SETAREH Transform
A sorting-assisted Haar wavelet energy compaction transform used for structured sparsification of high-dimensional input vectors. By sorting and applying the Haar wavelet transform, it concentrates energy and truncates small coefficients, thereby reducing the number of dimension pairs involved in computation while approximately preserving inner-product accuracy.
Timeline (last 90 days)
The full name of the SETAREH transform is "Sorting-assisted Energy-concentrating Transform for Approximate Representation using Haar wavelets."
The goal of SETAREH is, given a fixed high-dimensional coefficient vector and a series of input vectors, to approximate their inner products using fewer coefficient-input pairs.
SETAREH applies exactly the same sequence of permutations and Haar transforms to the coefficient vector and the input vector, so the inner product is exactly preserved before truncation.
SETAREH uses a convolutional neural network for apparent age estimation on the APPA-REAL dataset, and experiments show that inner products can be approximated with far fewer transform coefficient-input pairs than the original vector dimension while keeping error within an acceptable range.
Sorting and transformation require storing or transmitting permutation metadata, and when input vectors change frequently, the metadata overhead may offset part of the savings.
SETAREH belongs to data-side compression: model parameters remain fixed, but each incoming input vector undergoes a structured transform and sparsification, reducing the dimension pairs that actually participate in inner-product computation.
SETAREH and model-side compression (such as pruning, quantization, and knowledge distillation) are not mutually exclusive and can be used in combination.
The experimental scale that SETAREH has currently made public is limited, including only a synthetic Gaussian baseline and a single real task (apparent age estimation).
The paper does not discuss end-to-end accuracy loss and error propagation in depth.
All Facts (9)
The full name of the SETAREH transform is "Sorting-assisted Energy-concentrating Transform for Approximate Representation using Haar wavelets."
50%UnverifiedThe goal of SETAREH is, given a fixed high-dimensional coefficient vector and a series of input vectors, to approximate their inner products using fewer coefficient-input pairs.
50%UnverifiedSETAREH applies exactly the same sequence of permutations and Haar transforms to the coefficient vector and the input vector, so the inner product is exactly preserved before truncation.
50%UnverifiedSETAREH uses a convolutional neural network for apparent age estimation on the APPA-REAL dataset, and experiments show that inner products can be approximated with far fewer transform coefficient-input pairs than the original vector dimension while keeping error within an acceptable range.
50%UnverifiedSorting and transformation require storing or transmitting permutation metadata, and when input vectors change frequently, the metadata overhead may offset part of the savings.
50%UnverifiedSETAREH belongs to data-side compression: model parameters remain fixed, but each incoming input vector undergoes a structured transform and sparsification, reducing the dimension pairs that actually participate in inner-product computation.
50%UnverifiedSETAREH and model-side compression (such as pruning, quantization, and knowledge distillation) are not mutually exclusive and can be used in combination.
50%UnverifiedThe experimental scale that SETAREH has currently made public is limited, including only a synthetic Gaussian baseline and a single real task (apparent age estimation).
50%UnverifiedThe paper does not discuss end-to-end accuracy loss and error propagation in depth.
50%