Training Deep Neural Networks with Different Datasets In-the-wild: The Emotion Recognition Paradigm

Abstract

A novel procedure is presented in this paper, for training a deep convolutional and recurrent neural network, taking into account both the available training data set and some information extracted from similar networks trained with other relevant data sets. This information is included in an extended loss function used for the network training, so that the network can have an improved performance when applied to the other data sets, without forgetting the learned knowledge from the original data set. Facial expression and emotion recognition inthe-wild is the test bed application that is used to demonstrate the improved performance achieved using the proposed approach. In this framework, we provide an experimental study on categorical emotion recognition using datasets from a very recent related emotion recognition challenge.

D-PAttNet: Dynamic Patch-Attentive Deep Network for Action Unit Detection

“…Li et al (2017) propose an adaptive region cropping based multi-label learning with deep recurrent net, which is based on combining region-based CNN (RCNN) with LSTM. Although a few deep approaches considering dynamics for AU detection have been proposed, many efforts have been devoted to incorporate dynamics in deep models for emotion recognition (Fan et al, 2016;Vielzeuf et al, 2017;Kollias and Zafeiriou, 2018;Liu et al, 2018;Lu et al, 2018). However, focusing on detecting action units is crucial since FACS is a comprehensive, anatomically-based system which describes all visually discernible facial movement and provides an objective measure.…”

Unified deep learning approach for prediction of Parkinson's disease

“…Transfer learning is another approach usually adopted in deep learning methodologies [12, 13], according to which the DNN model trained with the original data set is used to initialise DNN re‐training with the new data set. However, a serious problem then arises: as the refined DNN learns to predict from the new data set, it tends to forget the old data that are not used in the retraining procedure; this is known as ‘catastrophic forgetting’.…”

Predicting Parkinson’s Disease using Latent Information extracted from Deep Neural Networks

“…Normally, the DNN (CNN, or CNN-RNN) training is performed through minimization of the error criterion in Eq. (21) in terms of the DNN weights:…”

“…It is desirable that the G(i, j) term -with a respective value of z(i, j) equal to one -is minimized, whilst the G(i, j) values -corresponding to the rest of the z(i, j) values, which are equal to zero -are maximized. Similarly to [21], we pass G(i, j) through a softmax f function and subtract its output from 1, so as to obtain the above-described respective minimum and maximum values.…”