Multimodal AI, especially the sub-field of visual question answering (VQA), has made a lot of progress in recent years. Multimodal systems, with access to both sensory and linguistic modes of intelligence, process information the way humans do.
What is multimodal interaction?
As human beings, we experience the world as multimodal: we can feel texture, hear sounds, see objects, smell odours and taste flavours.
However, standard AI systems are usually unimodal, meaning they are trained to do a specific task such as processing images or languages. The systems are fed a single sample of training data, from which they are able to identify corresponding images or words. While it is easier to work with a single source of information, it also means that the software lacks the context and supporting information to make the best possible deductions.
The advancement of AI relies on its ability to process multimodal signals simultaneously, just like humans. For instance, while we can understand how the meaning of texts and images change when they are paired or juxtaposed with each other, a unimodal AI system would be unable to infer how the meaning of an image could change when placed next to a contradictory piece of text.
Multimodal learning, on the other hand, pieces together disjointed data (collected from different sensors and data inputs) into a single model. Since multiple sensors are used to observe the same data, multimodal learning offers more dynamic predictions compared to a unimodal system–processing more datasets translates to more intelligent insights.
- Researchers at the Allen Institute of Artificial Intelligence, AI2, created an AI model capable of producing images from text captions.
- At the University of North Carolina, Chapel Hill, researchers improved on the reading comprehension of existing language models by incorporating images.
- Google AI created MURAL (“Multimodal, Multitask Representations Across Languages”) in order to combat the lack of direct translation between different languages across the world. To make translations more accurate, the software used multitask learning to match images to text—covering over 100 languages.
- Meta developed AV-Hubert— a speech recognition system with the ability to filter out background noise to better decipher the speaker’s voice. The model is 75% more accurate than existing models trained on an equal number of transcriptions.
- Open AI introduced DALL.E, a 12-billion parameter version of GPT-3 to generate images in response to text descriptions, utilising a dataset of text-image pairs. The model can anthropomorphize animals and objects, connect unrelated concepts in plausible ways, apply alterations to existing images, and make visual representations of text.
Sophisticated multimodal systems have multiple applications across industries including aiding advanced robotic assistants, empowering advanced driver assistance and driver monitoring systems, and extracting business insights through context driven data mining.
Over time, advancements in multimodal learning could also help overcome some of AI’s present challenges. For instance, Meta developed the Hateful Memes Challenge to combat harmful multimodal content. Many of the problems surrounding AI’s inability to understand context, could be combated with the development of strong multimodal algorithms.