Input: set of points or set of connected lines, non-intersecting, simple, convex, closed polygon
Output: another set of points or set of connected lines, open/branched polygons possible
Goal: Like PCA (Principal Component Analysis) or Curve fitting, for given set of points, wish to compute fitting curve (polygon: lines or points).
Wish to know if there is any Neural Network (Deep Learning) architecture suitable to such tasks.
This is Sequence to Sequence model. Input is sequence of variable length points or lines (m). Output is variable length points/lines (n). m may not be same as n, mostly less.
For each set of input/output, m and n would be different. ie. m1 is not same as m2.
RNN and LSTM seem to be for fixed length inputs/outputs. Another constraint is that I cannot use bucketing or padding to make input/output fixed length, as there are not good PAD values. PAD values like (0.0, 0.0) can be considered as VALID in set of given points.
Please suggest good neural network model. Any references/papers?
Dear Yogesh,
A suggestion may be combining Spectral Embedding methods:
- https://pdfs.semanticscholar.org/912a/60967992f4126f223a288e2c0a77b975327b.pdf
with Supervised/Unsupervised Kernel Regression ones:
- http://www.cs.utoronto.ca/~rfm/pubs/ukr_pami.pdf
- https://pub.uni-bielefeld.de/download/2305711/2305714 , page 34
in [Sparse] Deep Architectures ( https://arxiv.org/pdf/1312.6203.pdf )
An alternative solution (not preferred) may be combining multiple kernel regression-based ANNs by means of Ensemble Learning.
Finally, here is a paper of interest about computing representations (i.e., inference), and the geometrical connections between representation learning, density estimation and manifold learning: http://www.cl.uni-heidelberg.de/courses/ws14/deepl/BengioETAL12.pdf
Regards
Hello Stephane,
Thank you so much for such a detailed suggestion. I will surely go through the links and try to follow the methodology mentioned therein.
Thanks again,
Yogesh
You are welcome.
If you are interested in, some sources of information about the limitations of Deep Learning are available by following https://www.researchgate.net/post/What_are_the_limitations_of_deep_learning
Stéphane
- Badges
- Science topic