402 karma · joined March 23, 2014
Previously RStudio, The Linux Foundation, Microsoft Research, and SAP.
x.com/javierluraschi github.com/javierluraschi linkedin.com/in/javierluraschi
From the conclusion:
“We presented a practical implementation of an LSTM based lossless compressor. Although it is slow, its description is simple and the memory consumption is reasonable compared to compressors giving a similar compression ratio. It would greatly benefit from a dedicated hardware implementation.”
I only skimmed through the paper, but looks to me like this might be a 10X improvement over other LSTM compression methods, not state-of-the-art compression.
That said, I’m far from being an expert in compression and was hoping from someone in HN to explain how relevant this is in their discipline.
I still need to add support to add voxels from R and retrieve them, but should be straightforward, the hard part was putting together the rendering engine.
I built an R package to easily support nomnoml in R Markdown: https://github.com/javierluraschi/nomnoml
What is more interesting to me is that one can pull together this demo pretty easily using all the new tools that have just been made available in R. Those tools are much more interesting than this post.
Regarding, 'which is compounded by the nature of the former', the argument from the comment seems to be that creating a layer of something makes things worse, this is not necessarily the case. Abstraction layers are one of the most powerful and reliable tools in software engineering. For instance, I would much rather write a layer of C++ code that wraps assembly code. Others can choose to create layers over C++ and so on, it's all good.
If you are referring to the act of mixing "rmarkdown, keras and kerasjs" to run deep learning models inside a slide as a hack, that would make much more sense.
Could you please clarify?
Worth mentioning that this model uses a simple feedforward network with a few dense layers trained over the MNIST dataset; therefore, digit recognition classification is not very accurate.
I'll look later today into adding a tablet version that will allow them do draw over to avoid printing them. Give it another try tomorrow.
Basically, the ("lots of hardware") distributed implementation gets ~3100 points in the Elo rating against ~2900. ~2900 is still sufficient to win against Fan Hui. So I would say, that yes, this algorithm has most of the merit here.
_ _ G
G B B
_ _ G
A next step would be to use real computer vision over a 3D camera in the environment aligned to R2D@ and a deep neural network to interpret pixels instead of states. DeepMind has a paper describing this approach: http://arxiv.org/pdf/1312.5602.pdf