Attention to Proto-Objects: Grouping Based Saliency

Thu, June 30, 2011 3-4pm

Alexander Russell, Stefan Mihalas, Ernst Niebur and Ralph Etienne-Cummings
The Johns Hopkins University, Baltimore, MD

A visual saliency map is useful in any visual sensory system (be it artificial or biological) in that it helps deal with information overload by directing the application of computational resources to interesting parts of the image. Itti et al. [1] developed a saliency map which utilizes differences in features across an image to find interesting (salient) locations. The model was able to predict human eye fixations for images featuring pop out and predict human eye fixations better than chance in more complex scenes [1]. However, Gestalt psychologists argue that humans perceive the whole before they analyze individual features. A recent study [2], showing that objects predict eye-fixations better than features, supports this. The aim of the research conducted over the past year was to build the next generation saliency map which finds salient objects rather than features. First, the visual scene is deconstructed into intensity and color-opponency channels. The object based saliency algorithm then uses the principles of border ownership neurons discovered by Zhou et al [3] and grouping neurons, proposed by Craft et al [4], to extract objects from each channel at different scales. A normalization is then applied so that unique objects are awarded the highest saliency and common objects the lowest saliency. The algorithm design and a MATLAB implementation have been completed. The ability of the algorithm to predict human eye fixations is currently being tested. Preliminary results show a 300% increase in the ability to predict eye fixations over the Itti et al. feature based saliency map. Work to accelerate the object based saliency algorithm in hardware (FPGA) has begun. This will allow for the real-time application of the algorithm to large scenes. This will provide a mechanism to extract interesting objects from the scene and pass them to more computationally expensive algorithms such as object recognition.

[1] L. Itti, C. Koch, and E. Niebur. "A model of saliency-based visual attention for rapid scene analysis." IEEE Pattern Analysis and Machine Intelligence, (11):1254–1259, 1998

[2]W. Einhauser, M. Spain, and P. Perona. "Objects predict fixations better than early saliency". J.Vision, pages 1–26, 2008.

[3]H. Zhou, H. S. Friedman, and R. von der Heyt. "Coding of border ownership in monkey visual cortex." J. Neuroscience, pages 6594–6611, 2000

[4]E. Craft, H. Schutze, E. Niebur, and R. von der Heyt. "A neural model of figure ground organization". J. Neurophysiology, pages 4310 – 4326, 2007.

How Variances Change with Attention

Friday, July 1, 2011 2-3pm

Jude F. Mitchell

Systems Neurobiology Lab, The Salk Institute, La Jolla, CA

Although previous studies of attention often focus on increases in firing rate as the neural correlate of improved processing for attended items, it is also important to consider how attention alters the variability of response. In particular, if sources of variability are correlated at the population level then it is not possible to average out noise by pooling, thus limiting the accuracy with which sensory information can be decoded (Zohary et. al., 1994). Recently we found that attending to a sustained stimulus inside a V4 neuron's receptive field led to a reduction in the trial-to-trial variability of the response, as measured by the Fano Factor, and that reductions were much stronger among neurons with narrower action potentials, putative fast spiking interneurons (Mitchell et. al., 2007). Here we examine the timescales of the underlying firing rate fluctuations that contribute to this variability and whether or not they are correlated at the population level. The dominant source of noise is due to low frequency (< 8hz) fluctuations in firing rate that are reflected both in spike-LFP coherence reductions. We analyzed the correlations in spiking of simultaneously recorded pairs to evaluate if these sources of noise were correlated in the population. We find that the low frequency rate fluctuations are correlated across units and importantly that these correlations are reduced by attention (Mitchell et al., 2009). Low frequency fluctuations are ubiquitous to cortical firing and may represent one source of recurrent network noise that attention modulates to improve sensory signals.

Sparse approximation on a network of locally competitive integrate and fire neurons

Tue, July 5, 2011 3-4pm

'Samuel Shapero', Daniel Bruderle, 'Hasler', and Christopher Rozell

See abstract attached

Visual Attention at the Systems Level

Monday July 11, 2011 3-4PM

'Fred Hamker'

Modeling the physiology and psychophysics of visual attention at the systems level

I will give an overview about the systems-level circuits of visual attention and explain how attention emerges by competitive interactions between brain areas. In particular, it will be shown how attention will affect the receptive field structure and how this will lead to particular observations at the psychophysical level, such as the mislocalization of briefly flashed stimuli around saccade onset. Finally, an outline how to integrate attention with object recognition and cognitive control will be given.

Attention and Auditory Scene Analysis

Thursday, July 7, 2011 2-3PM

Kang Wang

The Ohio State University

Attention does not play a big role in Bregman’s original theory of auditory scene analysis (ASA). To what extent is ASA pre-attentive? Is attention required for auditory organization? What is the neural basis of ASA and auditory attention? This presentation attempts to address these issues. I will discuss an oscillatory correlation model of auditory attention, and ways of incorporating saliency into such models. I will also try to place attention in the larger context of perception, and give a computational theory analysis of ASA. This analysis has led to a new formulation of ASA as binary classification. This formulation shifts the emphasis from signal estimation to signal classification, and opens the ASA problem to modern classification techniques in machine learning.