Algorithm 1: Multi-Source M-FUDA Training

Data:N labelled source domains ${\left\{D_{s_i}\right\}}_{i=1}^N={\left\{{\{x_{s_i}^{\left(j\right)},y_{s_i}^{\left(j\right)}\}}_{j=1}^{M_s}\right\}}_{i=1}^N$ and unlabelled target domain $D_t={\left\{x_t^{\left(j\right)}\right\}}_{j=1}^{M_t}$ Result: Optimized pre-trained ResNet50 $\left(F\right)$ [80], N trained domain-specific feature extractors ${\left\{H_{i1}\right\}}_{i1=1}^N$, and N trained domain-specific classifiers ${\left\{C_{i2}\right\}}_{i2=1}^N$