Deep-learning methods are becoming key in the analysis of neutrino physics. Current neutrino experiments are leveraging these techniques, which have exhibited to outperform standard tools in several domains, such as identifying neutrino interactions or reconstructing particle energies. In this work, we show various deep-learning algorithms used in the context of the SuperFGD, a novel 3D-granular plastic-scintillator detector that will be used to upgrade the magnetised near detector of the T2K experiment. Specifically, we present different sparse convolutional networks for tasks including particle identification and electron/gamma separation, combined with an additional neural network trained to remove hits produced by multiple particles to minimise systematic biases. On top of that, we describe the design and application of a recurrent neural network for particle tracking. All the methods above report promising results and are planned to be integrated into the reconstruction chain of the experiment.