Position Prediction as an Effective Pretraining Strategy

Transformers have gained increasing popularity in a wide range of applications, including Natural Language Processing (NLP), Computer Vision and Speech Recognition, because of their powerful representational capacity. However, harnessing this representational capacity effectively requires a large amount of data, strong regularization, or both, to mitigate overfitting. Recently, the power of the Transformer has been unlocked by self-supervised pretraining strategies based on masked autoencoders which rely on reconstructing masked inputs, directly, or contrastively from unmasked content. This pretraining strategy which has been used in BERT models in NLP, Wav2Vec models in Speech and, recently, in MAE models in Vision, forces the model to learn about relationships between the content in different parts of the input using autoencoding related objectives. In this paper, we propose a novel, but surprisingly simple alternative to content reconstruction~-- that of predicting locations from content, without providing positional information for it. Doing so requires the Transformer to understand the positional relationships between different parts of the input, from their content alone. This amounts to an efficient implementation where the pretext task is a classification problem among all possible positions for each input token. We experiment on both Vision and Speech benchmarks, where our approach brings improvements over strong supervised training baselines and is comparable to modern unsupervised/self-supervised pretraining methods. Our method also enables Transformers trained without position embeddings to outperform ones trained with full position information.

Speech Emotion: Investigating Model Representations, Multi-Task Learning and Knowledge Distillation

Estimating dimensional emotions, such as activation, valence and dominance, from acoustic speech signals has been widely explored over the past few years. While accurate estimation of activation and dominance from speech seem to be possible, the same for valence remains challenging. Previous research has shown that the use of lexical information can improve valence estimation performance. Lexical information can be obtained from pre-trained acoustic models, where the learned representations can improve valence estimation from speech. We investigate the use of pre-trained model representations to improve valence estimation from acoustic speech signal. We also explore fusion of representations to improve emotion estimation across all three emotion dimensions: activation, valence and dominance. Additionally, we investigate if representations from pre-trained models can be distilled into models trained with low-level features, resulting in models with a less number of parameters. We show that fusion of pre-trained model embeddings result in a 79% relative improvement in concordance correlation coefficient CCC on valence estimation compared to standard acoustic feature baseline (mel-filterbank energies), while distillation from pre-trained model embeddings to lower-dimensional representations yielded a relative 12% improvement. Such performance gains were observed over two evaluation sets, indicating that our proposed architecture generalizes across those evaluation sets. We report new state-of-the-art "text-free" acoustic-only dimensional emotion estimation $CCC$ values on two MSP-Podcast evaluation sets.