Tony Vladusich

Prototypical category learning in high-functioning autism.

An ongoing debate in developmental cognitive neuroscience is whether individuals with autism are able to learn prototypical category representations from multiple exemplars. Prototype learning and memory were examined in a group of high‐functioning autistic boys and young men, using a classic paradigm in which participants learned to classify novel dot patterns into one of two categories. Participants were trained on distorted versions of category prototypes until they reached a criterion level of performance. During transfer testing, participants were shown the training items together with three novel stimulus sets manifesting variable levels of physical distortion (low, medium, or high distortion) relative to the unseen prototypes. Two experiments were conducted, differing only in the manner in which the physical distortions were defined. In the first experiment, a subset of autistic individuals learned categories more slowly than controls, accompanied by an overall diminution in transfer‐testing performance. The autism group did, however, manifest a typical pattern of performance across the testing conditions, relative to controls. In the second experiment, group means did not differ statistically in either the training or testing phases. Taken together, these data indicate that high‐functioning autistic individuals do not manifest gross deficits in prototypical category learning. A theoretical discussion is given in terms of how perceptual grouping may interact with category learning.

Fast, Simple and Accurate Handwritten Digit Classification by Training Shallow Neural Network Classifiers with the ‘Extreme Learning Machine’ Algorithm

Recent advances in training deep (multi-layer) architectures have inspired a renaissance in neural network use. For example, deep convolutional networks are becoming the default option for difficult tasks on large datasets, such as image and speech recognition. However, here we show that error rates below 1% on the MNIST handwritten digit benchmark can be replicated with shallow non-convolutional neural networks. This is achieved by training such networks using the ‘Extreme Learning Machine’ (ELM) approach, which also enables a very rapid training time (∼ 10 minutes). Adding distortions, as is common practise for MNIST, reduces error rates even further. Our methods are also shown to be capable of achieving less than 5.5% error rates on the NORB image database. To achieve these results, we introduce several enhancements to the standard ELM algorithm, which individually and in combination can significantly improve performance. The main innovation is to ensure each hidden-unit operates only on a randomly sized and positioned patch of each image. This form of random ‘receptive field’ sampling of the input ensures the input weight matrix is sparse, with about 90% of weights equal to zero. Furthermore, combining our methods with a small number of iterations of a single-batch backpropagation method can significantly reduce the number of hidden-units required to achieve a particular performance. Our close to state-of-the-art results for MNIST and NORB suggest that the ease of use and accuracy of the ELM algorithm for designing a single-hidden-layer neural network classifier should cause it to be given greater consideration either as a standalone method for simpler problems, or as the final classification stage in deep neural networks applied to more difficult problems.

Enhanced image classification with a fast-learning shallow convolutional neural network

We present a neural network architecture and training method designed to enable very rapid training and low implementation complexity. Due to its training speed and the absence of iteratively-tuned parameters, the method has strong potential for applications requiring frequent retraining or online training. The approach is characterized by (a) convolutional filters based on biologically inspired visual processing filters, (b) randomly-valued classifier-stage input weights, (c) use of least squares regression to train the classifier output weights in a single batch, and (d) linear classifier-stage output units. We demonstrate the efficacy of the method by applying it to image classification. Our results match existing state-of-the-art results on the MNIST (0.37% error) and NORB-small (2.2% error) image classification databases, but with very fast training times compared to standard deep network approaches. The networks performance on the Google Street View House Number (SVHN) (4% error) database is also competitive with state-of-the art methods.

Edge integration and the perception of brightness and darkness

How do induced brightness and darkness signals from local and remote surfaces interact to determine the final achromatic color percept of a target surface? An emerging theory of achromatic color perception posits that brightness and darkness percepts are computed by weighting and summing the induction signals generated at edges in a scene. This theory also characterizes how neighboring edges interact to modulate the gain of brightness and darkness signals induced from one another. Here we assess evidence for this edge integration theory by means of computational modeling and a psychophysical experiment. We quantitatively show how local and remote edge induction signals in disk-ring displays give rise to either contrast or assimilation effects. Spatial integration of same-polarity edge signals supports a contrast effect, whereas integration of opposite-polarity signals supports an assimilation effect, particularly when the remote induction signal is much stronger than the local induction signal. The results confirm a key prediction of edge integration theory, namely, that strong assimilation effects can lead subjects to ignore the polarity of local edge information when setting achromatic color matches. The conditions necessary for strong assimilation effects are also associated with greater difficulty in setting matches, suggesting that caution is required when interpreting matching data in terms of gain control. We describe several avenues for further study of contrast, assimilation, and gain control.

Simultaneous contrast and gamut relativity in achromatic color perception.

Simultaneous contrast refers to the respective whitening or blackening of physically identical image regions surrounded by regions of low or high luminance, respectively. A common method of measuring the strength of this effect is achromatic color matching, in which subjects adjust the luminance of a target region to achieve an achromatic color match with another region. Here I present psychophysical data questioning the assumption--built into many models of achromatic color perception--that achromatic colors are represented as points in a one-dimensional (1D) perceptual space, or an absolute achromatic color gamut. I present an alternative model in which the achromatic color gamut corresponding to a target region is defined relatively, with respect to surround luminance. Different achromatic color gamuts in this model correspond to different 1D lines through a 2D perceptual space composed of blackness and whiteness dimensions. Each such line represents a unique gamut of achromatic colors ranging from black to white. I term this concept gamut relativity. Achromatic color matches made between targets surrounded by regions of different luminance are shown to reflect the relative perceptual distances between points lying on different gamut lines. The model suggests a novel geometrical approach to simultaneous contrast and achromatic color matching in terms of the vector summation of local luminance and contrast components, and sets the stage for a unified computational theory of achromatic color perception.

A reinterpretation of transparency perception in terms of gamut relativity

Classical approaches to transparency perception assume that transparency constitutes a perceptual dimension corresponding to the physical dimension of transmittance. Here I present an alternative theory, termed gamut relativity, that naturally explains key aspects of transparency perception. Rather than being computed as values along a perceptual dimension corresponding to transmittance, gamut relativity postulates that transparency is built directly into the fabric of the visual systems representation of surface color. The theory, originally developed to explain properties of brightness and lightness perception, proposes how the relativity of the achromatic color gamut in a perceptual blackness-whiteness space underlies the representation of foreground and background surface layers. Whereas brightness and lightness perception were previously reanalyzed in terms of the relativity of the achromatic color gamut with respect to illumination level, transparency perception is here reinterpreted in terms of relativity with respect to physical transmittance. The relativity of the achromatic color gamut thus emerges as a fundamental computational principle underlying surface perception. A duality theorem relates the definition of transparency provided in gamut relativity with the classical definition underlying the physical blending models of computer graphics.

A Unified Account of Perceptual Layering and Surface Appearance in Terms of Gamut Relativity

When we look at the world—or a graphical depiction of the world—we perceive surface materials (e.g. a ceramic black and white checkerboard) independently of variations in illumination (e.g. shading or shadow) and atmospheric media (e.g. clouds or smoke). Such percepts are partly based on the way physical surfaces and media reflect and transmit light and partly on the way the human visual system processes the complex patterns of light reaching the eye. One way to understand how these percepts arise is to assume that the visual system parses patterns of light into layered perceptual representations of surfaces, illumination and atmospheric media, one seen through another. Despite a great deal of previous experimental and modelling work on layered representation, however, a unified computational model of key perceptual demonstrations is still lacking. Here we present the first general computational model of perceptual layering and surface appearance—based on a boarder theoretical framework called gamut relativity—that is consistent with these demonstrations. The model (a) qualitatively explains striking effects of perceptual transparency, figure-ground separation and lightness, (b) quantitatively accounts for the role of stimulus- and task-driven constraints on perceptual matching performance, and (c) unifies two prominent theoretical frameworks for understanding surface appearance. The model thereby provides novel insights into the remarkable capacity of the human visual system to represent and identify surface materials, illumination and atmospheric media, which can be exploited in computer graphics applications.