7 ms·
It's hard to do that with human hearing, but we can use neural networks to model the sound of their conversations and build a language model. The purpose of a l
by visarga 10y ago
It's hard to do that with human hearing, but we can use neural networks to model the sound of their conversations and build a language model. The purpose of a language model is to predict the next sound based on the past. If it learns to do this precisely enough, then we have access to a transcription from "dolphin audio" to word vector representations. Then we need to match those word vectors to concepts based on observing how they correlate to their actions.
- xg15 10y agoI agree building a language model would be the first step - and we need to apply some mapping to outside observations if we want to assign some meaning to the model. But especially the last step seems hard to me. Humans often do one thing and talk about something completely different at the same time. Even if their actions and conversation are related, you usually try to avoid redundancy and will talk about things that the recipient can not already infer from your body language and/or actions. This would make correlation hard for human languages. I don't see any rrason to believe that dolphins are any nicer in that regard. What I found more promising was a controlled experiment performed a time ago, where pairs of dolphins were trained to think up and perform a choreography. At some point, the dolphins would have to coordinate how to swim - and researchers were apparently able to pinpoint some particular "conversations" that looked like they were just that. So you might have some better chances to analyze those conversations because you can make concrete assumptions about their content. Still, even then you're kind of behaving like an alien that tries to learn both "the culture of humans" and the complete english language at the same time - by observing a handful of soccer games.
- visarga 10y agoThen we could identify concepts in those conversations and play them out on a speaker, to see if they do what we thought the sound means.