Brainspace

tjm 2023-10-03 12:41

It's... interesting? :sweat_smile: I was able to clear the first four problems just by playing around and a layperson's intuition of how neural networks work, but that was about my limit.

I do want to come back to it later and try to understand the underlying maths a little more, though; I think with maybe a little more in the way of supporting examples this could be a very useful teaching tool?

What's there works well and feels smooth to use (and I've written enough node-based editors to know how annoying it can be), so good work; Hopefully some other players can make it further than I did!

stmate03 2023-10-03 13:12

now I'm a logic oriented person, but I have not a clue what half of the nodes do, the other half I barely know what they do, but hey I passed level 1 with great results so yay!

endurion 2023-10-03 13:40

I do like programming games, but this is assuming too much knowledge from neural networks methinks. I did manage to pass level 1 with a little help though ;)

pegasys 2023-10-06 20:09

A very unique entry. I think this could be used as a base to turn into a more simplified tool to teach neural networks. If it were gamified a bit more (unlocking different nodes one by one, maybe a bit more of a graphical indication of what each node does etc.), this could easily become a really fun programming style game. Great job!

fisher2470 2023-10-06 20:18

This is addicting, but I definitely don't have the experience necessary to make it all the way to the end. Really smooth interface and a decent game to go about learning how neural networks actually work.

joey-gugler 2023-10-06 20:22

In technical terms, it may seem quite impressive, but I find myself unable to fully comprehend what I am doing. I managed to resolve the initial issue, but the methods and reasons behind it elude my understanding.

moebius 2023-10-06 20:37

Even though I had a machine learning lecture back in university, it is already a while back that I had to actually work with neural networks. Still after some tinkering around I think it's a really nice editor that you designed there. Also I like that I can play around with it and immediately see if it works or not. With bigger networks usually you have to use a super computer and that's pricey, so this is really nice for learning how to get a network to converge.

If it's really a game? I don't know. But did I have fun playing around with it? Definitely. So, well done! Especially considering you only had two days for that.

jason-shapiro 2023-10-07 00:49

I finished the Coursera Machine Learning Course by Andrew Yang - maybe that isn't enough background, but I still don't understand what is going on, here. I realize that different types of hidden layers do different things, but my problem is that I don't understand the feedback I'm getting when it's wrong. I get it when the shape is wrong. I like seeing the graph get low as I train letting me know I've trained enough. But I don't know how to fix what I'm doing wrong. I tried to read the info screen and I sadly ran out of patience. :( Still, I think this is really cool. If I could figure out how it works this might be a fantastic game. Thanks for making it!

godsboss 2023-10-07 07:42

brainspace_good_and_bad.png

I'm confused. Is this a bug? Seems like some shapes are both good and bad?

loveapplegames 2023-10-07 08:08

@godsboss each sample represents a series of shapes with certain properties. In this case, all the shapes in the series are red.

@jason-shapiro I was thinking of visualising the network in a postcompo version, both the shapes of the involved tensors as well as the values.

candlesan 2023-10-07 08:11

I realy liked the idea of a game that could "gameify" the teaching of neural networks. I think a non-jam version that taught the concepts but in an incremental way could be very popular.

vidarn 2023-10-08 17:37

This was really interesting and fun! I'm glad I got to play it. It felt very well made and professional. I feel like I might have needed a bit more guidance as I got stuck on level 4 and couldn't quite figure out what I was doing wrong. The easier levels were really fun to solve, however!

Here's my attempt at level 4. It didn't work at all, but I don't know how to tweak it. I must have missed something in how I should set these up Screenshot 2023-10-08 193252.png

loveapplegames 2023-10-08 20:13

@vidarn thanks for the feedback! Your network looks good, only it needs two layers at the end to process the data properly, with a ReLU in between. One layer only has limited capacity for learning different logical functions. If it still doesn't work, try increasing the number of features in the intermediate layers.

dan-str 2023-10-09 05:42

cool game, keep up the good work

vidarn 2023-10-09 06:32

@loveapplegames awesome, that makes a lot of sense! I solved it now :v: Thanks a lot for the help! I'm having a lot of fun with this :grin:

henk 2023-10-15 15:28

Engrossing and educational! It's kind of meta that the game is training my personal neural network to understand how neural networks work.

It'd be nice to have some more visual understanding of what's going on, though, since I'm often in a situation where I've come up with a wrong answer but have no clue _why_ it's wrong. Maybe some numbers or shaded cells or something so I can see what the weights are or what data looks like when it passes through the nodes.

Some other ideas: - I think there should be a "response function" dropdown on the "layer" node rather than just a "response function" node, since I'm almost always going to add one after each layer. - Maybe an option to manually set NN weights? I'm often desperate to know whether what I've built is just training poorly or if it's not capable of representing the solution at all, and I think a way to input my own weights would help with that.

Also: Since it's possible to create an exhaustive list of test cases for some of these challenges, it'd be interesting to see what would happen if you threw _literally everything_ into the training data and tried to make a neural network that's 100% correct.

crotonyl-alcohol 2023-10-20 10:22

Nice game! I'm currently learning cv and this game shows me clearly what a rookie I am.

wiredoverload 2023-10-20 20:16

Interesting peek into behind the scenes of neural net creation, but I think I would have had a bit more fun and learned more if each puzzle had a solution you could look at or hints, as at certain points I really had no idea what I was missing.