This post summarizes

Automatic Alignment of Local Representations (2003)
Yee Whye Teh , Sam Roweis
NIPS 2003.

The paper proposes a generic method to combine coordinates obtained from many dimensionality reduction algorithms into one global coordinate, essentially “aligning” local coordinates into a global one. The motivation of such a proposal starts from the inability of linear methods (e.g., PCA) to capture the structure of curved manifolds. A common way to deal with a curved manifold is to consider a mixture of local dimensionality reducers. The idea is analogous to approximating a curve with a piecewise linear function. The problem with this approach is that there is no globally coherent coordinate because each local region has its own.

Proposal

Assume that we are given a set of points $\mathcal{X}=\left[x_{1},\ldots,x_{N}\right]^{\top}$ from $D$-dimensional space sampled from a $d\ll D$ dimensional manifold. Let $\mathcal{Y}=\left[y_{1},\ldots,y_{N}\right]^{\top}$ be the reduced images of $\mathcal{X}$ in $d$ dimensional embedding space. Suppose that we have already trained $K$ local dimensionality reducers, each producing a representation $z_{nk}\in\mathbb{R}^{d_{k}}$ of point $x_{n}$. Associated with each output $z_{nk}$ is a responsibity $r_{nk}\geq0$ describing the reliability of the representation of $x_{n}$ from the $k^{th}$ reducer. The algorithm aligns \(\{z_{nk}\}_{n,k}\) and \(\{r_{nk}\}_{n,k}\) to produce $\mathcal{Y}$.

Assume that the global coordinate is a linear function of $z_{nk}$. The paper proposes to obtain $\mathcal{Y}$ from

\[\begin{aligned} y_{n} & =\sum_{k}r_{nk}(L_{k}z_{nk}+l_{k}^{0})=\sum_{k}\sum_{i=0}^{d_{k}}r_{nk}z_{nk}^{i}l_{k}^{i}=\sum_{j}u_{nj}l_{j},\end{aligned}\]

where $L_{k}=\left(l_{k}^{0}|\cdots|l_{k}^{d_{k}}\right)$ is a linear projection associated with the $k^{th}$ reducer, and $l_{k}^{0}$ is an offset. If we let $u_{nj}=r_{nk}z_{nk}^{i}$, $z_{nk}^{0}:=1$ and vectorizing the indeces $(i,k)$ into $j$, we have \(\mathcal{Y}=UL,\) which is a linear system with fixed $U$ and unknown $L$. To determine $L$, the authors proposed to use a cost function $\mathcal{E}$ based on the locally linear embedding (LLE) algorithm: \(\mathcal{E}(\mathcal{Y},W)=\sum_{n}\|x_{n}-\sum_{m\in\mathcal{N}_{n}}w_{nm}x_{m}\|^{2}=\mathrm{tr}(\mathcal{Y}^{\top}(I-W^{\top})(I-W)\mathcal{Y}),\) where $\mathcal{N}_{n}$ is the set of nearest neighbours of $x_{n}$ and \(\sum_{m\in\mathcal{N}_{n}}w_{nm}=1\).

The local linear reconstruction weights $W$ are constructed by solving $\min_{W}\mathcal{E}(\mathcal{X},W)$ subject to the normalization constraint. With $W$ determined, $\mathcal{E}(\mathcal{Y},W)$ can be solved by minimizing \(\mathcal{E}(\mathcal{Y},W)=\mathrm{tr}\left[L^{\top}U^{\top}(I-W^{\top})(I-W)UL\right],\) subject to two constraints, $\frac{1}{N}\boldsymbol{1}^{\top}\mathcal{Y}=0$ and $\frac{1}{N}\mathcal{Y}^{\top}\mathcal{Y}=I_{d}$ to break degeneracies. Solving for $\mathcal{Y}$ now amounts to solving for $L$. The objective together with the two constraints form a generalized eigenvalue problem whose solution is given by the $2^{nd}$ to $(d+1)^{st}$ smallest eigenvectors. The alignment problem has no local optima.