next up previous
Next: Discussion Up: The Bars Problem Previous: Learning Procedure


In the initial phases of learning, some of the sources represented multiple bars and there were multiple sources representing a single bar. There was also a source which was specialised to diminish variance in cases where both the horizontal and vertical orientations were inactive. These local minima were escaped, however, as the weights after 1200 sweeps in Figure 6 demonstrate.

The sources of the second layer are ordered for visualisation purposes according to the weights $ \mathbf{A}_2$ and $ \mathbf{B}_2$ using self-organising map. The two sources on the second layer correspond to the horizontal and vertical orientations and the 18 sources on the first layer correspond to the bars.

Regular bars, present in $ \mathbf{A}_{1}$, are reconstructed accurately but the variance bars in $ \mathbf{B}_{1}$ exhibit some noise. The distinction between horizontal and vertical orientations is clearly visible in $ \mathbf{A}_{2}$.

Figure 6: Posterior means of the weight matrices after 1200 sweeps. The matrices are organised in patches and dark shades represent positive values.
$\mathbf{A}_{2}$\ ($18 \times 2$) & $\mathbf{...
\epsfig{file=bar_B1.eps,width=0.20\textwidth} \end{tabular} \end{figure}

Harri Valpola 2001-10-01