-
Dropout right before max pooling
Why dropout followed by max pooling makes training and evaluation disagree, a fix from Wu & Gu (2015), and two toy experiments that test both.
-
Hiding emotions in plain sight
Using adversarial perturbations to hide emotion from voice assistants while speech-to-text keeps working.
-
Social navigation of a mobile robot
Planning robot paths through crowds with trajectory prediction and the social forces model.
-
Auto-ML with Auto-Keras
Letting Auto-Keras search for a CNN architecture, and how it compares with a hand-built one.
-
Classifying emotions with neural networks
Building a neural network that recognises emotions from voice, using the RML dataset and openSMILE features.