Aurelie Georgette Geraldine Herbelot
Using cognitively-inspired algorithms for small AI
Friday 31 December 2021
Can we use inspiration from biological algorithms to create more efficient AI? Fruit flies have got an amazing hashing algorithm for doing conceptual similarity, which has been tested in information retrieval settings by Dasgupta et al (2017): fruit flies, the authors showed, can improve Web search.
The concept representations used in the study above, however, were already pre-computed semantic vectors, meaning that the fruit fly algorithm was really only used as a dimensionality reduction method over quality input. In the meantime, fruit flies have been shown to be able to learn word meanings from scratch, given an input corpus. An interesting question is whether their ability also reaches to learning document representations – leading the way for more efficient indexing of Web content.
In this project, you will test whether the fruit fly's algorithm can produce document representations that rival the state of the art in deep learning.