Encoder-decoder transformer models have achieved great success on various vision-language (VL) and language tasks, but they suffer from high inference latency. Typically, the decoder takes up most of the latency because of the auto-regressive decoding. To accelerate the inference, we propose an approach of performing Dynamic Early Exit on Decoder (DEED). We build a multi-exit encoder-decoder transformer model which is trained with deep supervision so that each of its decoder layers is capable of generating plausible predictions. In addition, we leverage simple yet practical techniques, including shared generation head and adaptation modules, to keep accuracy when exiting at shallow decoder layers. Based on the multi-exit model, we perform step-level dynamic early exit during inference, where the model may decide to use fewer decoder layers based on its confidence of the current layer at each individual decoding step. Considering different number of decoder layers may be used at different decoding steps, we compute deeper-layer decoder features of previous decoding steps just-in-time, which ensures the features from different decoding steps are semantically aligned. We evaluate our approach with three state-of-the-art encoder-decoder transformer models on various VL and language tasks. We show our approach can reduce overall inference latency by 20%-74% with comparable or even higher accuracy compared to baselines.

In this paper, we describe an ongoing effort in collecting and annotating a multilingual speech database of natural stress emotion from university students. The goal is to detect natural stress emotions and study the stress expression differences in different languages, which may help psychologists in the future. We designed a common questionnaire of stress-inducing and non-stress-inducing questions in English, Mandarin and Cantonese and collected a first ever, multilingual corpus of natural stress emotion. All of the students are native speakers of the corresponding language. We asked native language speakers to annotate recordings according to the participants' self-label states and obtained a very good kappa inter labeler agreement. We carried out human perception tests where listeners who do not understand Chinese were asked to detect stress emotion from the Mandarin Chinese database. Compared to the annotation labels, these human perceived emotions are of low accuracy, which shows a great necessity for natural stress detection research.