Seminar for
DECISION MAKING – THEORY, TECHNOLOGY AND PRACTICE

PROGRAM

Predavanja možete pratiti i online putem MITEAM stranice Seminara Odlučivanje - teorija, tehnologija, praksa:
https://miteam.mi.sanu.ac.rs/asset/sEL32w8mjmruyeEqW

Plan rada Seminara Odlučivanje - teorija, tehnologija, praksa za APRIL 2026.

Četvrtak, 16.04.2026. u 13:00, Pariske Komune bb, Niš i Online
Aleksa Bogdanović, Mathematical Institute of the Serbian Academy of Sciences and Arts
GROUPED DEPTHWISE SEPARABLE TEMPORAL CONVOLUTIONAL NETWORKS WITH SQUEEZE-AND-EXCITATION ATTENTION FOR EFFICIENT MULTIVARIATE FORECASTING
TCNs have been shown to be highly effective alternatives to RNNs in sequence modeling tasks. They provide stable training, parallelization, and efficient learning of long-term dependencies using dilated convolutions. However, their classical implementation has some inefficiencies, such as large-width convolutional layers that significantly increase computational complexity with increasing network depth. To solve this problem, by redesigning the inner workings of a TCN block, grouped depthwise separable filters are employed that divide the learning process into two parts: channel-wise spatial filters followed by pointwise cross-channel mixing filters, resulting in a drastic reduction in the number of parameters without sacrificing representational capacity. Furthermore, the Squeeze-and-Excitation technique introduces an adaptive weighting scheme for channel-wise information at each block, allowing the model to selectively attenuate less relevant channels while enhancing more important ones. The proposed method retains all beneficial properties of the classical TCN framework, while achieving up to multiple times fewer trainable parameters than the baseline.