Asymmetric Autoencoders: An NN alternative for resource-constrained devices in IoT networks

Local computation and communication are known challenges for energy -constrained devices that can become even more complex if we consider data acquisition with noise. Thus, developing systems that address these problems is fundamental when implementing sensing nodes in IoT networks. Fortunately, sensed data has intrinsic redundancies that allow compression with little or no information loss, which can even be used to suppress the collected noise. Many solutions using Neural Networks (NNs) have emerged to address both issues, resorting to autoencoders to extract these redundancies to reduce data transmissions in IoT networks and to remove noise from data in general. However, solutions that resort to NNs often rely on increasing the number of NN layers to achieve performance improvements, which can be tricky when deploying them in resource -constrained devices. Models with multiple layers require more space to store their parameters and more computations. To address these problems, we propose Asymmetric Autoencoders (AAEs), a model that modifies the typical autoencoder, which adopts a symmetric encoder -decoder architecture, in favour of a design that has fewer NN and other resources in the encoder than in the decoder. Our experiments with single -sensor temporal -data compression show that our proposed AAEs can offer a similar or smaller reconstruction error compared to the symmetric AEs while using encoders with fewer parameters and that require fewer floatingpoint operations (FLOPs) with each compression operation. For instance, the proposed AAEs can outperform the best symmetrical implementations by executing five to seven times fewer FLOPs. Given their inherently IoT-friendly design and positive results, we show that AAEs are a valuable model for NN deployment in sensor nodes, as they can achieve similar or better performance than symmetric autoencoders while saving sensor node resources.