Privacy Enhanced Multimodal Neural Representations for Emotion Recognition

Authors: Mimansa Jaiswal, Emily Mower Provost7985-7993

AAAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate this paradigm on multiple datasets and show that we can improve the privacy metric while not significantly impacting the performance on the primary task.
Researcher Affiliation Academia Mimansa Jaiswal, Emily Mower Provost University of Michigan {mimansa, emilykmp}@umich.edu
Pseudocode No The paper describes the network architecture and training process in text and with a diagram, but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about releasing source code or provide links to a code repository for the methodology described.
Open Datasets Yes We use four common emotion recognition datasets: MSPImprov (Busso et al. 2017), MSP-Podcast (Lotfian and Busso 2017), Interactive Emotional Dyadic MOtion Capture (IEMOCAP) dataset (Busso et al. 2008), and Multimodal Stressed Emotion (Mu SE) dataset (Jaiswal et al. 2019).
Dataset Splits Yes We use validation samples for hyper-parameter selection and early stopping. The hyper-parameters that we use for the main network include: number of convolutional layers {3, 4}, width of the convolutional layers {2, 3}, number of convolutional kernels {32, 64, 128}, number of GRU layers {2, 3}, GRU layers width {32}, number of dense layers {1, 2}, dense layers width {32, 64}, GRL λ {0.3, 0.5, 0.75, 1}. ... We report the average across five-fold speaker-independent crossvalidation in Table 1a and Table 1b.
Hardware Specification No The paper does not provide any specific hardware details (like GPU/CPU models, memory, or specific computing environments) used for running the experiments.
Software Dependencies No The paper mentions implementing models using 'the Keras library' but does not specify a version number for Keras or any other software dependencies.
Experiment Setup Yes The hyper-parameters that we use for the main network include: number of convolutional layers {3, 4}, width of the convolutional layers {2, 3}, number of convolutional kernels {32, 64, 128}, number of GRU layers {2, 3}, GRU layers width {32}, number of dense layers {1, 2}, dense layers width {32, 64}, GRL λ {0.3, 0.5, 0.75, 1}. For the adversarial emotion classification setups, we use the hyper-parameters that maximize the validation emotion classification performance while minimizing the validation gender classification performance. For the attacker s model, we use the following hyper-parameters: number of dense layers {2, 3, 4}, dense layers width {32, 64}.