🏅 Top traders

I need help resolving this. Implementation obviously exists, but there are no CIFAR results yet. The question may be interpreted as "the implementation as of today should be good enough to get the same results on CIFAR, but the test doesn't have to be done by today" or "... and the test is included in the results that must exist by today".

What is usually done in such cases?

@l8doku I bet based on what was written, given I assumed that there'd be no CIFAR implementation by Jan 7th. It seemed to me that the market was irrationally enthusiastic for some vague "community" to fill in and dot all those i's. As written I think it is a NO.

@l8doku I see my comment earlier assumed a CIFAR test would be done in a month

@l8doku Also the missing RNN pieces means the public implementations themselves are not complete, not just not tested.

@EdwardKmett a comment by @SiddharthHiregowdara links to a version with the RNN part. However, the repository mentions other parts which aren't implemented.

@EdwardKmett At the end of the day, I bet based on the intersection of assuming apathy on the part of my fellow man would ensure that this didn't get to done, and off your own previous comment that lacking CIFAR results was enough for you to not stop and call this a YES. Do as you will.

@EdwardKmett Of course, you could morally resolve this YES, by just rushing off and trying to get CIFAR-10 comparable results in the next 4 hours, whereupon I'd feel justified in having posted effectively a bounty for that effort. ;)

@EdwardKmett The only thing that has kept me from betting harder on this market, frankly, is resolution risk, given you have a position on the other side.

@EdwardKmett

I bet the same way. I understood the criteria as “if there are not both MNIST and CIFAR results by Jan 7, then the resolution is NO.”

@QuantumObserver Same, I pulled out based on the lack of CIFAR.

@EdwardKmett My position is one of the reasons I started this discussion. I don't trust myself to be unbiased in resolving the market.

However, a large part of my motivation is other people who bought "YES" and haven't pulled out. If they bought assuming the wording where CIFAR-10 results can appear later (using the code from Jan 6th), resolving "NO" would feel unfair towards them in the same way as resolving "YES" is unfair towards you and other people who commented in this thread.

@l8doku Except they always had the excuse that at any point along the way this could be something that could happen. That only works in the YES direction, after all you can't unring that bell. It just never rang.

@EdwardKmett fair, and my old comment about CIFAR-10 is a perfect argument for this point.

I'll wait for a day to see if there are more comments and then resolve it.

I'll make my case for "YES", too, if nobody does it by that time (not doing it right now because I don't trust myself). In this case I think I'll resolve "NO" after making my arguments.

If someone else makes a case for "YES", I'll think again.

In any case, thank you for showing me the importance of clear resolution criteria in practice.

@l8doku No one replied, I'm resolving it "NO".
I started arguing my case for "YES" but ended up convincing myself that "NO" is correct (basically, running it on CIFAR isn't as easy as just substituting the dataset, you need to modify enough code that it should count as too much modification). If it was just dataset substitution, I think it would be "YES".

https://github.com/nebuly-ai/nebullvm/tree/main/apps/accelerate/forward_forward

@SiddharthHiregowdara Still no CIFAR. =)

https://twitter.com/mpezeshki91/status/1600610571485089792

https://github.com/mohammadpz/pytorch_forward_forward

Seems like an implementation was released. This is a simple version without the RNN part, and there is no CIFAR-10 test, so I won't resolve "Yes" for now.

It's probably safe to assume that a CIFAR-10 test can be done easily, and will be done within a month. I'm not as sure about the RNN part and whether an implementation without it counts.

This is my first time resolving a market, so please tell me if I'm doing something wrong or not doing something obvious.

@l8doku what is the RNN part? I haven't checked out the paper, only read the twitter thread

@JackOBrien In the simple version each layer only "sees" the layer below it. So you can't train the first layer in a way that provides optimal inputs to the second layer. The second layer has to work with whatever weights the first one ended up having. This is a huge problem as you add more layers.

A potential solution is to include the second layer's weights as inputs to the first layer. So the first layer not only tries to remember data, but also considers the second layer while doing so.

Usually this would lead to a loop (the first layer learns from the second one and the second one learns from the first one) and would require a complex solution (if it's even possible to solve). Instead of updating everything at the same time, these connections use past weights, which are fixed and don't lead to weird equations.

In the initial configuration there is no "past", there is only one pass of the data through the network, so you can't apply this method. So the data gets passed through the network several times, using the weights from the previous pass to learn better weights in the current one. The paper says they used 8-10 iterations on MNIST.

There is a literal RNN part later, and I think it's a similar architecture, but for real sequences of data, not a pseudo-sequence for more efficient learning.

Will a Forward-Forward neural network implementation be publicly available before Jan 7th 2022?, 8k, beautiful, illustration, trending on art station, picture of the day, epic composition