Neural network architectures have demonstrated to achieve impressive results across a wide range of different domains. The availability of very large datasets makes possible to overcome the limitation of the training stage thus achieving significant level of performance. On the other hand, even though the advancements in GPU hardware, training a complex neural network model still represents a challenge. Long time is required when the computation is demanded to a single machine. In this work, a distributed training approach for 3DPyraNet model built for a specific domain, that is the emotion recognition from videos, is discussed. The proposed work aims at distributing the training procedures over the nodes of the Intel DevCloud Platform and demonstrating how the training performance are affected in terms of both computational demand and achieved accuracy compared to the use of a single machine. The results obtained in an experimental design suggests the feasibility of the approach for challenging computer vision tasks even in presence of limited computing power based on exclusive use of CPUs.