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.

4.7. ORIGINAL CONTRIBUTIONS IN THIS CHAPTER<br />

used by IJtvak and Ullmaii (2009). George and Hawkins (2009). This approximalion is<br />

justified when the total number <strong>of</strong> incoming messages to a node is relatively high, as is<br />

the case <strong>of</strong> the present model<br />

• The CPT P{X |t/i, • . Ufj) is approximated as the weighted sum <strong>of</strong> N CPTs <strong>of</strong> the form<br />

f'{X\Ui). The method has been justified geometrically as providing a good model <strong>of</strong> the<br />

combination <strong>of</strong> information from multiple experts {parent nodes) and has been success­<br />

fully employed on other probabilistic models thai require reasoning under uncertainty<br />

(Das 2004).<br />

• For the calculation <strong>of</strong> the belief and the A messages, only ki„„ax highest-valued samples<br />

from ihc N,„a, ^ messages with the highest variance are employed. The method has been<br />

empirically demonstrated In provide a relatively good til to die exact distribution, given<br />

moderate values o( N„aj i"id ^umai-<br />

• To avoid the excessive computational cost associated to updating the beliefs and output<br />

messages <strong>of</strong> the nodes in all layers, beliefs are for a single layer at each time step, starl­<br />

ing from the bottom layer and moving upwards sequentially. The rationale behind this<br />

approximation to loopy belief propagation is that evidence arrives at Ihe network from<br />

dummy nodes connected to the bottom layer.<br />

4.7 Original contributions in tliis chapter<br />

• A Bayesian network that captures the structure <strong>of</strong> the HMAX model, a hierarchical object<br />

recognition model based on anatomical and physiological cortical data.<br />

• An approximalion to the selectivity and invariance operations <strong>of</strong> Ihe HMAX model using<br />

the belief propagation algorithm over ihe proposed Bayesian network.<br />

• An inherent extension <strong>of</strong> the static feedforward HMAX model to include dynamic and<br />

recursive feedback ba.sed on the loopy belief propagation algorithm.<br />

• A particularization <strong>of</strong> the CPT learning method proposed by Das (2004) to the hierarchi­<br />

cal object recognition domain. The method simplifies the generation <strong>of</strong> the CPT parame-<br />

189

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

Saved successfully!

Ooh no, something went wrong!