Searcharxiv⌕ Search

arXiv subjects

Paweł Korus

Publications and source records attributed to Paweł Korus.

2 recordsLinked to original sources

Hard-Attention for Scalable Image Classification

Can we leverage high-resolution information without the unsustainable quadratic complexity to input scale? We propose Traversal Network (TNet), a novel multi-scale hard-attention architecture, which traverses image scale-space in a top-down fashion, visiting only the most informative image regions along the way. TNet offers an adjustable trade-off between accuracy and complexity, by changing the number of attended image locations. We compare our model against hard-attention baselines on ImageNet, achieving higher accuracy with less resources (FLOPs, processing time and memory). We further test our model on fMoW dataset, where we process satellite images of size up to $896 \times 896$ px, getting up to $2.5$x faster processing compared to baselines operating on the same resolution, while achieving higher accuracy as well. TNet is modular, meaning that most classification models could be adopted as its backbone for feature extraction, making the reported performance gains orthogonal to benefits offered by existing optimized deep models. Finally, hard-attention guarantees a degree of interpretability to our model's predictions, without any extra cost beyond inference. Code is available at $\href{https://github.com/Tpap/TNet}{github.com/Tpap/TNet}$.

cs.CV↗

Correction Trees as an Alternative to Turbo Codes and Low Density Parity Check Codes

The rapidly improving performance of modern hardware renders convolutional codes obsolete, and allows for the practical implementation of more sophisticated correction codes such as low density parity check (LDPC) and turbo codes (TC). Both are decoded by iterative algorithms, which require a disproportional computational effort for low channel noise. They are also unable to correct higher noise levels, still below the Shannon theoretical limit. In this paper, we discuss an enhanced version of a convolutional-like decoding paradigm which adopts very large spaces of possible system states, of the order of $2^{64}$. Under such conditions, the traditional convolution operation is rendered useless and needs to be replaced by a carefully designed state transition procedure. The size of the system state space completely changes the correction philosophy, as state collisions are virtually impossible and the decoding procedure becomes a correction tree. The proposed decoding algorithm is practically cost-free for low channel noise. As the channel noise approaches the Shannon limit, it is still possible to perform correction, although its cost increases to infinity. In many applications, the implemented decoder can essentially outperform both LDPC and TC. This paper describes the proposed correction paradigm and theoretically analyzes the asymptotic correction performance. The considered encoder and decoder were verified experimentally for the binary symmetric channel. The correction process remains practically cost-free for channel error rates below 0.05 and 0.13 for the 1/2 and 1/4 rate codes, respectively. For the considered resource limit, the output bit error rates reach the order of $10^{-3}$ for channel error rates 0.08 and 0.18. The proposed correction paradigm can be easily extended to other communication channels; the appropriate generalizations are also discussed in this study.

cs.IT↗