The goal is to demonstrate hyperdimensional computing with a simple example.
When working in hyperspace (say 10000 dimensions), some interesting properties arise. Example:
- We can pick any number of random vector from that space and each one will be approximately orthogonal to all other ones previously picked.
- We can ADD two vectors together to obtain a third vector that is similar to the original ones.
- We can multiply two vectors to obtain a third vector that is dissimilar to the original ones.
- We can reconstruct the meaning of a vector by taking any 60% of its values (will simply be noisier).
Using these properties, we can encode patterns, correlate them and compare them easily, while distributing information througout 10K values (holistically), therefore having a highly redundant and robust system.
The example will show how the algorithm can guess the language of any phrase or text from the set of languages that it has previously learned.
You can modify the test file "data/testing/test_file" content with your own input. The languages available are English and French.
The algorithm for this example is as follows:
- Encode each letter: generate a random 10K bipolar vector (1s and -1s) for each letter and store them in a map.
- Encode trigrams using rotate and multiply: trigrams are sequences of 3 letters. Example, "something" will generate trigrams: "som", "ome", "met", "eth", "thi", "hin", "ing".
- Encode language profile: add each of the trigrams to form a similar vector to all of the trigram vectors.
- Repeat for any other language and for the test string.
- Compare the test vector with each language vector using cosine similarity.
docker build -t hyperdimensional-computing .
docker run --rm hyperdimensional-computing- Paper on hyperdimensional computing by Pentti Kanerva: http://www.rctn.org/vs265/kanerva09-hyperdimensional.pdf
- Another repository implementing language recognition: https://github.com/abbas-rahimi/HDC-Language-Recognition