Grandmother Cells Research Articles

Statistical mechanics of learning via reverberation in bidirectional associative memories

We study bi-directional associative neural networks that, exposed to noisy examples of an extensive number of random archetypes, learn the latter (with or without the presence of a teacher) when the supplied information is enough: in this setting, learning is heteroassociative – involving couples of patterns – and it is achieved by reverberating the information depicted from the examples through the layers of the network. By adapting Guerra’s interpolation technique, we provide a full statistical mechanical picture of supervised and unsupervised learning processes (at the replica symmetric level of description) obtaining analytically phase diagrams, thresholds for learning, a picture of the ground-state in plain agreement with Monte Carlo simulations and signal-to-noise outcomes. In the large dataset limit, the Kosko storage prescription as well as its statistical mechanical picture provided by Kurchan, Peliti, and Saber in the eighties is fully recovered. Computational advantages in dealing with information reverberation, rather than storage, are discussed for natural test cases. In particular, we show how this network admits an integral representation in terms of two coupled restricted Boltzmann machines, whose hidden layers are entirely built of by grand-mother neurons, to prove that by coupling solely these grand-mother neurons we can correlate the patterns they are related to: it is thus possible to recover Pavlov’s Classical Conditioning by adding just one synapse among the correct grand-mother neurons (hence saving an extensive number of these links for further information storage w.r.t. the classical auto-associative setting). Likewise, it is enough to equip the BAM’s synaptic coupling with its transpose matrix (and alternate among these two, e.g., by introducing a clock) to let these networks be able count sequences of patterns (still preserving Detailed Balance, at difference w.r.t. the auto-associative case).

ClueReader: Heterogeneous Graph Attention Network for Multi-Hop Machine Reading Comprehension

Multi-hop machine reading comprehension is a challenging task in natural language processing as it requires more reasoning ability across multiple documents. Spectral models based on graph convolutional networks have shown good inferring abilities and lead to competitive results. However, the analysis and reasoning of some are inconsistent with those of humans. Inspired by the concept of grandmother cells in cognitive neuroscience, we propose a heterogeneous graph attention network model named ClueReader to imitate the grandmother cell concept. The model is designed to assemble the semantic features in multi-level representations and automatically concentrate or alleviate information for reasoning through the attention mechanism. The name ClueReader is a metaphor for the pattern of the model: it regards the subjects of queries as the starting points of clues, takes the reasoning entities as bridge points, considers the latent candidate entities as grandmother cells, and the clues end up in candidate entities. The proposed model enables the visualization of the reasoning graph, making it possible to analyze the importance of edges connecting entities and the selectivity in the mention and candidate nodes, which is easier to comprehend empirically. Evaluations on the open-domain multi-hop reading dataset WikiHop and drug–drug interaction dataset MedHop proved the validity of ClueReader and showed the feasibility of its application of the model in the molecular biology domain.

Sparse ensemble neural code for a complete vocal repertoire.

The categorization of animal vocalizations into distinct behaviorally relevant groups for communication is an essential operation that must be performed by the auditory system. This auditory object recognition is a difficult task that requires selectivity to the group identifying acoustic features and invariance to renditions within each group. We find that small ensembles of auditory neurons in the forebrain of a social songbird can code the bird's entire vocal repertoire (∼10 call types). Ensemble neural discrimination is not, however, correlated with single unit selectivity, but instead with how well the joint single unit tunings to characteristic spectro-temporal modulations span the acoustic subspace optimized for the discrimination of call types. Thus, akin to face recognition in the visual system, call type recognition in the auditory system is based on a sparse code representing a small number of high-level features and not on highly selective grandmother neurons.

Are there any ‘object detectors’ in the hidden layers of CNNs trained to identify objects or scenes?

Various methods of measuring unit selectivity have been developed with the aim of better understanding how neural networks work. But the different measures provide divergent estimates of selectivity, and this has led to different conclusions regarding the conditions in which selective object representations are learned and the functional relevance of these representations. In an attempt to better characterize object selectivity, we undertake a comparison of various selectivity measures on a large set of units in AlexNet, including localist selectivity, precision, class-conditional mean activity selectivity (CCMAS), the human interpretation of activation maximization (AM) images, and standard signal-detection measures. We find that the different measures provide different estimates of object selectivity, with precision and CCMAS measures providing misleadingly high estimates. Indeed, the most selective units had a poor hit-rate or a high false-alarm rate (or both) in object classification, making them poor object detectors. We fail to find any units that are even remotely as selective as the ‘grandmother cell’ units reported in recurrent neural networks. In order to generalize these results, we compared selectivity measures on units in VGG-16 and GoogLeNet trained on the ImageNet or Places-365 datasets that have been described as ‘object detectors’. Again, we find poor hit-rates and high false-alarm rates for object classification. We conclude that signal-detection measures provide a better assessment of single-unit selectivity compared to common alternative approaches, and that deep convolutional networks of image classification do not learn object detectors in their hidden layers.

Commentary: The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience.

Commentary: The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience.

The Value of Failure in Science: The Story of Grandmother Cells in Neuroscience.

The annals of science are filled with successes. Only in footnotes do we hear about the failures, the cul-de-sacs, and the forgotten ideas. Failure is how research advances. Yet it hardly features in theoretical perspectives on science. That is a mistake. Failures, whether clear-cut or ambiguous, are heuristically fruitful in their own right. Thinking about failure questions our measures of success, including the conceptual foundations of current practice, that can only be transient in an experimental context. This article advances the heuristics of failure analysis, meaning the explicit treatment of certain ideas or models as failures. The value of failures qua being a failure is illustrated with the example of grandmother cells; the contested idea of a hypothetical neuron that encodes a highly specific but complex stimulus, such as the image of one’s grandmother. Repeatedly evoked in popular science and maintained in textbooks, there is sufficient reason to critically review the theoretical and empirical background of this idea.

Researchers Keep Rejecting Grandmother Cells after Running the Wrong Experiments: The Issue Is How Familiar Stimuli Are Identified.

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.

Approximating Bayesian inference with a sparse distributed memory system

Approximating Bayesian inference with a sparse distributed memory system

Neglected but Exciting Concepts in Developmental and Neurobiological Psychology

This review provides an evaluative overview of five concepts specific to developmental and neurobiological psychology that are found to be largely overlooked in current textbooks. A sample of 19 introductory psychology texts was surveyed to develop a list, including glial cell signaling, grandmother cells, memory reconsolidation, brain plasticity, and moral judgements by infants. These topics are relatively new, have proven to be of high impact in their respective fields, but are rarely discussed in psychological textbooks or by instructors in the classroom. In addition to a brief, but detailed background on each of the concepts, potential textbook chapters and classroom topics that would benefit from a discussion of these concepts are identified. Finally, this review briefly addresses possible ways for textbook authors to incorporate these new topics in future editions of texts without drastically increasing the overall length of the text.

Grandmother cells and localist representations: a review of current thinking

ABSTRACTThere is now a large literature in neuroscience highlighting how some neurons respond highly selectively to high-level information (e.g. cells that respond to specific faces) and a growing literature in psychology and computer science showing that artificial neural networks often learn highly selective representations. Nevertheless, the vast majority of neuroscientists reject “grandmother cell” theories out of hand, and many psychologists reject localist models based on neuroscience. In this review, I detail some of the conceptual confusions regarding grandmother cells that have contributed to this state of affairs, and review the literature of single-unit recording studies in artificial neural networks that may provide insights into why some neurons respond in a highly selective manner. I then briefly review the contributions from leading theorists in psychology and neuroscience. My hope this special issue contributes to a more productive debate on an important issue that has often been characterised by misunderstandings between disciplines.

Comparing single-unit recordings taken from a localist model to single-cell recording data: a good match

ABSTRACTSingle-cell recording studies show that some neurons respond to complex visual information (e.g. words, objects, faces) in a highly selective manner, with individual neurons responding to about 0.5% of presented images. Such data have often been taken as inconsistent with “grandmother cell” theories as well as with localist models in psychology. In particular, it is commonly assumed that units in localist models respond to only one input, resulting in greater levels of selectivity than seen in single-cell results. To test this assumption, we recorded unit activity from a localist model of word identification. Our results show that the model can capture the levels of selectivity reported in neuroscience. Accordingly, single-cell data do not rule out localist coding schemes. We propose that the term grandmother cell should be reserved for the hypothesis that the brain implements localist representations: neurons that represent one and only one thing but respond to multiple things.

Transient localist representations in critical branching networks

ABSTRACTThere is a long-standing debate between localist and distributed representations in neural network research. Different connectionist models have employed one or the other type of representation, but all such models assume that representations are either always stable, or stable after learning converges. The present article explores an alternate possibility, which is that representations continually shift and change, even on relatively fast timescales, and even after learning has stabilised. This possibility of transient representation is demonstrated in a spiking neural network model, along with supporting neuroscientific evidence. The model learns localist representations that change over a wide range of timescales. Representations are made transient because synaptic connections are continually enabled and disabled to regulate spike propagation to a homeostasis of critical branching. Experiments and models suggest that so-called grandmother cells may shift on relatively fast timescales between different mental representations in different contexts and conditions.

Grandmother cells: much ado about nothing

ABSTRACTWe do not dispute the possibility of the existence in the brain of “grandmother cells”, which are very finely tuned neurons that fire only in the presence of specific objects or categories. However, we question the causal efficacy of such neurons at the functional or behaviour level. We claim that, even though very familiar items, such as “my grandmother”, may well have associated grandmother neurons, these neurons have very little, or no impact on the actual recognition of my grandmother. A study by Thomas, Van Hulle, and Vogels [(2002). Encoding of categories by noncategory-specific neurons in the inferior temporal cortex. Journal of Cognitive Neuroscience, 13, 190–200. doi:10.1162/089892901564252] found finely tuned, category-specific neurons in the inferior temporal cortex of monkeys, but also found that when these neurons were removed from their analysis, this had no effect on categorisation performance. Further, we have found no reported cases of the loss of recognition of single, highly familiar objects, which also argues for a lack of causal efficacy of grandmother-cell neurons.

Grandmother cohorts: multiple-scale brain compression dynamics during learning of object and sequence categories

ABSTRACTThis article summarises neural models of how invariant object categories are learned under conditions where eye movements freely scan a scene, and of how list categories, or chunks, are learned in response to sequences, or lists, of events that are temporarily stored in working memory. During both kinds of category learning, distributed information is compressed by a category learning process into compact groups of cells that can selectively respond to this information. Adaptive Resonance Theory models describe these learning processes and enable, in a limiting case, winner-take-all grandmother cells to be learned, even in a single trial, and rapidly activated during recall trials by bottom-up direct access. The degree of compression can vary with the amount of contextual evidence, and a distributed network of multiple categories may be needed to represent an object or event, all computed using localist mechanisms. Such a network is called a grandmother cohort in the current article.

Grandmother cells and the distinction between local and distributed representation

ABSTRACTA grandmother cell is a neural or cognitive unit that selectively responds to just one particular physical object (one’s maternal grandmother, for example, or one’s own right hand, or one’s own car). Grandmother cells should be distinguished from gnostic units, which are neural or cognitive units that respond selectively to all members of categories of physical objects (all elderly ladies, for example, or all hands, or all cars). Both grandmother cells and gnostic units count as local representations, but not all local representations count as grandmother cells.

Depression, Ketamine Treatment, and the Theater of Consciousness

In severe depression thoughts of suicide can develop over weeks or months. Traditional pharmacologic or psychotherapeutic treatments take several weeks before they begin to improve symptoms. This gradual improvement is consistent with our everyday experience of the self. However, a new treatment, ketamine, can alter depressive symptoms within hours or even minutes. Ketamine can stop suicidal thoughts in as little as 20 minutes and is being developed to treat suicidal patients in the emergency room. Rapid changes in a belief as complex as suicide raise new questions about the self and identity. More complicated models of the self are required to understand how rapid changes in complex beliefs are possible without invoking unrealistic specificity in the brain (i.e. grandmother cells for belief). If the brain has multiple competing selves within modules then ketamine could work by either turning a self module on or off. This is consistent with a modified version of the global workspace theory of consciousness where a self module provides the context for the spotlight of attention.

Travels with Charlie

Charlie Gross is a maverick of modern neuroscience. Never content to tread the straight and worn path, never bridled by accepted views and traditional approaches, Charlie has often veered off the tracks, mapped much new terrain, and repeatedly opened our eyes to new ways of thinking. His scientific contributions are prodigious, transformative, and sometimes—as in his discovery of “face cells” in the cerebral cortex—the stuff of legend. Charlie is also deeply loved by the generations of students and colleagues with whom he has worked. We love Charlie for his wit and wisdom; for the passionate joie de vivre that fills his personal and professional spheres; for the devotion he extends to his students, their work, and their lives; for a certain (often bizarre) unaffected goofiness and unpredictability; and for those seemingly stray provocative comments that tilt the frame of your world. On May 25 and 26, 2013, a great many of Charlieʼs former and current students, colleagues, and friends assembled in Princeton to honor him, to celebrate his contributions to science and society, and to express our gratitude to a man who has profoundly influenced our careers, our lives, and the way we think. “Charlie Fest” was both scientific symposium and party: We began Saturday with a series of scientific lectures in old Green Hall. This special volume of the Journal of Cognitive Neuroscience is largely a compendium of those presentations. Saturday evening consisted of a banquet at which Charlie was both praised and gently ridiculed by his students and colleagues. Sunday was appropriately marked by a celebration with traditional pig roast (more about that later) at Charlieʼs home. Much has been written about the age in which Charlie worked—five remarkable decades in which the field of neuroscience not only blossomed but resonated with a public eager to understand why we do the things we do—and about his life and contributions. Charlieʼs autobiography (“Being Charlie Gross”) appears in the Society for Neuroscience-sponsored collection known as The History of Neuroscience (Vol. 6). The engaging story of his Brooklyn youth, his family, and the unique factors that influenced his development as child of Jewish intellectual Communist Party members is told in “Wait: A Memoir of a Red Diaper Baby” (Boulevard, 2009). Charlieʼs historical impact on the field of neuroscience is also covered in his “Genealogy of the Grandmother Cell” (The Neuroscientist, 2002), “Single neuron studies of the inferior temporal cortex” (Neuropsychologia, 2007), and “Processing the Facial Image: A Brief History” (American Psychologist, 2005). The scientific articles in this special issue of the Journal of Cognitive Neuroscience reveal just a hint of the scientific world that Charlie has spawned. After a brief summary of my connection to Charlie and a sketch of major episodes in his career, to help set the stage, what I aim to do here is convey a sense of the man—that scruffy, shuffling, frenetic, boisterous, seriously intellectual, mischievous, inspired, ardent, radical, bon vivant provocateur. To convey a sense of what he has meant to us through the years and how he has helped make us who we are—not just as neuroscientists but also as people and members of society. At the same time, I canʼt help but note that these character manifestations of Charlie are precisely the reasons he has deeply impacted

The Emergence of Mind: Personal Knowledge and Connectionism

The Emergence of Mind: Personal Knowledge and Connectionism

Motor Learning Mechanism on the Neuron Scale

Based on existing data, we wish to put forward a biological model of motor system on the neuron scale. Then we indicate its implications in statistics and learning. Specifically, neuron firing frequency and synaptic strength are probability estimates in essence. And the lateral inhibition also has statistical implications. From the standpoint of learning, dendritic competition through retrograde messengers is the foundation of conditional reflex and grandmother cell coding. And they are the kernel mechanisms of motor learning and sensory motor integration respectively. Finally, we compare motor system with sensory system. In short, we would like to bridge the gap between molecule evidences and computational models.

Gnostic cells in the 21st century.

In this short review, I revise the notion of gnostic cells posited by Konorski, together with similar arguments by James, Lettvin and Barlow--namely, the idea of pontifical, grandmother and cardinal cells, respectively. I then discuss whether the characteristics of the recently discovered concept cells, i.e. neurons in the human medial temporal lobe with a very high degree of specificity and invariance, fit the conjecture of gnostic or grandmother cells and then discuss the key role of concept cells in memory formation.