Рет қаралды 103
The singular value decomposition of a square (real, complex, or quaternionic) matrix A is a factorization A=UDV where U,V are unitary matrices and D is a diagonal matrix with positive non-increasing entries. The diagonal entries of the singular value decomposition are known as the singular values of A, and these singular values are unique. Whenever the diagonal entries are distinct (this happens almost everywhere), the unitary matrices U,V are unique as well.
On the left side of this visualization, we show the real matrix A where A changes over time, and on the right side, we show the orthogonal matrix U in the singular value decomposition U.
In the visualization, the matrix A begins as a random real matrix with independent standard Gaussian entries, but for the second half of the visualization, the matrix A gradually becomes more symmetric.
While the singular value decomposition is always unique, this decomposition is unstable
The gradients of the eigenvalues and eigenvectors of a matrix were computed in the 1985 paper ON DIFFERENTIATING EIGENVALUES AND EIGENVECTORS by Jan Magnus. We can apply those calculations to produce expressions for the gradients of the singular vectors of matrix.
We observe that in this visualization, much of the instability of the singular value decomposition is a result of two singular values swapping their ordering, but this is not the only cause of instability. In a future visualization, I will probably show how much the swapping of the ordering of singular values contributes to the instability of the singular value decomposition.
The notion of a singular value decomposition is not my own. I am making this visualization to demonstrate a potential stability issue with the singular value decomposition. In addition to stability issues, one also encounters difficulty in generalizing the notion of a singular value decomposition to higher order objects such as tensors, collections of square matrices, and quantum channels/completely positive superoperators.
It is known that if A(t) is an analytic Hermitian valued matrix, then there are analytic vector valued functions r_1,...,r_n and analytic functions s_1,...,s_n where (r_1(t),...,r_n(t)) is the set of all eigenvectors of A(t) and (s_1(t),...,s_n(t)) are their corresponding eigenvalues counting multiplicity, and r_1,...,r_n,s_1,...,s_n are essentially unique by analytic continuation. In other words, if the matrix A varies smoothly enough, then we can make its singular vectors also vary smoothly.
Unless otherwise stated, all algorithms featured on this channel are my own. You can go to github.com/spo... to support my research on machine learning algorithms. I am also available to consult on the use of safe and interpretable AI for your business. I am designing machine learning algorithms for AI safety such as LSRDRs. In particular, my algorithms are designed to be more predictable and understandable to humans than other machine learning algorithms, and my algorithms can be used to interpret more complex AI systems such as neural networks. With more understandable AI, we can ensure that AI systems will be used responsibly and that we will avoid catastrophic AI scenarios. There is currently nobody else who is working on LSRDRs, so your support will ensure a unique approach to AI safety.