Adversarial autoencoder for visualization and classification of human activity: Application to a low-cost commercial force plate

Adversarial Autoencoder and Multi-Armed Bandit for Dynamic Difficulty Adjustment in Immersive Virtual Reality for Rehabilitation: Application to Hand Movement

“…Hernandez et al [24] proposed a data dimensionality reduction approach to visualize high-dimensional data from the Wii Balance Board during upper and lower body exercises. This study uses a data dimensionality reduction approach with deep learning models called adversarial autoencoder to visualize time series in a 2D latent space.…”

“…In this study, two different approaches are considered. The first one considers an Unsupervised AAE (UAAE) model trained in a fully unsupervised manner with p(z) defined as a set of six two-dimensional Gaussian distributions representing N = 6 movements with the ith movement as z i ∼ N (µ, σ 2 ) [24,31]. The second approach considers a Semi-Supervised AAE (SSAAE) approach by integrating the label information at the input of the discriminator Equation ( 9) during the adversarial phase according to the following modification for the discriminator:…”

“…The prior distribution used to train the AAE consists of six two-dimensional Gaussian distributions uniformly distributed in a polar coordinate system, and its shape is similar to a polar rose [24].…”

“…Finding the proper combination of hyperparameters is important because deep learning models can have different degrees of performance depending on their hyperparameters. Various combinations of hyperparameters were tested during a hyperparameter search performed by a genetic algorithm (GA); studies [24,33] indicate the hyperparameters considered.…”

“…The original shape of the signal is represented in a lower-dimensional space that preserves the similarity between the motions. This allows for the creation of a simplified visualization that allows the current motion to be represented among other motions and helps detect outliers and patterns [24]. In this study, each motion is represented by a single point in a two-dimensional latent space generated by AAE, which is a powerful and efficient approach to perform the dimensionality reduction of nonlinear data [24].…”