08.02.2013 Views

Bernal S D_2010.pdf - University of Plymouth

Bernal S D_2010.pdf - University of Plymouth

Bernal S D_2010.pdf - University of Plymouth

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

6.4. WTUREWURK<br />

by Chikkerur et al. (2009) also implements exact inference on a Bayesian network but models<br />

exclusively high-level atteniion, such that the lower half <strong>of</strong> the network is non-Bayesian and<br />

strictly feedforward.<br />

The type <strong>of</strong> input image used by the mtxlel is more complex and detailed than that <strong>of</strong> previous<br />

ones that were purely theoretical (Lee and Mumford 2003) or employed simplistic toy exam­<br />

ples ((•riston and Kiebel 2(H)9. Lewicki and Scjnowski 1997. Hinton et al. 2006), Those with<br />

comparable input images fail to account for other properties that have been implemented by<br />

the proposed model, such as position and scale invariance (Rao and Ballard 1997, Murray and<br />

Kreutz-Delgado 2007, Chikkerur et al. 2009) or illusory contour completion (Hinlon et al. 2006,<br />

Epshtein et al. 2008).<br />

6.4 Future work<br />

A number <strong>of</strong> potential improvements and extensions to the proposed model are listed below:<br />

• Run simulations with Kanizsa tij;ure controls that fully lest the hypothesis that the model<br />

performs illusory contour completion. Althe moment, the control data give an ambiguous<br />

answer to the model's performance.<br />

• Perform a systematic analysis <strong>of</strong> the model parameters for both feedforward and feedback<br />

processing. Some <strong>of</strong> the key parameters to study arc the number <strong>of</strong> features per group.<br />

Ihe sparseness <strong>of</strong> the connectivity matrices and the sampling parameters.<br />

• lj;am heterogeneous feedback weights for features within a group and allow features to<br />

belong to different groups. This should improve the feedback disambiguation capacity<br />

and could lead to improved contextual modulation through lateral interactions.<br />

• Improve the categorization <strong>of</strong> Kanizsa figures and the feedback C2-S2 reconstruction<br />

to allow automatic illusory contour completion without clamping feedback. This could<br />

also lead to an improvement <strong>of</strong> the categorization performance over time as a result <strong>of</strong><br />

feedback modulation.<br />

• Include adaptation mechanisms that could lead naturally to phenomena such as sensitivity<br />

259

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!