• Introduces quantum-inspired methods for modeling sequential data via tensor networks. • Uses bipartite graph representations of matrices to visualize linear maps. • Applies quantum probability concepts to estimate unknown distributions from limited samples. • Demonstrates approach on bitstring and natural language datasets. • Generates new sequences by sampling from learned quantum-state distributions. • Paper available on arXiv, collaboration between Flatiron Institute and CUNY.
Article Summaries:
- A new paper, “Modeling sequences with quantum states: a look under the hood,” proposes a generative modeling framework that treats probability distributions over finite‑alphabet sequences as quantum states. The authors, the author of the blog post, Miles Stoudenmire (Flatiron Institute) and John Terilla (CUNY/Tunnel), show how linear‑algebraic tools-bipartite‑graph representations, tensor‑network diagrams, and quantum‑probability concepts-can be used to estimate an unknown distribution from sample data and then generate new sequences. They illustrate the idea with bit‑string and natural‑language examples, outline a training algorithm, and discuss theoretical bounds on model performance given a limited training set.
Sources: