Adaptive optimizers constitute core components for training deep learning models.However, mainstream optimizers suffer from irreconcilable inherent flaws. AdamWtends to induce over-preconditioning due to long-term accumulation of second-ordermoments, trapping models in sharp local minima and resulting in stagnant updates inthe late training stage. Muon delivers powerful capability to escape loss surfaces viaorthogonal spectral normalization, yet it lacks dimension-wise curvature adaptationand suffers from persistent oscillations in the late convergence phase. MomentumSGD achieves strong generalization and refined convergence, but it fails to cope withill-conditioned Hessian matrices and exhibits extremely slow convergence at the earlystage.
This project introduces a <b>risk-conscious perception and decision framework</b> for Advanced Driver Assistance Systems (ADAS). By combining multi-view camera inputs (rear, left, and right) with YOLO object detection, the system replaces traditional fixed confidence thresholds w…
This study proposes a THz-TDS-based framework for spectral optimisation and polyethylene (PE) ageing identification. First, a filtering-pooling and peak attention network (FPAN) is developed to mitigate water vapour interference and system noise under conventional conditions. By…
Providing a comprehensive synthetic genomic reference panel for clinical-grade diagnostic assay validation and research applications across 55 important cancer-associated genes, this project offers a full complement. The dataset consists of 100,000 pairs of normal and mutated seq…
<i>Deep learning (DL) methods show promising potential for single-cell data analysis, yet required tremendous efforts in building the models. </i><i>To streamline the application of sequence-based DL methods in single-cell genomics, we established a two-layer CNN model as a basel…