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.

2.2. HIGH-LEVEL FEEDBACK<br />

PREDICTIVE CODING SHARPENING<br />

Li J .Li 11.1 A I<br />

HighLevRi Model High-I B\ie\ Model<br />

I • I. \u\Al<br />

Res dual Sharpened Response<br />

UiX.ljl. uUll'iL<br />

Law-level Inpul PaHem Low-level Input Patient<br />

Figure 2.10: Comparison <strong>of</strong> predictive coding (left) and sharpening (righl)i;ffects in mediating<br />

response reduclimi in lower levels. In predictive coding, a high-level prediction<br />

<strong>of</strong> the expected inpui is fed back and subiracied at the inpul level. What is sent<br />

forward is the dillerenee heiween the expected value and ihe ueiual input. With<br />

sharpening (presenl in biased compelilion and adaptive re.sonance models), the<br />

same high-level predieiion is fed back but is instead used loenhance those aspects<br />

<strong>of</strong> ihe input thai are consi.sleni wilh ihc high-level percepi and reduce all other<br />

aspects. The result, in both eases, can he a reduciion in activity. (Murray el al.<br />

20(M).<br />

error. While the error population will show reduction with high-level feedback, the prediction<br />

population may show enhancement. Predictive coding theories can be misleading, as they place<br />

a stronger focus on the error-delecting nodes and consequently under-emphasize or omit predic­<br />

tion nodes. A second requisite to reconcile biased competition and predictive coding theories<br />

regards the connectivity <strong>of</strong> feedback. In most biased competition mtxlels. nodes al each level<br />

compete by inhibiting Ihe output <strong>of</strong> neighhotiring nodes, while feedback in predictive coding<br />

typically acts on the level below. Therefore, the bia,sed conipeiiiion model thai was shown to<br />

be mathematically equivalent to predictive coding (Spratling 2008b). required an alternative<br />

implementation thai suppressed the inputs <strong>of</strong> neighbouring nodes,<br />

It is not therefore surprising that predictive coding models are also compatible with theories<br />

<strong>of</strong> feedback as attention. Both Rao (2005). by extending his original model, and .Spralling<br />

(2008a). using the previously described architecture, demonstrated thai predictive coding could<br />

account for spatial and feature attentional effecls. Turihermore, Spratling (2()08a) hypothe-<br />

45

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

Saved successfully!

Ooh no, something went wrong!