Sorbet: A neuromorphic hardware-compatible transformer-based spiking model
57 points
2 months ago
| 6 comments
| arxiv.org
| HN
magicalhippo
2 months ago
[-]
I found this[1] article to give a nice overview over spiking neural networks and their connections to the more "traditional" neural networks of modern fame.

In particular the connection between the typical weighted-sum plus activation function and a simplistic spiking model where one considers the output simply by the spiking rate was illuminating (section 3).

[1]: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9313413/ Spiking Neural Networks and Their Applications: A Review

reply
satvikpendem
2 months ago
[-]
I wonder how well this model can typecheck Ruby code.
reply
thih9
2 months ago
[-]
Context: Sorbet is also the name of a popular Ruby type checker[1], built by Stripe.

[1]: https://sorbet.org

reply
fnordpiglet
2 months ago
[-]
Realistically only at static time
reply
Alifatisk
2 months ago
[-]
Can’t wait for the follow up paper, RBS
reply
krasin
2 months ago
[-]
There's no code or weights released => no way to reproduce their results.
reply
ajb
2 months ago
[-]
They may not have linked the source in the paper, but it's not hard to find: https://github.com/Kaiwen-Tang/Sorbet

Haven't checked if there's enough there to build it.

reply
krasin
2 months ago
[-]
Thank you very much for posting this! The code is indeed there, that's great.
reply
sanxiyn
2 months ago
[-]
Results get reproduced without code or weights all the time. I note that in this case training data and evaluation benchmarks are public.
reply
hiddencost
2 months ago
[-]
The ML community has historically held themselves to a higher standard, IME.
reply
evanwolf
2 months ago
[-]
sometimes it seems folks are just making up words.
reply
a-dub
2 months ago
[-]
neuromorphic hardware is just hardware that has biologically inspired designs.

spiking neural networks are artificial neural networks that actually simulate the dynamics of spiking neurons. rather than sums, ramps and squashing, they simulate actual spike trains and the integration of energy that occurs in the dendrites.

neuromorphic hardware can range from specialized asics for doing these simulations efficiently to more experimental hybrid analog-digital systems that use analog elements to do more of the computation.

it's all very cool stuff, but i tend to think of snns as similar to the wings on the avion 3 where simplified unit functions look more like a modern jet wing.

but who knows, maybe the neuromorphic route will open the door to far more efficient computations. personally, i'm very excited about potential wins that could come from novel computational substrates!

reply
ImHereToVote
2 months ago
[-]
I wonder how far we will move the goalposts once we have a multimodal transformer type model running on neuromorphic hardware.
reply
welferkj
2 months ago
[-]
Lots of people involved in explaining away AI are labouring under the axiom that intelligence is mysterious. Therefore, if I can understand how a system works, it logically follows that it can't be intelligent.
reply
lucubratory
2 months ago
[-]
I predict that many of those people will continue to believe that up until human cognition is mechanistically understood, at which point there will be some other reason that humans are "real" thinkers and machines are not. The problem is that theoretical opposition to the existence of AIs is incompatible with materialism and thus just doesn't fit with our world, which is very much built using the scientific truths that materialism enables us to discover.
reply
welferkj
2 months ago
[-]
It is insane to me that views of consciousness and cognition other than physicalism still exist in mainstream scientific and philosophical discourse. As far as I can tell, no matter how much discourse you dress it up in, any alternative boils down to "it's magic, I ain't gotta explain shit".
reply
dartos
2 months ago
[-]
We… can’t really understand how neural networks work, but we can definitely tell they’re not intelligent beyond making good sounding word soup (as demonstrated in their minimal practical reasoning abilities)

I wouldn’t call pagerank intelligent, even though I can give it a text prompt and get relevant information back.

In my view, the only difference between that and an llm is the natural language interface.

I’m no expert on intelligence, but I’d expect being able to introspect and continually learn to be part of it.

reply
welferkj
2 months ago
[-]
You're engaging in explaining away intelligence.

One way to help you notice this is to try and estimate how many billions of people you've defined out of "being intelligent" with your latest goalpost movement.

Be honest, how many people do you think "introspect and continually learn" on a daily basis?

reply
cootsnuck
2 months ago
[-]
> Be honest, how many people do you think "introspect and continually learn" on a daily basis?

That's wild if you think that isn't quite literally one of the defining features of human consciousness (and many would say other animals as well).

If you think people thinking differently than you means they don't still indeed...think...then I don't know what tell you.

reply
welferkj
2 months ago
[-]
Unfortunately for the intelligence denial crowd, introspection and learning capability is something we can measure, as opposed to the vibes-based discourse you prefer to engage in. If that's what you've picked for your threshold of "intelligence", you've reduced the majority of the bell curve's left side to soulless automatons. Again, your definition, not mine.
reply
dartos
2 months ago
[-]
My definition (again, as a layman on the subject of intelligence philosophy), but your incorrect analysis.

I guess I just think more highly of my fellow humans.

reply
exe34
2 months ago
[-]
> I guess I just think more highly of my fellow humans.

As an article of faith, yes. But I don't see what this adds to the discussion.

reply
dartos
2 months ago
[-]
> If that's what you've picked for your threshold of "intelligence", you've reduced the majority of the bell curve's left side to soulless automatons

I disagree with this statement, which is the crux of their argument against my definition of intelligence.

I don’t think any credible survey of the intelligence or lack of a large enough population exists (due to there not being a common binary measure of intelligence), so it’s an issue you kind of need to take on faith.

reply
ImHereToVote
2 months ago
[-]
>so it’s an issue you kind of need to take on faith.

Thanks for playing.

reply
dartos
2 months ago
[-]
It cuts both ways… you need to believe that most humans aren’t actually intelligent as we don’t have any data that suggests that most humans aren’t.
reply
exe34
2 months ago
[-]
you have no data that suggests I can't fly either.
reply
dartos
2 months ago
[-]
… but… we do have data that proves that humans can’t just fly?

I think… I’m done talking with you now.

reply
exe34
2 months ago
[-]
you just like your assumptions to go unchallenged. goodbye!
reply
dartos
2 months ago
[-]
It’s crazy to me that people would rather believe that we can create intelligence by feeding the text of the internet into a statistics machine, than believe that the people making that text are intelligent.
reply
welferkj
2 months ago
[-]
It's crazy to me people would rather deny the intelligence of a large segment of the human population than admit to the increasing overlap between it and AI.

That's what you're doing when you keep moving the goalposts of "real intelligence" further and further right on the bell curve. You're denying the intelligence and consciousness of billions of people (and counting) just so you don't have to admit there's nothing magical about intelligence.

Sometime in the next 10 years, you'll have to start thinking of yourself as a soulless automaton to keep up the delusion. Good luck with that.

reply
dartos
2 months ago
[-]
> It's crazy to me people would rather deny the intelligence of a large segment of the human population than admit to the increasing overlap between it and AI.

You’re the one here denying. I think the vast majority (if not all) of humans are intelligent under my definition. You do not.

I don’t think LLMs or other statistical models are.

reply
welferkj
2 months ago
[-]
Under the latest definition you made up on the spot, yes. And definitely not all.

So what's your plan when the fraction keeps shrinking? When you're no longer in it?

This is simple interpolation. It is plainly obvious that at some point soon, you will be faced with the fact that there's nothing magical about intelligence. When that happens, will you concede that, or start thinking of yourself as a soulless automaton?

If you can't project that far forward, I question whether you meet any meaningful definition of "intelligent" right now.

reply
dartos
2 months ago
[-]
> So what's your plan when the fraction keeps shrinking?

What fraction? How would it shrink?

I don’t think that humans, as a species, are becoming non intelligent en masse. In fact, I think that we are, by default, intelligent.

That’s where our opinions seem to irreconcilably differ.

> you will be faced with the fact that there's nothing magical about intelligence.

I dont think there’s anything “magical” about anything. I just don’t think that a statistical model can achieve intelligence as we think of it with regards to humans.

You may see the recent trend of text generation models as new intelligent machines, but I’ve been studying and working these kinds of statistical models for about a decade (since 2016) and have seen these opinions spouted only to quiet down once the logarithmic improvement curve is reached. I don’t see any reason why these LLMs wouldn’t follow the same pattern.

> This is simple interpolation

Interpolation of what? You’re assuming that the goalpost will always shift, but in reality we just don’t have a generally agreed upon definition all. Either way, any definition of intelligence that rules out the majority of humanity is an incorrect definition off the bat, as pretty much all humans are intelligence.

There exists some accurate definition of intelligent such that almost any human satisfies it, but statistical models do not. I’m sure if I studied the philosophy of intelligent I could put one into words, but I’m ill equipped to do so.

> If you can't project that far forward, I question whether you meet any meaningful definition of "intelligent" right now.

Are you just trying to be mean, or do you actually believe that people who disagree with you are not intelligent?

We’ll see in 5 years that this intelligence hype will fade just like that last 2 AI booms.

This isn’t at all to say that we will never make a machine with intelligence that rivals humans, just that I don’t think the statistical model route will get us there… and it hasn’t.

reply
dartos
2 months ago
[-]
One last pretty funny thing:

Here a short chatgpt convo about how personification bias can cause people to believe that statistical models are intelligent. I think it's what's fooling so many people.

https://chatgpt.com/share/670549a3-2f9c-8001-81c1-d950c626ad...

reply
dartos
2 months ago
[-]
> how many people do you think "introspect and continually learn" on a daily basis?

At the very least, every single person who plays sports, video games, tries finding a way around traffic, a faster route home, a way to do less work, take a longer break, or a way to save some extra money getting food.

Literally any optimization task at all requires an observation, analysis (read: introspection,) and adjustment. That’s why we model training loops as optimization problems.

We spoof that with REACT prompts in LLMs, but it becomes clear after a few iterations that there’s no real optimization going on, just guessing at tokens (a gross oversimplification, as this guessing has real uses). It’s doing what it was trained to do, completes text. Not to mention that those steps all disappear when the prompt is changed.

reply
exe34
2 months ago
[-]
> One way to help you notice this is to try and estimate how many billions of people you've defined out of "being intelligent" with your latest goalpost movement.

love this, I will use this in future rants.

reply
dartos
2 months ago
[-]
That argument only works if your audience already thinks of humans as mostly automotons…
reply
exe34
2 months ago
[-]
It's an operational definition: if you claim AI is not intelligent because it cannot do X, then you necessarily exclude a whole lot of humans who also can't do X.

There used to be a strident faction that would say "but AI can't produce original art/a symphony/novel/etc". My answer was usually (correctly), "neither can you."

reply
dartos
2 months ago
[-]
> It's an operational definition: if you claim AI is not intelligent because it cannot do X, then you necessarily exclude a whole lot of humans who also can't do X.

Sure, but I think most people are intelligent according to my definition, but AI is not…

You’re already coming from the assumption that people are “souless automatons,” which is probably why the idea of a machine being “intelligent” is so easy for you to accept.

> There used to be a strident faction that would say "but AI can't produce original art/a symphony/novel/etc". My answer was usually (correctly), "neither can you."

This is a dumb apples and oranges comparison. AI as a concept is different than a concrete person.

AI as a concept can do anything, it’s a conceptual placeholder for an everything machine.

reply
dartos
2 months ago
[-]
I can’t reply to the other comment, but “soulless” was to quote the other commenter. Having a soul (whatever that might mean) holds no bearing on what I’m saying.
reply
exe34
2 months ago
[-]
again, do they obey the laws of physics? can they decide to go against what the physical interactions in their brain guide them to do?
reply
exe34
2 months ago
[-]
> You’re already coming from the assumption that people are “souless automatons

do your people obey the laws of physics? is the soul magical or physical?

reply
joloooo
2 months ago
[-]
I think there is a difference from people upset around over hyped LLMs and arguing about intelligence in "A.I.". Most of the "intelligence" arguments I've seen are fighting against putting too much stock in chatgpt and Sam's fever dreams.
reply
a-dub
2 months ago
[-]
the goalposts for what?
reply
ImHereToVote
2 months ago
[-]
Literal goalpost in a game of football of course. Or soccer if you are an American.
reply
a-dub
2 months ago
[-]
thanks. you seem to think that a spiking multimodal variant of transformers on neuromorphic hardware would demarcate a goal of some sort, which one?

for as far as i can see, the achievement would just be a spiking multimodel variant of transformers on neuromorphic hardware.

reply
ImHereToVote
2 months ago
[-]
I bet you are great at playing blackjack, but suck at Texas hold 'em.
reply
ithkuil
2 months ago
[-]
To be fair, all words are made up.

Words are useful to the extent they effectively communicate with the intended audience.

This can be accomplished by a mix of familiarity (has this word been already used enough in the target audience with the intended meaning) and the ability to evoke new meanings by intuitive derivation rules (word composition, affixes, ...)

In the case of this title, fwiw, it was perfectly clear to me what this was about because I'm already familiar with related topics and they were using the same terminology

reply
Arubis
2 months ago
[-]
And even with a willingness to make up words, it’s STILL hard to name tech projects uniquely: https://github.com/sorbet/sorbet
reply
fnordpiglet
2 months ago
[-]
I’m flashing back to bapi
reply
allendave6945
2 months ago
[-]
Hello, as a newbie to cryptocurrency trading, I lost a lot of money trying to navigate the market on my own, then in my search for a genuine and trusted trader/broker, i came across Trader Bernie Doran who guided and helped me retrieve my lost cryptocurrencies and I made so much profit up to the tune of $60,000. I made my first investment with $2,000 and got a ROI profit of $25,000 in less than 2 week. You can contact this expert trader Mr Bernie Doran via Gmail : BERNIEDORANSIGNALS@GMAIL.COM or WhatsApp +14242850682 and be ready to share your experience , tell him I referred you
reply
remon
2 months ago
[-]
I definitely know what "A" and "model" means.
reply