Sorry, you need to enable JavaScript to visit this website.

End-to-end learned image compression (LIC) has become promising alternatives for lossy image compression. However, deployments of LIC models are restricted, due to excessive network parameters and high computational complexity. Existing LIC models realized throughout with integer networks are significantly degraded in rate-distortion (R-D) performance. In this paper, we propose a novel fully integerized model for LIC that leverages channel-wise weight and activation quantization.

Quantization of one deep neural network to multiple compression rates (precisions) has been recently considered for flexible deployments in real-world scenarios. However, existing methods for network quantization under multiple compression rates leverage fixed-precision bit-width allocation or heuristically search for mixed-precision strategy and cannot well balance efficiency and performance.

This paper studies compression techniques for parallel in-memory sparse tensor algebra. We find that applying simple existing compression schemes can lead to performance loss in some cases. To resolve this issue, we introduce an optimized algorithm for processing compressed inputs that can improve both the space usage as well as the performance compared to uncompressed inputs. We implement the compression techniques on top of a suite of sparse matrix algorithms generated by taco, a compiler for sparse tensor algebra.

This paper proposes an effective universal "on-the-fly" mechanism for stochastic codebook generation in lossy coding of Markov sources.

In this paper, we propose a neural implementation of a companded quantization scheme allowing to train and implement optimal scalar quantization in data compression systems based on neural networks. The advantage of companded quantization lies in the fact that it allows to implement optimal non-linear quantization in a simpler form based on uniform quantization. In our work, we consider two different models of uniform quantization. Further on, in order to verify the effectiveness of the proposed approach, we made a series of experiments on natural grayscale images.

The wavelet tree is a data structure that indexes a text over an integer alphabet for efficient rank and select queries. Using the Huffman encoding, it can be stored in zero-order entropy-compressed space. We present a highly engineered open source implementation of an efficient sequential construction algorithm that makes use of bit parallelism via vector instructions. On hardware featuring ultrawide registers of up to 512 bits, it outperforms the currently fastest known practical sequential construction algorithms by a factor of up to 2.5.

It is known that a context-free grammar (CFG) that produces a single string can be derived from the compact directed acyclic word graph (CDAWG) for the same string. In this work, we show that the CFG derived from a CDAWG is deeply connected to the maximal repeat content of the string it produces and thus has O(m) rules, where m is the number of maximal repeats in the string. We then provide a generic algorithm based on this insight for constructing the CFG from the LCP-intervals of a string in O(n) time, where n is the length of the string.

Canonical binary AIFV coding contains two trees $T_0$ and $T_1$. We show the method to compress $T_0$, and the method to compress $T_1$ is with a similar way. We provide a new method to store the number of leaves, master nodes and complete internal nodes in each layer and compactly encode the string of numbers according to the specific property between the nodes.

Given a string S over an alphabet of size σ, we consider practical implementations of extended compressed RAM on S, which supports access, replace, insert, and delete operations on S while maintaining S in compressed form. In this paper, we proposed two implementations where each of them is based on the compressed RAM of Jansson et al. [ICALP 2012], and Grossi et al. [ICALP 2013], respectively. Experimental results show that our implementations support the operations efficiently while keeping the space proportional to the entropy of the input during the updates.

In this paper, we propose a novel approach to succinct coding of permutations taking advantage of the “divide-and-conquer” strategy. In addition, we provide a theoretical analysis of the proposed approach leading to formulations allowing to calculate precise bounds (minimum, average, maximum) to the length of permutation coding expressed in a number of bits per permutation element for various sizes of permutations n being integer powers of 2.