Geoffrey Hinton on Future Deep Learning Excitements: Machine Translation and Semantic Document Search

Geoffrey Hinton expresses his current excitement in deep learning, particularly focusing on two key areas: machine translation and document understanding. While he hasn't directly contributed to machine translation recently, he notes that his students are actively involved, indicating ongoing research and development in this field. He anticipates significant, short-term improvements in the quality and capability of machine translators. Hinton also highlights the immense potential for progress in document understanding over the next five to ten years. He envisions a future where search engines move beyond simple keyword matching to a more sophisticated semantic understanding of content. This shift would allow users to search for documents based on their underlying themes and claims, rather than just the specific words they contain. This advanced capability would represent a fundamental change from current search paradigms. Instead of inputting specific keywords and retrieving documents containing those exact terms, users would be able to query for concepts or arguments, and the system would identify relevant documents even if they express those ideas using different vocabulary. This semantic search would enable a much deeper and more intuitive interaction with information. The broader implication of these advancements is a significant enhancement in how humans access and process information. Better machine translation would break down language barriers, facilitating global communication and knowledge sharing. Semantic document search would empower users to extract meaning and claims from vast datasets more efficiently, leading to improved research, decision-making, and overall knowledge discovery, making information retrieval "kind of useful" in a profoundly new way.

I am excited about this machine translation

some of my students are doing it

I think over the fairly short term we're going to get much better translators

I am excited for things like understanding documents

I think there's going to be huge progress there over the next five to ten years

instead of giving Google some words and they they'll find you all the documents have those words in you give Google some themes and it'll find the documents that are saying that even if they're saying it in different words

you're going to be able to search for documents by what it is the documents are claiming independent whatever the works are

that would be kind of useful