Efficient MATLAB Computations with Sparse and Factored Tensors

Airborne and landing phases of a simplified back somersault movement

“…In the latter part of the airborne phase, vestibular canal signals [20] are used for negative feed forward. The net effect is negative torso acceleration feedback (assuming that the head does not move relative to the body).…”

Scalable Tensor Mining

“…For example, line 3 of Algorithm 1, if naively implemented, would generate too much intermediate data; this is known as the 'intermediate data explosion' problem. Bader et al [14] proposed an efficient method for the computation by exploiting the characteristic of Kronecker multiplication, Kang et al [15,16] proposed distributed ALS algorithms for the computation using HADOOP, and Beutel et al [17] proposed a distributed stochastic gradient descent algorithm using HADOOP.…”

Parallel Algorithms for Constrained Tensor Factorization via Alternating Direction Method of Multipliers

“…flops, by exploiting sparsity and the structure of the Khatri-Rao product [27], [28], [29]. With 5nF flops, it is possible to parallelize this computation [21].…”

“…Rewriting the update of A k in terms of partitioned matrices, we obtain (27) at the top of this page. If we focus on a certain block of A k+1 in (27), then we obtain the corresponding update of the distributed algorithm (see (26)).…”

“…If we focus on a certain block of A k+1 in (27), then we obtain the corresponding update of the distributed algorithm (see (26)). We observe that the matrix inverse in the second line of (27) is common to all blocks, and should be computed once.…”