Hacker News Reader

Comment by gardenfelder | Hacker News Reader

The piece is chock full of interesting findings, using terms biologists routinely use. But, do those terms they use, e.g. neuron, synapse mean the same thing they do for biologists? For instances, we know that synapses can be one of excitatory or inhibitory, and we know that neurons are bathed in a wash of hormones. Neurons make hormones which serve other functions throughout the brain. For instance "Neuron-Derived Estrogen Regulates Synaptic Plasticity and Memory" [1]. How does the linked work stack up against that?

[1] https://www.jneurosci.org/content/39/15/2792

View on HN

Short answer is no. The field is full with tenuous analogies. Then again „Neural Networks“ are also at best metaphorically related. More accurate existing Neuron models are actually also plagued by lots of limitations among them that they are typically implemented in 3 ancient domain specific languages with lots and lots of hardcoded constants copied from research papers.

spiking neural networks are interesting though, and new computational substrates that allow for experiments at larger scales could produce some interesting results.

today's sum n' squash (sometimes not even squash) graph networks were just kind of a curiosity before gpus turned them into a new very successful computational paradigm. maybe we'll see something similar with these high element count optical spiking graphs, even if they aren't great approximations of the real biology.

i like to think that a new analog computational substrate (or mixed analog and digital system) will be what drives the next leap in machine computation.

I'm excited to see where spiking NNs go. Something like that is needed to progress now that conventional NNs on GPUs are pretty much tapped out (in terms of non-incremental advances) from their power consumption and the end of Moore's law. Things really do need to be more hardware-efficient.

> But, do those terms they use, e.g. neuron, synapse mean the same thing they do for biologists

They are not meant to. This is not "brain simulation" or similar - which exists, but is a different matter. This context is instead about neuromorphic computing, as hardware implementation of components for Artificial Neural Networks. And results seem to be remarkable:

> They calculated that the synapses are capable of spike rates exceeding 10 million hertz while consuming roughly 33 attojoules of power per synaptic event (an attojoule is 10-18 of a joule)

The comparison with biological neuro-transmission is just indicative - for trivia, for curiosity.

Edit:

on the contrary, these devices aim to be in a way simpler than ANN's neurons (far from aiming to be as complex as cerebral neurons):

> By only rarely firing spikes, these devices shuffle around much less data than typical artificial neural networks and, in principle, require much less power and communication bandwidth

That is because the underlying aim is to achieve using a single photon for communication, with an immediate potential practical use in ANNs.

We should also note that the article only references the over-dramatic comparison to biological neurons in passing in the first paragraph (wonder if it is an "author gets it, but doesn't get to pick the title" type problem).

All that, more, and it seems like every computational biology analogy just completely forgets about the most common cell type in the brain: astrocytes. And then there are things like axo-axonal transmission that totally blow up the simple models, https://www.cell.com/neuron/fulltext/S0896-6273(22)00656-0

Can you elaborate on how the axo-axonal transmissions blow up our simple models? I couldn't understand anything from that summary.

Would it be the equivalent of edges communicating between each other in artifical neural networks?

Biologists just won't allows us to have any fun. It is always this kind of rhetoric: "what, are you modeling the brain without considering the influence of <insert obscure type of cell> on the hormone regulated blood flow around ion pump circuits during chinese new year neuron firing patterns? you are obviously bounded to fail..."

This whole AI field keeps on failing because people like to overthink things. Did Michelangelo need to know molecular chemistry to make sculptures? Why do people pretend there is no artistic component to building AI? Rant finished.

I don't think this is the case; the whole field of AI seems pretty healthy at the moment, not failing, and not all that worried about the inaccuracy of their model (It's only a model /Patsy -- but it can still host the whole song-and-dance).

While I agree there is a large contingent of people who seem to believe every detail of real neurons are necessary to cognition (up to and including the atoms), and while I tend to agree those people are way off the mark about which details are actually relevant to the algorithms of cognition versus which are specific to natures implementation thereof, I also think the existing AI field goes too far in the other direction and oversimplifies more than it can get away with.

So far we've been able to gloss over our mistake through the raw brute force of voluminous training data and GPU power. It works in the same way using a hammer to drive a screw into a wall works. Sure, we can do it to an extent, but there's a much better way. We need to figure out how to use a screwdriver. And by screwdriver I mean slightly more sophisticated artificial neuron.

The main idea is oscillatory neurodynamics. When you have a bunch of tuned oscillators, you can produce incredible computational capacity.

Reply on news.ycombinator.com