SDRs/NLP at constellation.ai
Current & Past Projects | | Links: Constellation.ai | Paper
I have developed a generic library transforming multiple data types into Sparse Distributed Memories (binary high dimensional vectors). Code became semantic engine.
I have developed a generic library transforming multiple data types into Sparse Distributed Memories (binary high dimensional vectors). Code became semantic engine.
The technology was applied as:
- semantic agent for NLP (with better performances than W2Vec before universal sentence encoders)
- Fast retrieval with SDRs and clustering
- Improved performances on the RNN architecture (pytorch, LSDM)
In order to demonstrate the technology, I have applied it on benchmarks
- NLP classification when merged with NNets (We beat google research on MRDA dataset)
- Good performance on various multinomial datasets as well
Furthermore, we also had a potential publication written when I left. The main result is that, in many cases, we have found that classification scores could be improved preprocessing the data with HD computing.
