ASG-TDM: A Graph-Enhanced Transformer With Spiking Preprocessing and a Dendritic Head for Multi-Task Facial Analysis

Facial analysis supports many real applications such as expression understanding, attribute and age estimation, and driver monitoring, and it benefits from models that capture spatial structure, temporal change and global context together. Conventional Neural Network CNNs model local texture well but struggle with long range relations, while standard vision transformers demand heavy computation and do not include facial priors. Pure spiking or dendritic models give efficiency or interpretability but are usually attached as separate modules rather than being trained in a unified system. This paper presents ASG TDM, an adaptive spiking graph transformer with dendritic modules, built as a single trainable pipeline. A light spiking front end turns short term frame differences into compact rate features. A landmark graph carries anatomical relations through a graph attention branch that runs with the transformer patch branch. A small cross attention step fuses node and patch streams. The dendritic head uses simple gates from spike and attention signals to combine branch outputs into a shared representation for all tasks. We evaluate on FER2013, RAF DB, AffectNet (eight class split), CelebA, AgeDB and a driver monitoring set, and compare against strong CNN, transformer, spiking transformer and graph guided baselines. Across five runs per setting, the model reaches 88.1% on FER2013, 92.4% on RAF DB, 65.7% on AffectNet (eight class), 96.2% mean attribute accuracy on CelebA, 3.95 mean absolute error on AgeDB and 92.3% on the driver set. Improvements over a Vision Transformer (ViT) Base baseline are consistent and statistically significant in most cases. Under thirty percent random occlusion the accuracy drop is about two points, smaller than for the non spiking baselines. Model size and Floating Point Operations FLOPs are in the same range as ViT Base and a recent spiking transformer, with a measured throughput close to real time on a single V100.