[nabla9]

Interestingly the Gene Regulatory Network (GRN) - the interaction of genes and proteins that control the prosess inside the cell is computationally very similar to recurrent neural network. Gene expression levels is controlled by proteins that are produced by active genes. https://en.wikipedia.org/wiki/Gene_regulatory_network

Harnessing this mechanism directly for neural computation would be grand project.

[daveguy]

I think it's a bit of a stretch to say GRN is "computationally similar" to neural networks. That seems to be a shoehorn of the most popular technology of one field into another field.

Just because the GRN contains feedback loops with multiple influences doesn't mean it's suited to NN computation. The GRN is orders of magnitude more complex than computational NNs and it is orders of magnitude slower than signal transduction of axons.

I agree it will be awesome to be able to harness molecular machinery to "do stuff" for us. The "nanotechnology" in our bodies is way beyond any current manufactured nanotechnology.

We currently have only crude control of the cellular functionality -- think point a wind up toy in the direction we want or modify it by taping a flashlight to the top. We are pretty far from being able to use the manufacturing equipment to make cars instead of windup toys. (Not to mention the manufacturer is already making rockets)

[nabla9]

> That seems to be a shoehorn of the most popular technology of one field into another field.

Recurrent neural network is used to model gene regulatory network. It's not a shoehorn.

See for example:

[1]: Reconstruction of Gene Regulatory Networks from Gene Expression Data Using Decoupled Recurrent Neural Network Model https://link.springer.com/chapter/10.1007/978-4-431-54394-7_...

[2]: Gene regulatory networks inference with recurrent neural network models https://ieeexplore.ieee.org/document/1555844/

[3]: Recurrent Neural Network Based Modeling of Gene Regulatory Network Using Bat Algorithm https://arxiv.org/pdf/1509.03221.pdf

> The GRN is orders of magnitude more complex than computational NNs and it is orders of magnitude slower than signal transduction of axons.

It's possible that we can reduce relevant complexity to the RNN subset that it useful. Feedback loop speeds are slower but they can be below second.

In many search and optimization problems the ability to run say 100 trillion large stochastic RNN's in parallel in a 100 liter tank could be huge. Especially if all you need is glucose and few cheap nutrients to power it.

[daveguy]

The articles you cite start with experimental data about Gene regulatory networks (eg from dna microarray) and then use rnn to characterize or produce the networks known or elucidated experimentally.

None of the sources claim functional equivalence of the GRN by the RNN or vice versa.

From a "big O" computational complexity perspective the gap between what you are describing and the actual case is the gap between P and NP. Just because we can confirm the results of a GRN with an RNN doesn't mean we can produce those results.

Yes, biological computing could harness very powerful parallelism. We are nowhere close to harnessing that power. (See toy manufacturing analogy)