Models and theory
Experiment Toolbox: An Approach To Intuitive Experimental Design And Programming
Daniel Reutter and Michael Zehetleitner
Nowadays, most visual experiments are conducted with the aid of computers. The programming of said computers is left for the experimenter. Libraries and toolboxes do exist to support the scientists by and large, but in the end, the task of developing the software running their specific experiment is left to themselves. Our Experiment Toolbox (ETB) defines a set of development tools that pick up the experimenters from a point of view they understand best: The methodical description of the experiment itself. At the same time, the ETB avails existing backends for speed and accuracy, thereby relying on the stability of proven frameworks. Definitions for constants and variables are put in the ETB human-readable and are easily edited. While the constant parameters of the experiment are stored in text files as key-value pairs, the independent variables of the experiment are listed in tabular form, where each column represents a variable and each line one trial of the experiment. This intuitive storage reflects the idea behind an experiment, to explore certain conditions of a given problem. The experiment's structure is defined with a simple set of functions in the backend's language. The functions represent the mindset of the experimenter, letting him define displays and measurements - the same way the experiment would be described in a paper. Simultaneously, the ETB does not hinder the experimenter by occluding the underlying programming language. The ETB furthermore defines interfaces for providing interchangeable code collections. The current implementation of the ETB provides all functions needed to interact with the files for parameters and the experiment structure and ensures the correct execution of the so defined experiment. It is written in MATLAB and relies on the established Psychophysics Toolbox for display handling and timing.
A neural network model of perceptual grouping ability for 2-dimensional patterns
Maria Pietronilla Penna, Anna Montesanto and Eliano Pessa
This contribution reports about a number of computer simulations of the behaviour of a model of emergence of perceptual grouping ability. This model is a simplified version of a neural network model of visual cortex operation already proposed some years ago by Grossberg and coworkers and belonging to the class of LAMINART models. Relying on previous simulations, performed on 1-dimensional patterns, we studied the case of 2-dimensional patterns, in order to evidence the influence of model parameter values on perceptual grouping ability. The simulations evidenced the crucial role played by a particular parameter modulating the strength of the signal coming from the input layer to an intermediate layer of inhibitory cells. The value of this parameter is critical in allowing the existence of perceptual grouping ability. It is possible to speculate about the possible adaptive origin of the allowed values of this parameter in biological vision systems. References Leveille, J., Versace, M., Grossberg, S. (2010). Running as fast as it can: How spiking dynamics form object groupings in the laminar circuits of visual cortex. Journal of Computational Neuroscience, 28: 323-346. Yazdanbakhsh, A., Grossberg, S. (2004). Fast synchronization of perceptual grouping in laminar visual cortical circuits. Neural Networks, 17: 707-718.
Variability in pupil size during a cognitive task using nonlinear dynamics analysis
Olival Lago, Giuliano Ginani, Sabine Pompeia and Paulo Schor
Measuring the dynamics of pupillary motility is a challenge for many researchers due to its nonlinear behavior and produce many outliers and noise. Our goal is to introduce method to analyze the variability in pupil size during the cognitive task. From the data vector X=(X1,X2,...,Xn+1), we define the auxiliary vectors x=(x1,x2,...,xn) and y=(y2,y3,...,yn+1), which are analyzed in ordered pairs (xi,yi). Taking as reference the line y=tan(pi/4)*x+0, we measured the perpendicular distance from each pair (xi,yi) to the line, and calculate its dispersion by the standard deviation (SDa). Also calculate the standard deviation of the points perpendicular to the line y=tan(3pi/4)*x. The SDa estimates the variations of the points in successive intervals (short-term variation) while SDb estimates the dispersion of points along the line of reference (long-term variation). We analyzed subjects performing Random Number Generation test and compare the pupil size during the test with the baseline, we evaluated their responses by parameters SDa and SDb. The results showed that there were consistente changes in these parameters associated cognitive exertion, suggesting a pupillary dynamics, with variations of short and long term well-defined. There are reasons to believe that the method is promising for analysis pupillary in cognitive tasks. [Supported by FAPESP(2010/00455-0) to PS, OCL, and (2011/01286-0) to GEG, SP]
Cooperative dynamics of motion binding
Guillermo Aguilar, Maurizio Mattia, Paolo Del Giudice, Alexander Pastukhov and Jochen Braun
Many aspects of visual perception are characterized by 'cooperative behavior' such as order-disorder transitions, multi-stable states, and 'hysteresis' (Williams, Phillips, Sekuler, 1986, Nature), presumably reflecting recurrent excitation and inhibition within the underlying neural populations (Buckthought, Kim, Wilson, 2008, Vision Res). We have studied hysteresis in the transition between ordered and disordered percepts induced by random-dot-kinematograms (RDK). Our aim was to establish the probability and dynamics of order-disorder transitions and to characterize their dependence on the number of display elements (dots). Six observers viewed continuous RDK for 5 s - during which the fraction of orderly dots varied unpredictably (low-pass-filtered random walk) - and reported their initial and final percepts (orderly/disorderly). Orderly dots produced a coherent motion flow (clockwise or anti-clockwise, converging or diverging spirals, varying angular velocity). We obtained pronounced hysteresis for 'orderly fractions' from 0.4 to 0.6. Random walks terminating at 0.5 consistently yielded different percepts, depending on the orderly fraction 1 to 2 s prior termination, revealing a surprisingly pronounced influence of prior history. We introduce a novel method for characterizing the cooperative dynamics of order-disorder transitions in motion binding. These observations provide powerful constraints for abstract neural models based on discrete stochastic integration (Braun, Mattia, 2010, Neuroimage).
Did I do that? Causal Inference of Authorship in goal-directed actions for impoverished stimuli
Tobias F Beck, Carlo Wilke, Barbara Wirxel, Dominik Endres, Axel Lindner and Martin Giese
The perception of own actions is affected by visual information and internal predictions [Wolpert et al,1995,Science,269,1880-1882]. Integration of these cues depends on their accuracies [Burge et al,2008,Journal of Vision,8(4:20),1-19], including the association of visual signals with one's own action or with unrelated external changes [Körding et al,2007,PLOSOne,2(9)]. This attribution should thus depend on both consistency between predicted and actual visual consequences and their signal accuracy. METHODS. We used a virtual-reality setup to manipulate the consistency between pointing movements and their visual consequences. We investigated the influence of this consistency on self-action perception for varying visual accuracies. We asked whether a causal inference model accounts for the empirical data, assuming a latent authorship-variable: if the visual stimulus was attributed to one's own action, visual and internal information should fuse in a Bayesian optimal manner. If the stimulus was attributed to external influences, there should be no fusion. RESULTS & CONCLUSION. The model, fitted to the motor responses, correctly predicts authorship ratings, showing attribution of visual signals to one's own action for small, and stronger reliance on internal information for large deviations. Presently, we test predictions of the model for a variation of the accuracy of the visual information. ACKNOWLEDGEMENTS. This work was supported by: BMBF FKZ: 01GQ1002, EC FP7-ICT grants TANGO 249858, AMARSi 248311, and DFG GI 305/4-1, DFG GZ: KA 1258/15-1.
Spatial remapping during combined eye and body motion
Ivar Clemens, Luc Selen and W. Pieter Medendorp
A typical characteristic of human vision is that the position of the retina is constantly changing due to eye, head, or body movements. These changes cause the retinal image to shift. Recent studies on the retinal image shifts caused by saccades have suggested that the brain corrects for these perturbations by remapping the neural image. To date, it is not known how and to which extent this spatial remapping is implicated during whole-body motion. We used a novel 2AFC paradigm to investigate this question. Participants (n=8) were translated sinusoidally (15 or 30 cm, 1.6 Hz) in lateral direction while fixating a stationary point. When the direction of motion reversed a reference LED was presented (50ms). At the next motion reversal, a probe LED was flashed and participants had to judge its position (left or right) relative to the remembered reference. We found systematic biases in the spatial updates, which depended on the retinal location of the reference LED. Furthermore, the bias was in opposite directions for references presented in front versus behind fixation. We show that our results can be explained by a gaze-centered model, with an underestimation of translation amplitude.
A new measure of motor-estimated size early in the grasp
Rachel M. Foster, Joep Vanlier and Volker H. Franz
Foster and Franz [VSS 2012] demonstrated a mathematical side-effect of aggregating grasping trajectories: Temporal noise creates misalignment of the trajectories, which induces a dependence of grip aperture variability (ApVar) and aperture velocity (ApVel). Consequently, ApVar at early time points (when the fingers are opening and ApVel is large) must be higher than at later time points (near the maximum grip aperture, when ApVel is small). This presents a problem for studies that used ApVar to investigate whether Weber's Law exists early but not late in grasping, because only at early time points does ApVel and therefore ApVar depend on object size. An alternative approach is therefore necessary to compare early and late time points. We fit a simple force-inertia based grasp model to trajectories of N=18 participants and used the resulting parameter estimates as dependent variables. We discovered that, after an initial acceleration period, a constant velocity is reached which varies linearly with object size, similar to MGA. This is achieved quite early in the grasping movement. We therefore suggest using this model-based parameter as dependent variable when discerning motor-estimated size at early time points. We also apply this method to the original questions related to Weber's Law.
The abstraction of visual numerosity from continous visual properties
Ivilin Peev Stoianov and Marco Zorzi
Visual numerosity estimation is an evolutionary ancient ability foundational to mathematical learning supported by an Approximate Number System (ANS; Nieder & Dehane, 2009, Annual Review of Neuroscience, 32, 185-208). Based on novel results about the influence of continuous visual properties on numerosity estimation, Gebuis and Reynvoet (2012, JEP:General) suggested that number-judgment is the result of the weighting of several distinct visual cues and challenged the ANS hypothesis. We investigated this issue with our neurocomputational model of ANS (Stoianov & Zorzi, 2012, Nature Neuroscience, 15, 194-196) in which numerosity is an emergent high-order statistical image property coded by specialized number-sensitive neurons. Here we show that the model can perfectly account for the data of Gebuis and Reynvoet. At the same time, we demonstrate that the response of the model's numerosity detectors is largely invariant to both cumulative surface and occupied area, consistent with neurophysiological data (Roitman, Brannon, & Platt, 2007 PLoS Biology, 5, e208). Thus, the modulations induced by continuous visual cues do not challenge the notion of abstrac coding of numerosity in the ANS and are in agreement with the 'Visual Number Sense' hypothesis (Burr & Ross, 2008, Current Biology, 18, 424-428).
A contrast energy model for numerosity coding by human observers
Sabine Raphael, Barbara Dillenburger and Michael Morgan
We investigated the ability of human observers to discriminate and identify differences in numerosity, size or texture density between a test and a standard pattern. In the first experiment the patterns had an approximately circular envelope and the standard contained 64 dots. The number of dots in the test was altered by changing either the radius (size) of the pattern or the dot density. Changes in size were more accurately reported than changes in density. Numerosity changes were no better reported than changes in density. In a second experiment the shape outlines consisted of irregular polygons, making size estimation more difficult. Size discrimination thresholds were raised relative to those in Experiment 1 but those for density and numerosity were little affected. Changes in density could be accurately discriminated from changes in size. The data were fitted by a model in which number is computed from size and density and an unsigned 'Max' rule is used to decide whether size or density had changed. We also considered a mechanism for numerosity based on the quantity of 'stuff'in the texture, computed from the contrast energy integrated over spatial-frequency tuned channels.
Dynamic scenes as an experimental tool to provide behavioral evidence for a Coarse-to-Fine processing sequence during scene categorization
Louise Kauffmann, Benoit Musel, Nathalie Guyader, Carole Peyrin and Alan Chauvin
Studies agree that visual scenes are analyzed by the visual system in terms of spatial frequencies following a predominant coarse-to-fine (CtF) processing (low spatial preceding high spatial frequencies). In Musel et al, [submitted PlosOne]), someresults provided behavioral evidences that scene categorization was facilitated when scenes where displayed in a dynamic sequence following a CtF sequence compared to a fine-to-coarse (FtC) sequence. They used a bank of Gabor-like filters to decompose scenes in different spatial frequencies and they built dynamic sequences presenting the same filtered scene in different spatial frequencies either in CtF or FtC sequence. In this study, 72 participants performed a categorization task (indoor vs. outdoor scenes), using dynamic scene sequences, in which they resorted to either CtF or FtC sequences. Ten different types of sequences were built using different filters, central frequency samplings, normalizations and also, different durations for the filtered images that composed the dynamic sequences. Results showed that participants categorize CtF sequences more quickly than FtC sequences when sequences were built with constant band-pass filters, with the null frequency removal and normalization of the filtered scenes These dynamic scenes seem therefore well appropriate to highlight the neural substrates of the CtF categorization of scenes.
Comparisons of static and dynamic paradigms to investigate the Coarse-to-Fine categorization of scenes
Jessica Bourgin, Cécile Maire, Louise Kauffmann, Alan Chauvin, Nathalie Guyader and Carole Peyrin
Visual scene perception seems to be based on a Coarse-to-Fine (CtF) process where low spatial frequency luminance is rapidly extracted followed by higher spatial frequency chromatic information. However, a question still remains: The quality and quantity of information displayed in a filtered scene could favor low frequency oppose to other frequencies band. These studies used filtered scenes in different spatial frequency bands in order to disentangle the information available from the diagnostic information. We evaluated the contribution of different spatial frequency bands through several dynamic and static paradigms on scene categorization. Participants performed a categorization task (indoor vs. outdoor scenes), using static scene with 6 different spatial frequency content and dynamic natural scene sequences (in this case they resorted to either a CtF sequence or a reverse FtC sequence), displayed during 150 ms. Our results showed that diagnostic information resides in the middle frequency even though low frequency band is sufficient for an effective categorization. Thus, CtF strategy seems to be a mandatory processing but visual system capacities [Musel et al, accepted PlosOne], perceptual constraints (Mermillod et al 2005 Brain & Cognition, 58(2), 246-248] and task demands [Schyns & Oliva 1997 Perception, 26, 1027-1038] modulate the diagnosticity of specific spatial frequency bands.
The role of higher-order statistical dependency of natural images in visual perception
Hadi Maboudi and Mehdi Abouzari
In recent studies on computational vision from the efficient coding viewpoint, it assumed that visual sensory neurons are adjusted to the statistical properties of natural environment during sensory evolution. Thus understanding the statistics of natural images can help us to comprehend the function of visual sensory processing and perception. Natural images have non-random structures that reflect causal differences in the world. However, decorrelated natural images contain obvious structures but many of the important forms in the natural images require higher-order statistics to describe. It has been demonstrated that higher-order statistical structures of images are basis for the visual perception and object recognition and investigating such regularities could assist to clarify the spatiotemporal function of neurons in V1 and beyond. Nonetheless, higher-order statistics properties of natural scenes and their representation in the neural population are still unclear. In this study we extract spectra features from natural vs. non-natural images by using statistical methods to construct an statistical model of natural image space. Our findings indicate that there are some significant differences between natural images and random spaces which are critical for visual perception. Our results will be useful for optimal modeling of visual system and can help to develop hierarchical models for learning non-linear regularities in natural images.