Funny to see this in light of our conversation yesterday [1]

This should be read with extreme skepticism. I have experienced first hand how 'science' is done in the Turaga lab. I have on multiple occasions been pressured to cut corners and do shady things in the name of results.

In light of my experience, I require extra-extraordinary evidence to believe anything coming out of that lab....

[1] https://news.ycombinator.com/item?id=17306673

Well, also this claim is definitely something that would represent an extraordinary advance. Any detailed correspondence between machine learning neural nets and actual neurology would be quite a step and not something neurologist or machine learning experts expect.

Neurology is the study of disorders of the nervous system. This is theoretical neuroscience. FTFY.

Not entirely true, as highlighted by a CVPR 2017 key note talk modeling/comparing neural activations to deep learning activations has been done already so this is not wholly novel: https://www.youtube.com/watch?v=ilbbVkIhMgo

It's theoretical neuroscience. If you have philosophical difference write a position paper.

?? I don't have a philosophical difference. I just don't believe they did their due diligence. Based on my personal experience in the lab where I was constantly pressured to do garbage science in the name of shiny results...

This sounds a little like a lab I was once part of. I’m now in an entirely different field of research.

That's a philosophical difference. I don't disagree, but given the various constraints involved what you're asking isn't possible. Additionally, I'd note that it sounds like they are practicing Lean methodology - which is justifiable.

https://plato.stanford.edu/entries/scientific-method/#MetPra

If I work in a kitchen with unsanitary working practices, does that mean I have a philosophical difference with the restaurant because I don't want to eat their filthy food?

The comparison falls flat in too many ways.

I'm sorry, what am I asking?

In your own words, you are asking for 'science'. And if you click the link through to the SEP, you'll see that it's well documented that this is, in the words of Stephen Wolfram, A New Kind of Science.

If you're interested in this kind of thing, check out Openworm and their c302 neural network.

http://docs.openworm.org/en/0.9/Projects/muscle-neuron-integ...

This is not dissimilar to research done on artificial regulatory networks driving paterning in drsosophila. Turns out the topological circuitry, not the biological details explains the behavior. The network actually didn’t work initially until they added a then unknown gene that they later discovered. They even used this artificial network to predict phenotypic mutations that were later confirmed emirically.

One of the reasons why we have such a problem getting our heads around neural networks is that we don’t test and remove the spurious interactions in our topological visualization. Do that and the underlying circuit will reveal itself in the same way that any second year EE can identify the circuit topology of a 3-bit adder. Neural networks don’t have binary logic gates, rather you can have a large number of inputs and it works on a threshold basis.

> Our work is the first demonstration, that knowledge of the connectome can enable in silico predictions of the functional properties of individual neurons in a circuit, leading to an understanding of circuit function from structure alone.

That's kind of huge. If they're right, the connectome from a real organism is a good geometry for an ANN trained to do the same job.

Two kind of performance metrics they seem to suggest on top of accurate tracking:

+ stability of the network, the neurons should not change their function willy nilly

+ robustness against multiplicative noise

I'm curious if that might lead to a more physiological plausible network.

This is really the holy grail that we will be able to translate the connectome into a software representation of it and be able to quickly acquire working neural networks from living organisms.

How do they measure the synaptic weights? I didn’t think even electron imagery could see synaptic receptors and neurotransmitters.

They don't. It's purely topology, for which they use EM [1]

Uncle Howard has spent a lot of money on modeling the fly connectome at this point!! Pressure is on to show that connectome data is actually useful.

I am skeptical. It's not not useful. But it's just one piece of the puzzle...

[1] https://www.janelia.org/project-team/flyem

You can see synaptic receptors (T-bars and post-synaptic densities) in EM, particularly if the staining is good. Neurotransmitter identification, not so much. So you combine the EM connectomics with other studies like RNA sequencing or Fluorescent In-Situ Hybridization. In insects, you can have multiple transmitters with multiple receptors (possibly differing in sign) in the same cell, so there's that complication.

Using recurrent neural nets to map and train ai based on part of the visual connectome of a fly is interesting but lacks scalable experimental feasibility. Every new neuron in the algorithm introduces an exponential amount of complexity to the ai which takes away from it's scalability.

Is this implying these flies use gradient descent?

No.

Does anyone have a link to the supplemental material for this?

I can't find one, and weirdly the NIPS 2017 site doesn't seem to know about this paper either... https://nips.cc/Conferences/2017

so I can say that they may submitted it to arxiv using a nips LaTeX styling template - even if it wasn't in NIPS (could have been rejected or something else and they forgot to change the style template)