Large-scale representation learning from visually grounded untranscribed speech
This work addresses the challenge of visually grounded language learning for AI systems, representing an incremental advance with strong specific gains in audio-image retrieval.
The paper tackled the problem of learning multimodal representations from untranscribed speech and images by developing a scalable method to generate audio for image captioning datasets and using a dual encoder with a masked margin softmax loss. The result was state-of-the-art performance on the Flickr8k Audio Captions Corpus, improving recall in the top 10 from 29.6% to 49.5%.
Systems that can associate images with their spoken audio captions are an important step towards visually grounded language learning. We describe a scalable method to automatically generate diverse audio for image captioning datasets. This supports pretraining deep networks for encoding both audio and images, which we do via a dual encoder that learns to align latent representations from both modalities. We show that a masked margin softmax loss for such models is superior to the standard triplet loss. We fine-tune these models on the Flickr8k Audio Captions Corpus and obtain state-of-the-art results---improving recall in the top 10 from 29.6% to 49.5%. We also obtain human ratings on retrieval outputs to better assess the impact of incidentally matching image-caption pairs that were not associated in the data, finding that automatic evaluation substantially underestimates the quality of the retrieved results.