Learning by active nonlinear diffusion

Figure 1.  Data colored by class label. Both the data in (a) and (b) exhibit cluster structure, which can guide active learning in the case that the labels are constant on these clusters. Indeed, on the left, using a simple clustering algorithm such as $ K $-means suggests that only 3 labels are necessary to correctly label the entire dataset. For the data on the right, many more than three labels are necessary if $ K $-means is used for the underlying clustering, since the clusters are highly elongated and nonlinear. Indeed, $ K $-means will split the annular and elongated clusters. On the other hand, if pairwise comparisons are made with distances other than Euclidean distances, it may be possible that active learning achieves near perfect results with only 3 labels. The proposed active learning scheme gains robustness to class shape via diffusion geometry, and is suitable for data in both (a) and (b)

Figure 2.  In (a) and (b), the values of $ \log_{10}( \mathcal{D}_{t}) $ are shown for synthetic geometric data from Figure 1 (b) for $ \log_{10}(t) = 2 $ and $ \log_{10}(t) = 9 $, respectively. We see that for small values of $ t $, the mode estimation incorrectly places the first three modes on the highly elongated cluster. For larger time values, the underlying random walk reaches a mesoscopic equilibrium and correct mode estimation is achieved. The emergence of mesoscopic equilibria is apparent in (c), (d), which show the matrix of diffusion distances at time $ \log_{10}(t) = 2 $ and $ \log_{10}(t) = 9 $, respectively. When $ \log_{10}(t) = 2 $, $ P^{t} $ has not mixed, and there are still substantial within-cluster distances. For $ \log_{10}(t) = 9 $, $ P^{t} $ has reached mesoscopic equilibria, so that within-cluster distances are quite small, yet between-cluster distances are still large [40]

Figure 3.  Top row: Three different synthetic datasets in two dimensions are shown, categorized as geometric, bottleneck and Gaussian. Bottom row: Plots of node purity for three different multiscale, hierarchical methods of clustering: average linkage clustering (ALC), single linkage clustering (SLC) and learning by unsupervised nonlinear diffusion (LUND). As the number of leaves/clusters increases, purity is non-decreasing. The purity of the LUND clusters converges more rapidly to the optimal value 1, indicating that high accuracy can be gained by correctly labeling a smaller number of clusters, compared to ALC and SLC

Figure 4.  In (a), data with natural hierarchical structure is exhibited. The four Gaussians have means $ (0, 0), (0, 2), \left(\frac{3}{2}, 0\right), \left(\frac{3}{2}, 2\right) $. While at one level of granularity there are 4 clusters (shown in (b)), at a coarser level of granularity the top 2 and bottom 2 Gaussians form clusters, leading to a clustering with only 2 clusters (shown in (c))

Figure 5.  The matrix of pairwise diffusion distances for $ \log_{10}(t) = 1.5 $ and $ \log_{10}(t) = 5 $ are shown in (a) and (b), respectively, illustrating the hierarchical cluster structure in the data. This hierarchical structure introduces ambiguities into the estimation of the number of clusters $ K $ in LUND, as shown (c). For small time, $ \hat{K} = 4 $, while for larger time $ \hat{K} = 2 $

Figure 6.  LAND labelings of the data with four queries under two different scenarios: small diffusion time (top row) and large diffusion time (bottom row), and four latent clusters (first column) and two latent clusters (second column). The clusters are closer in the horizontal direction than in the vertical direction, from whence the hierarchical structure is derived. In all cases, LAND is able to to correctly label the data with just four queries, one for each Gaussian

Figure 7.  Experimental results on the synthetic datasets introduced in Figure 3. We see that LAND achieves rapid convergence to perfect labeling accuracy, compared to much slower convergence for the two comparison methods

Figure 8.  The Salinas A dataset consists of $ 83 \times 86 = 7138 $ pixels in $ D = 224 $ dimensions. The image has spatial resolution $ 3.7 $m/pixel, and was recorded over Salinas, USA by the Aviris sensor. The six labelled classes are arranged in diagonal rows, and are quite spatially regular. The sum across all spectral bands is shown in (a), and the labeled ground truth is shown in (b), with pixels having the same class being given the same color

Figure 9.  Pavia data consists of a $ 270\times 50 = 13500 $ subset of the full Pavia data set. The image has spatial resolution $ 1.3 $m/pixel, and was recorded over Pavia, Italy by the ROSIS sensor. It consists of 6 spatial classes, some of which are quite well-spread out in the image. The sum across all spectral bands is shown in (a), and the labeled ground truth is shown in (b), with pixels having the same class being given the same color

Figure 10.  The active learning results for the Salinas A and Pavia datasets are shown in (a) and (b), respectively. In both cases, the LAND algorithm strongly outperforms the modified LAND variant using randomly selected training data, and the CBAL algorithm. In particular, LAND is able to achieve a significant improvement in accuracy with a very small number of labels