Fragment-based learning of visual object categories in non-human primates

When we perceive a visual object, we implicitly or explicitly associate it with an object category we know. Recent research has shown that the visual system can use local, informative image fragments of a given object, rather than the whole object, to classify it into a familiar category. We have previously reported, using human psychophysical studies, that when subjects learn new object categories using whole objects, they incidentally learn informative fragments, even when not required to do so. However, the neuronal mechanisms by which we acquire and use informative fragments, as well as category knowledge itself, have remained unclear. Here we describe the methods by which we adapted the relevant human psychophysical methods to awake, behaving monkeys and replicated key previous psychophysical results. This establishes awake, behaving monkeys as a useful system for future neurophysiological studies not only of informative fragments in particular, but also of object categorization and category learning in general.     

Visual categorization: when categories fall to pieces

We cannot help but categorize the visual world into objects like cats and faces. An intriguing new study shows that observers automatically discover informative fragments of visual objects during category learning.     

A Representational Similarity Analysis of the Dynamics of Object Processing Using Single-Trial EEG Classification

The recognition of object categories is effortlessly accomplished in everyday life, yet its neural underpinnings remain not fully understood. In this electroencephalography (EEG) study, we used single-trial classification to perform a Representational Similarity Analysis (RSA) of categorical representation of objects in human visual cortex. Brain responses were recorded while participants viewed a set of 72 photographs of objects with a planned category structure. The Representational Dissimilarity Matrix (RDM) used for RSA was derived from confusions of a linear classifier operating on single EEG trials. In contrast to past studies, which used pairwise correlation or classification to derive the RDM, we used confusion matrices from multi-class classifications, which provided novel self-similarity measures that were used to derive the overall size of the representational space. We additionally performed classifications on subsets of the brain response in order to identify spatial and temporal EEG components that best discriminated object categories and exemplars. Results from category-level classifications revealed that brain responses to images of human faces formed the most distinct category, while responses to images from the two inanimate categories formed a single category cluster. Exemplar-level classifications produced a broadly similar category structure, as well as sub-clusters corresponding to natural language categories. Spatiotemporal components of the brain response that differentiated exemplars within a category were found to differ from those implicated in differentiating between categories. Our results show that a classification approach can be successfully applied to single-trial scalp-recorded EEG to recover fine-grained object category structure, as well as to identify interpretable spatiotemporal components underlying object processing. Finally, object category can be decoded from purely temporal information recorded at single electrodes.     

A stable biologically motivated learning mechanism for visual feature extraction to handle facial categorization

The brain mechanism of extracting visual features for recognizing various objects has consistently been a controversial issue in computational models of object recognition. To extract visual features, we introduce a new, biologically motivated model for facial categorization, which is an extension of the Hubel and Wiesel simple-to-complex cell hierarchy. To address the synaptic stability versus plasticity dilemma, we apply the Adaptive Resonance Theory (ART) for extracting informative intermediate level visual features during the learning process, which also makes this model stable against the destruction of previously learned information while learning new information. Such a mechanism has been suggested to be embedded within known laminar microcircuits of the cerebral cortex. To reveal the strength of the proposed visual feature learning mechanism, we show that when we use this mechanism in the training process of a well-known biologically motivated object recognition model (the HMAX model), it performs better than the HMAX model in face/non-face classification tasks. Furthermore, we demonstrate that our proposed mechanism is capable of following similar trends in performance as humans in a psychophysical experiment using a face versus non-face rapid categorization task.     

Creating Objects and Object Categories for Studying Perception and Perceptual Learning

In order to quantitatively study object perception, be it perception by biological systems or by machines, one needs to create objects and object categories with precisely definable, preferably naturalistic, properties¹. Furthermore, for studies on perceptual learning, it is useful to create novel objects and object categories (or object classes) with such properties².Many innovative and useful methods currently exist for creating novel objects and object categories^3-6 (also see refs. 7,8). However, generally speaking, the existing methods have three broad types of shortcomings.First, shape variations are generally imposed by the experimenter^5,9,10, and may therefore be different from the variability in natural categories, and optimized for a particular recognition algorithm. It would be desirable to have the variations arise independently of the externally imposed constraints.Second, the existing methods have difficulty capturing the shape complexity of natural objects^11-13. If the goal is to study natural object perception, it is desirable for objects and object categories to be naturalistic, so as to avoid possible confounds and special cases.Third, it is generally hard to quantitatively measure the available information in the stimuli created by conventional methods. It would be desirable to create objects and object categories where the available information can be precisely measured and, where necessary, systematically manipulated (or ''tuned''). This allows one to formulate the underlying object recognition tasks in quantitative terms.Here we describe a set of algorithms, or methods, that meet all three of the above criteria. Virtual morphogenesis (VM) creates novel, naturalistic virtual 3-D objects called ''digital embryos'' by simulating the biological process of embryogenesis¹⁴. Virtual phylogenesis (VP) creates novel, naturalistic object categories by simulating the evolutionary process of natural selection^9,12,13. Objects and object categories created by these simulations can be further manipulated by various morphing methods to generate systematic variations of shape characteristics^15,16. The VP and morphing methods can also be applied, in principle, to novel virtual objects other than digital embryos, or to virtual versions of real-world objects^9,13. Virtual objects created in this fashion can be rendered as visual images using a conventional graphical toolkit, with desired manipulations of surface texture, illumination, size, viewpoint and background. The virtual objects can also be ''printed'' as haptic objects using a conventional 3-D prototyper.We also describe some implementations of these computational algorithms to help illustrate the potential utility of the algorithms. It is important to distinguish the algorithms from their implementations. The implementations are demonstrations offered solely as a ''proof of principle'' of the underlying algorithms. It is important to note that, in general, an implementation of a computational algorithm often has limitations that the algorithm itself does not have.Together, these methods represent a set of powerful and flexible tools for studying object recognition and perceptual learning by biological and computational systems alike. With appropriate extensions, these methods may also prove useful in the study of morphogenesis and phylogenesis.     

Do infant Japanese macaques (<Emphasis Type="Italic">Macaca fuscata</Emphasis>) categorize objects without specific training?

In the present study, we examined whether infant Japanese macaques categorize objects without any training, using a similar technique also used with human infants (the paired-preference method). During the familiarization phase, subjects were presented twice with two pairs of different objects from one global-level category. During the test phase, they were presented twice with a pair consisting of a novel familiar-category object and a novel global-level category object. The subjects were tested with three global-level categories (animal, furniture, and vehicle). It was found that they showed significant novelty preferences as a whole, indicating that they processed similarities between familiarization objects and novel familiar-category objects. These results suggest that subjects responded distinctively to objects without training, indicating the possibility that infant macaques possess the capacity for categorization.     

Visual artificial grammar learning: comparative research on humans, kea (Nestor notabilis) and pigeons (Columba livia)

Artificial grammar learning (AGL) provides a useful tool for exploring rule learning strategies linked to general purpose pattern perception. To be able to directly compare performance of humans with other species with different memory capacities, we developed an AGL task in the visual domain. Presenting entire visual patterns simultaneously instead of sequentially minimizes the amount of required working memory. This approach allowed us to evaluate performance levels of two bird species, kea (Nestor notabilis) and pigeons (Columba livia), in direct comparison to human participants. After being trained to discriminate between two types of visual patterns generated by rules at different levels of computational complexity and presented on a computer screen, birds and humans received further training with a series of novel stimuli that followed the same rules, but differed in various visual features from the training stimuli. Most avian and all human subjects continued to perform well above chance during this initial generalization phase, suggesting that they were able to generalize learned rules to novel stimuli. However, detailed testing with stimuli that violated the intended rules regarding the exact number of stimulus elements indicates that neither bird species was able to successfully acquire the intended pattern rule. Our data suggest that, in contrast to humans, these birds were unable to master a simple rule above the finite-state level, even with simultaneous item presentation and despite intensive training.     

Category selectivity in the ventral visual pathway confers robustness to clutter and diverted attention

Are objects coded by a small number of neurons or cortical regions that respond preferentially to the object in question, or by more distributed patterns of responses, including neurons or regions that respond only weakly? Distributed codes can represent a larger number of alternative items than sparse codes but produce ambiguities when multiple items are represented simultaneously (the "superposition" problem). Recent studies found category information in the distributed pattern of response across the ventral visual pathway, including in regions that do not "prefer" the object in question. However, these studies measured neural responses to isolated objects, a situation atypical of real-world vision, where multiple objects are usually present simultaneously ("clutter"). We report that information in the spatial pattern of fMRI response about standard object categories is severely disrupted by clutter and eliminated when attention is diverted. However, information about preferred categories in category-specific regions is undiminished by clutter and partly preserved under diverted attention. These findings indicate that in natural conditions, the pattern of fMRI response provides robust category information only for objects coded in selective cortical regions and highlight the vulnerability of distributed representations to clutter and the advantages of sparse cortical codes in mitigating clutter costs.     

Development of neural mechanisms for machine learning

The goal of this work is to develop a humanoid robot's perceptual mechanisms through the use of learning aids. We describe methods to enable learning on a humanoid robot using learning aids such as books, drawing materials, boards, educational videos or other children toys. Visual properties of objects are learned and inserted into a recognition scheme, which is then applied to acquire new object representations - we propose learning through developmental stages. Inspired in infant development, we will also boost the robot's perceptual capabilities by having a human caregiver performing educational and play activities with the robot (such as drawing, painting or playing with a toy train on a railway). We describe original algorithms to extract meaningful percepts from such learning experiments. Experimental evaluation of the algorithms corroborates the theoretical framework.     

Visual interpolation for contour completion by the European cuttlefish (Sepia officinalis) and its use in dynamic camouflage

Cuttlefish rapidly change their appearance in order to camouflage on a given background in response to visual parameters, giving us access to their visual perception. Recently, it was shown that isolated edge information is sufficient to elicit a body pattern very similar to that used when a whole object is present. Here, we examined contour completion in cuttlefish by assaying body pattern responses to artificial backgrounds of 'objects' formed from fragmented circles, these same fragments rotated on their axis, and with the fragments scattered over the background, as well as positive (full circles) and negative (homogenous background) controls. The animals displayed similar responses to the full and fragmented circles, but used a different body pattern in response to the rotated and scattered fragments. This suggests that they completed the broken circles and recognized them as whole objects, whereas rotated and scattered fragments were instead interpreted as small, individual objects in their own right. We discuss our findings in the context of achieving accurate camouflage in the benthic shallow-water environment.     

Emerging Object Representations in the Visual System Predict Reaction Times for Categorization

Recognizing an object takes just a fraction of a second, less than the blink of an eye. Applying multivariate pattern analysis, or “brain decoding”, methods to magnetoencephalography (MEG) data has allowed researchers to characterize, in high temporal resolution, the emerging representation of object categories that underlie our capacity for rapid recognition. Shortly after stimulus onset, object exemplars cluster by category in a high-dimensional activation space in the brain. In this emerging activation space, the decodability of exemplar category varies over time, reflecting the brain’s transformation of visual inputs into coherent category representations. How do these emerging representations relate to categorization behavior? Recently it has been proposed that the distance of an exemplar representation from a categorical boundary in an activation space is critical for perceptual decision-making, and that reaction times should therefore correlate with distance from the boundary. The predictions of this distance hypothesis have been born out in human inferior temporal cortex (IT), an area of the brain crucial for the representation of object categories. When viewed in the context of a time varying neural signal, the optimal time to “read out” category information is when category representations in the brain are most decodable. Here, we show that the distance from a decision boundary through activation space, as measured using MEG decoding methods, correlates with reaction times for visual categorization during the period of peak decodability. Our results suggest that the brain begins to read out information about exemplar category at the optimal time for use in choice behaviour, and support the hypothesis that the structure of the representation for objects in the visual system is partially constitutive of the decision process in recognition.     

Category learning induces position invariance of pattern recognition across the visual field

Human object recognition is considered to be largely invariant to translation across the visual field. However, the origin of this invariance to positional changes has remained elusive, since numerous studies found that the ability to discriminate between visual patterns develops in a largely location-specific manner, with only a limited transfer to novel visual field positions. In order to reconcile these contradicting observations, we traced the acquisition of categories of unfamiliar grey-level patterns within an interleaved learning and testing paradigm that involved either the same or different retinal locations. Our results show that position invariance is an emergent property of category learning. Pattern categories acquired over several hours at a fixed location in either the peripheral or central visual field gradually become accessible at new locations without any position-specific feedback. Furthermore, categories of novel patterns presented in the left hemifield are distinctly faster learnt and better generalized to other locations than those learnt in the right hemifield. Our results suggest that during learning initially position-specific representations of categories based on spatial pattern structure become encoded in a relational, position-invariant format. Such representational shifts may provide a generic mechanism to achieve perceptual invariance in object recognition.     

Influence of Emotionally Charged Information on Category-Based Induction

Categories help us make predictions, or inductions, about new objects. However, we cannot always be certain that a novel object belongs to the category we are using to make predictions. In such cases, people should use multiple categories to make inductions. Past research finds that people often use only the most likely category to make inductions, even if it is not certain. In two experiments, subjects read stories and answered questions about items whose categorization was uncertain. In Experiment 1, the less likely category was either emotionally neutral or dangerous (emotionally charged or likely to pose a threat). Subjects used multiple categories in induction when one of the categories was dangerous but not when they were all neutral. In Experiment 2, the most likely category was dangerous. Here, people used multiple categories, but there was also an effect of avoidance, in which people denied that dangerous categories were the most likely. The attention-grabbing power of dangerous categories may be balanced by a higher-level strategy to reject them.     

Shape Similarity,Better than Semantic Membership,Accounts for the Structure of Visual Object Representations in a Population of Monkey Inferotemporal Neurons

The anterior inferotemporal cortex (IT) is the highest stage along the hierarchy of visual areas that, in primates, processes visual objects. Although several lines of evidence suggest that IT primarily represents visual shape information, some recent studies have argued that neuronal ensembles in IT code the semantic membership of visual objects (i.e., represent conceptual classes such as animate and inanimate objects). In this study, we investigated to what extent semantic, rather than purely visual information, is represented in IT by performing a multivariate analysis of IT responses to a set of visual objects. By relying on a variety of machine-learning approaches (including a cutting-edge clustering algorithm that has been recently developed in the domain of statistical physics), we found that, in most instances, IT representation of visual objects is accounted for by their similarity at the level of shape or, more surprisingly, low-level visual properties. Only in a few cases we observed IT representations of semantic classes that were not explainable by the visual similarity of their members. Overall, these findings reassert the primary function of IT as a conveyor of explicit visual shape information, and reveal that low-level visual properties are represented in IT to a greater extent than previously appreciated. In addition, our work demonstrates how combining a variety of state-of-the-art multivariate approaches, and carefully estimating the contribution of shape similarity to the representation of object categories, can substantially advance our understanding of neuronal coding of visual objects in cortex.     

Object discrimination by pigeons: effects of object color and shape

Can nonhuman animals attend to visual stimuli as whole, coherent objects? We investigated this question by adapting for use with pigeons a task in which human participants must report whether two visual attributes belong to the same object (one-object trial) or to different objects (two-object trial). We trained pigeons to discriminate a pair of differently colored shapes that had two targets either on a single object or on two different objects. Each target equally often appeared on the one-object and two-object stimuli; therefore, a specific target location could not serve as a discriminative cue. The pigeons learned to report whether the two target dots were located on a single object or on two different objects; follow-up tests demonstrated that this ability was not entirely based on memorization of the dot patterns and locations. Additional tests disclosed predominate stimulus control by the color, but not by the shape of the two objects. These findings suggest that human psychophysical methods are readily applicable to the study of object discrimination by nonhuman animals.     

Using FMRI brain activation to identify cognitive states associated with perception of tools and dwellings

Previous studies have succeeded in identifying the cognitive state corresponding to the perception of a set of depicted categories, such as tools, by analyzing the accompanying pattern of brain activity, measured with fMRI. The current research focused on identifying the cognitive state associated with a 4s viewing of an individual line drawing (1 of 10 familiar objects, 5 tools and 5 dwellings, such as a hammer or a castle). Here we demonstrate the ability to reliably (1) identify which of the 10 drawings a participant was viewing, based on that participant's characteristic whole-brain neural activation patterns, excluding visual areas; (2) identify the category of the object with even higher accuracy, based on that participant's activation; and (3) identify, for the first time, both individual objects and the category of the object the participant was viewing, based only on other participants' activation patterns. The voxels important for category identification were located similarly across participants, and distributed throughout the cortex, focused in ventral temporal perceptual areas but also including more frontal association areas (and somewhat left-lateralized). These findings indicate the presence of stable, distributed, communal, and identifiable neural states corresponding to object concepts.     

Egocentric Representation Acquired from Offline Map Learning

It is widely accepted that people establish allocentric spatial representation after learning a map. However, it is unknown whether people can directly acquire egocentric representation after map learning. In two experiments, the participants learned a distal environment through a map and then performed the egocentric pointing tasks in that environment under three conditions: with the heading aligned with the learning perspective (baseline), after 240° rotation from the baseline (updating), and after disorientation (disorientation). Disorientation disrupted the internal consistency of pointing among objects when the participants learned the sequentially displayed map, on which only one object name was displayed at a time while the location of “self” remained on the screen all the time. However, disorientation did not affect the internal consistency of pointing among objects when the participants learned the simultaneously displayed map. These results suggest that the egocentric representation can be acquired from a sequentially presented map.     

Border collie comprehends object names as verbal referents

Four experiments investigated the ability of a border collie (Chaser) to acquire receptive language skills. Experiment 1 demonstrated that Chaser learned and retained, over a 3-year period of intensive training, the proper-noun names of 1022 objects. Experiment 2 presented random pair-wise combinations of three commands and three names, and demonstrated that she understood the separate meanings of proper-noun names and commands. Chaser understood that names refer to objects, independent of the behavior directed toward those objects. Experiment 3 demonstrated Chaser's ability to learn three common nouns - words that represent categories. Chaser demonstrated one-to-many (common noun) and many-to-one (multiple-name) name-object mappings. Experiment 4 demonstrated Chaser's ability to learn words by inferential reasoning by exclusion - inferring the name of an object based on its novelty among familiar objects that already had names. Together, these studies indicate that Chaser acquired referential understanding of nouns, an ability normally attributed to children, which included: (a) awareness that words may refer to objects, (b) awareness of verbal cues that map words upon the object referent, and (c) awareness that names may refer to unique objects or categories of objects, independent of the behaviors directed toward those objects.     

Object Learning Improves Feature Extraction but Does Not Improve Feature Selection

A single glance at your crowded desk is enough to locate your favorite cup. But finding an unfamiliar object requires more effort. This superiority in recognition performance for learned objects has at least two possible sources. For familiar objects observers might: 1) select more informative image locations upon which to fixate their eyes, or 2) extract more information from a given eye fixation. To test these possibilities, we had observers localize fragmented objects embedded in dense displays of random contour fragments. Eight participants searched for objects in 600 images while their eye movements were recorded in three daily sessions. Performance improved as subjects trained with the objects: The number of fixations required to find an object decreased by 64% across the 3 sessions. An ideal observer model that included measures of fragment confusability was used to calculate the information available from a single fixation. Comparing human performance to the model suggested that across sessions information extraction at each eye fixation increased markedly, by an amount roughly equal to the extra information that would be extracted following a 100% increase in functional field of view. Selection of fixation locations, on the other hand, did not improve with practice.     

Distributed neural plasticity for shape learning in the human visual cortex

Expertise in recognizing objects in cluttered scenes is a critical skill for our interactions in complex environments and is thought to develop with learning. However, the neural implementation of object learning across stages of visual analysis in the human brain remains largely unknown. Using combined psychophysics and functional magnetic resonance imaging (fMRI), we show a link between shape-specific learning in cluttered scenes and distributed neuronal plasticity in the human visual cortex. We report stronger fMRI responses for trained than untrained shapes across early and higher visual areas when observers learned to detect low-salience shapes in noisy backgrounds. However, training with high-salience pop-out targets resulted in lower fMRI responses for trained than untrained shapes in higher occipitotemporal areas. These findings suggest that learning of camouflaged shapes is mediated by increasing neural sensitivity across visual areas to bolster target segmentation and feature integration. In contrast, learning of prominent pop-out shapes is mediated by associations at higher occipitotemporal areas that support sparser coding of the critical features for target recognition. We propose that the human brain learns novel objects in complex scenes by reorganizing shape processing across visual areas, while taking advantage of natural image correlations that determine the distinctiveness of target shapes.