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

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

3.4. EXISTING MODFXS<br />

using two <strong>of</strong> the implementation methods described in this section {Lilvak and LHIman 2009,<br />

Steimer et al. 2009), have already been implemented using the slate-depcndenl VLSI technol­<br />

ogy (Emre Neftci, personal communication).<br />

Although the technology is still ai a very early stage and the scalability <strong>of</strong> the V1.SI spiking neu­<br />

ral networks is limited, it provides a slarting point for the development neuromorphic hardware<br />

capable <strong>of</strong> reproducing graphical models with cortical functionality.<br />

3.4.2 Functional models <strong>of</strong> visual processing<br />

This subsection focuses on mtxJels based on generative modelling approaches, which employ<br />

Bayesian networks/belief propagation or similar implementation methods. Specifically, we de­<br />

scribe models which deal with visual perception (recognition, reconstruction, etc) and have<br />

biologically grounded architectures. The literature in this area is very extensive so only mcxlels<br />

most relevant to this thesis are included. To facililalc comparison between models, they have<br />

been grouped according to the inference method employed (exact inference, sampling approx­<br />

imation, or variational approximation), although the classitication is not .strict as some models<br />

share characteristics <strong>of</strong> several methods. A summary and comparison <strong>of</strong> the models is included<br />

al the end <strong>of</strong> this subsection.<br />

3.4.2.1 Models based on exact inference mclhods<br />

The model proposed by Epshiein et aJ. (2008) extends a well-known feedforward object recog­<br />

nition model, namely Ultman's fragment-based hierarchical model described in Section 2.1.2.<br />

A hierarchy <strong>of</strong> informative Iragments and its corresponding smaller sub-fragments are learned<br />

for each class <strong>of</strong> objects. This information is stored using a factor graph where each variable<br />

represents an object fragment, which can take A' different values/.slates indicating the po.sition<br />

<strong>of</strong> that fragment within the image (a value <strong>of</strong> 0 indicates the fragment is not present). The<br />

relation (conditional probability function) between a sub-fragment and its parent fragments de­<br />

pends on the coordinate difference between the locations <strong>of</strong> child and parent fragments, and not<br />

on their absolute position. This allows the model to perform recognition with certain position<br />

invariance.<br />

121

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

Saved successfully!

Ooh no, something went wrong!