approach to model visual word recognition is offered. The dimensions of a possible universal model of reading, which outlines the common cognitive operations involved in orthographic processing in all writing systems, are discussed. (shrink)

A fundamental claim associated with parallel distributed processing theories of cognition is that knowledge is coded in a distributed manner in mind and brain. This approach rejects the claim that knowledge is coded in a localist fashion, with words, objects, and simple concepts, that is, coded with their own dedicated representations. One of the putative advantages of this approach is that the theories are biologically plausible. Indeed, advocates of the PDP approach often highlight the close parallels between distributed representations learned (...) in connectionist models and neural coding in brain and often dismiss localist theories as biologically implausible. The author reviews a range a data that strongly challenge this claim and shows that localist models provide a better account of single-cell recording studies. The author also contrast local and alternative distributed coding schemes and argues that common rejection of grandmother cell theories in neuroscience is due to a misunderstanding about how localist models behave. The author concludes that the localist representations embedded in theories of perception and cognition are consistent with neuroscience; biology only calls into question the distributed representations often learned in PDP models. (shrink)

Sibley et al. (2008) report a recurrent neural network model designed to learn wordform representations suitable for written and spoken word identification. The authors claim that their sequence encoder network overcomes a key limitation associated with models that code letters by position (e.g., CAT might be coded as C‐in‐position‐1, A‐in‐position‐2, T‐in‐position‐3). The problem with coding letters by position (slot‐coding) is that it is difficult to generalize knowledge across positions; for example, the overlap between CAT and TOMCAT is lost. Although we (...) agree this is a critical problem with many slot‐coding schemes, we question whether the sequence encoder model addresses this limitation, and we highlight another deficiency of the model. We conclude that alternative theories are more promising. (shrink)

On several key issues we agree with the commentators. Perhaps most importantly, everyone seems to agree that psychology has an important role to play in building better models of human vision, and (most) everyone agrees (including us) that deep neural networks (DNNs) will play an important role in modelling human vision going forward. But there are also disagreements about what models are for, how DNN–human correspondences should be evaluated, the value of alternative modelling approaches, and impact of marketing hype in (...) the literature. In our view, these latter issues are contributing to many unjustified claims regarding DNN–human correspondences in vision and other domains of cognition. We explore all these issues in this response. (shrink)

Two additional sources of evidence are provided in support of localist coding within connectionist networks. First, only models with localist codes can currently represent multiple pieces of information simultaneously or represent order among a set of items on-line. Second, recent priming data appear problematic for theories that rely on distributed representations. However, a faulty argument advanced by Page is also pointed out.

Levelt et al. describe a model of speech production in which lemma access is achieved via input from nondecompositional conceptual representations. They claim that existing decompositional theories are unable to account for lexical retrieval because of the so-called hyperonym problem. However, existing decompositional models have solved a formally equivalent problem.

The project of coordinating perception, comprehension, and motor control is an exciting one, but I found it hard to follow some of Pickering & Garrod's (P&G's) arguments as presented. Consequently, my comment is not so much a disagreement with P&G but a query about the logic of forward models: It is not clear how they are supposed to work, nor why they are needed in this (or many other) contexts, and toward that end I present an alternative idea.

We argue that Bayesian models are best categorized as methodological or theoretical. That is, models are used as tools to constrain theories, with no commitment to the processes that mediate cognition, or models are intended to approximate the underlying algorithmic solutions. We argue that both approaches are flawed, and that the Enlightened Bayesian approach is unlikely to help.

A universal property of visual word identification is position-invariant letter identification, such that the letter is coded in the same way in CAT and ACT. This should provide a fundamental constraint on theories of word identification, and, indeed, it inspired some of the theories that Frost has criticized. I show how the spatial coding scheme of Colin Davis can, in principle, account for contrasting transposed letter priming effects, and at the same time, position-invariant letter identification.

There is widespread agreement in neuroscience and psychology that the visual system identifies objects and faces based on a pattern of activation over many neurons, each neuron being involved in representing many different categories. The hypothesis that the visual system includes finely tuned neurons for specific objects or faces for the sake of identification, so‐called “grandmother cells”, is widely rejected. Here it is argued that the rejection of grandmother cells is premature. Grandmother cells constitute a hypothesis of how familiar visual (...) categories are identified, but the primary evidence against this hypothesis comes from studies that have failed to observe neurons that selectively respond to unfamiliar stimuli. These findings are reviewed and it is shown that they are irrelevant. Neuroscientists need to better understand existing models of face and object identification that include grandmother cells and then compare the selectivity of these units with single neurons responding to stimuli that can be identified. (shrink)

According to Pylyshyn, the early visual system is able to categorize perceptual inputs into shape classes based on visual similarity criteria; it is also suggested that written words may be categorized within early vision. This speculation is contradicted by the fact that visually unrelated exemplars of a given letter (e.g., a/A) or word (e.g., read/READ) map onto common visual categories.