OpenTalks.AI - Александр Артеменков, Towards Scalable Dimensionality Reduction via Noise Contrastive Estimation

Transcript

1. 1.

   Towards Scalable Dimensionality Reduction Aleksandr Artemenkov joint work with Maxim

   Panov via Noise Contrastive Estimation
2. 2.

   News Headlines Of India 3 million topics you are unfamiliar

   with Kaggle Dataset
3. 3.

   None
4. 4.

   Just repeat 3 million times! (not a good idea)

5. 5.

   Is there a better approach? (visualization)

6. 6.

   None
7. 7.

   Noise Contrastive Estimation (without a single formula) (check out the

   poster for them)
8. 8.

   Building Representation “cold” “frosty” “cold” “warm” “cold” “first” similar different

   unrelated (noisy) ? ? ?
9. 9.

   Simple Insight • similar points = closer than predefined noisy

   distance • different points = further than predefined noisy distance Any benefits from such approach?
10. 10.

    Parallelized Batch Optimization • noisy distance is predefined • repel

    different points • attract similar points • adjustments are independent of each other to distinguish them from noise Meet NCVis! (Noise Contrastive Visualization) GitHub Repository
11. 11.

    Why use NCVis? (there are reasons) Paper

12. 12.

    None
13. 13.

    None
14. 14.

    Thank you for your attention! (find me near the poster)