Deep Molecular Programming

arxiv.org

130 points·groar··11 comments

A warning shot.

I'm bracing for what I think will be rapid acceleration in joint organic-machine experimentation going forward.

Over the next decade, I surmise it's likely we will see everything ranging from constructing and training large-scale organic/molecular deep neural networks with stochastic gradient descent (which is what these guys have done at a small scale) to integrating artificial networks with organic ones (which is what startups like Neuralink are doing) to who knows what else.

We live in interesting times.

This paper mentions CRN++, a compiler that translates an imperative language into Chemical Reaction Networks. I had no idea this field existed. This stuff is mind bogglingly cool!

https://arxiv.org/abs/1809.07430

I skimmed the paper but I still wonder how a (practical) memory would be implemented, and how many bits it could contain.

I think these computations all happen in a "soup", and this means that the memory can only exist as a superposition of concentrations, which I suppose has its limits.

Nucleic acids store bits pretty well.

Programmable matter. Now we are getting closer to the future we were promised.

Well, what do you think life is? It’s a product of chemical computation.

Until when do I have to wait for an Organic Raspberry Pi to play with?

How can a binary-valued network be differentiable?

The loss function needs to be differentiable. The paper uses square hinge loss which is differentiable.

Thanks, and for those who are wondering how squared hinge loss is differentiable, I found:

https://www.quora.com/Why-is-squared-hinge-loss-differentiab...

future virus and worms will be scary...

Reply on news.ycombinator.com