Found another terrific one while looking at cards with X in their text (not necessarily cost):
Flavour-wise he is terrific; kill off his guildmates, and the more he kills the more powerful he gets. Even the cardname is perfect. The only pity here is the colour (and the rarity and subtype, but those don't matter as much). This is pretty much the epitome of a black card, though white has some sacrifice-and-benefit effects (though not many, like Caregiver, and none of them do buffs). In fact, this is more an Orzhov card, I feel; somewhat like Maw of the Obzedat. Or alternately, a more sacrificial and personal version of Kjeldoran War Cry.
Still, this is one of my favourite cards to emerge from the network.
Also; I'm analysing 16,000 cards or so, and about 50 have been generated with the text 'choose a colour', and not one of them has mentioned that colour ever again. Clearly, though this network is powerful, it has much to learn.
Salt Rock Horror 3BB
Creature - Horror (uncommon)
Whenever Salt Rock Horror deals combat damage to a player, put a +1/+1 counter on it.
Remove a +1/+1 counter from Salt Rock Horror: put a +1/+1 counter on Salt Rock Horror.
(3/3)
... isn't going to understand that Salt Rock Horror's ability isn't tactically useful.
Ok, so after some prodding, I discovered that MSE2 does in fact run perfectly fine on Ubuntu 14.04 with WINE 1.7.
Yaay! Quality renders are a go!
I've noticed the NN lacks a proper set symbol, which is something I'm actually up to the task of fixing. (The main project is beyond me, as I'm not fluent in any programming languages.) Attached is a mockup of the symbol at different rarities. The MSE symbol file can be located here. If this doesn't suit your fancy, I'm happy to take another crack at it.
Yaay! Quality renders are a go!
I've noticed the NN lacks a proper set symbol, which is something I'm actually up to the task of fixing. (The main project is beyond me, as I'm not fluent in any programming languages.) Attached is a mockup of the symbol at different rarities. The MSE symbol file can be located here. If this doesn't suit your fancy, I'm happy to take another crack at it.
If I may make a suggestion? Add a fourth circle at the top and separate the middle ones so you have a Magic-style pentagon in the middle.
Telcos, you said you were using mana costs to develop your color classifier. Have you considered using color identity instead?
It seems to me that your classifiers might see a creature with "GG: @ gains trample until end of turn" and classify it as a green card, because it mentions trample. In reality, this card could be mono-blue, or any other color(s).
Actually, it might be best to strip out off-color abilities altogether, and run them through their own classifier. Otherwise your classifier might still accept cards like:
Green Flier
3G
Creature ~ Beast
Flying
UU: @ gains trample until end of turn.
The generator may be smart enough to avoid this, but the order-agnostic classifier might let it through.
I liked your post last night to indicate that I had read it. And yes, I agree with your assessment. Good points.
Yaay! Quality renders are a go!
I've noticed the NN lacks a proper set symbol, which is something I'm actually up to the task of fixing. (The main project is beyond me, as I'm not fluent in any programming languages.) Attached is a mockup of the symbol at different rarities. The MSE symbol file can be located here. If this doesn't suit your fancy, I'm happy to take another crack at it.
That is so amazing. Wow! Input layer, hidden layer, output layer. Genius!! Text cannot describe how big of a grin appeared upon my face when I saw this. I mean, if you want to change it, that's fine too, the output will be equally interesting I'm sure. How does it look at the size it'll actually be rendered at? That'd be one concern. It might be a bit cluttered. You may consider moving things about like Zaratustra suggested.
Salt Rock Horror 3BB
Creature - Horror (uncommon)
Whenever Salt Rock Horror deals combat damage to a player, put a +1/+1 counter on it.
Remove a +1/+1 counter from Salt Rock Horror: put a +1/+1 counter on Salt Rock Horror.
(3/3)
... isn't going to understand that Salt Rock Horror's ability isn't tactically useful.
Flavour-wise he is terrific; kill off his guildmates, and the more he kills the more powerful he gets. Even the cardname is perfect. The only pity here is the colour (and the rarity and subtype, but those don't matter as much). This is pretty much the epitome of a black card, though white has some sacrifice-and-benefit effects (though not many, like Caregiver, and none of them do buffs). In fact, this is more an Orzhov card, I feel; somewhat like Maw of the Obzedat. Or alternately, a more sacrificial and personal version of Kjeldoran War Cry.
Still, this is one of my favourite cards to emerge from the network.
Also; I'm analysing 16,000 cards or so, and about 50 have been generated with the text 'choose a colour', and not one of them has mentioned that colour ever again. Clearly, though this network is powerful, it has much to learn.
First, I love Murderous Guildmage. Second, the network is very unclear about the meaning of "choose a color". The problem is that cards with that text have no regularity to them whatsoever, so it's not apparent to the network just where that clause joins up with the next.
That being said, it is possible to get cards that make use of the chosen color... just in very messy, ill-defined ways. Here are some examples that I forced the network to make by compelling it to include "ETB, choose a color"/"choose a color"
|elvish fire seer||creature||elf druid|N|&^/&^|{GG}|when @ enters the battlefield, choose a color.\@ is the chosen color is a creature card other than @ in addition to its other types.|
|grasp of souls||instant|||O||{BB}|choose a color. target creature becomes the color of your choice until end of turn.|
|beast charge||instant|||O||{WW}|choose a color. creatures you control gain protection from the chosen color until end of turn.|
|triton burst||instant|||O||{UUUU^}|choose a color. creatures you control gain protection from the color of your choice until end of turn.|
|eye of the fireheart||instant|||O||{^^WW^}|choose a color with the most common color among permanents you control and target creature with power &^^ or less. it can't be regenerated.|
The problem is exacerbated by the fact that "choose a color" and other such choices aren't always phrased the same way. Sometimes it's the "color of your choice" instead.
Also, sometimes "choose a color" is separated by a newline character (Alloy Golem) and sometimes it is not (Addle).
Altogether it's a very messy situation for the network to learn from. If this were a natural language corpus, we'd likely have more data to go on to figure out how to use the construct, and if this was a more regular formal language corpus, we wouldn't have to worry about such varied templating. Magic is somewhere in between. The templating needs to be consistent enough so players have no problem understanding the cards, yet not so strict that the wording comes off as extremely stilted and unnatural. Given the small size of the corpus, we can have situations like "choose a color" where the phrasing is difficult for the network to master.
Ok, so after some prodding, I discovered that MSE2 does in fact run perfectly fine on Ubuntu 14.04 with WINE 1.7.
Yaay! Quality renders are a go!
I've noticed the NN lacks a proper set symbol, which is something I'm actually up to the task of fixing. (The main project is beyond me, as I'm not fluent in any programming languages.) Attached is a mockup of the symbol at different rarities. The MSE symbol file can be located here. If this doesn't suit your fancy, I'm happy to take another crack at it.
Ok, so after some prodding, I discovered that MSE2 does in fact run perfectly fine on Ubuntu 14.04 with WINE 1.7.
Yaay! Quality renders are a go!
I've noticed the NN lacks a proper set symbol, which is something I'm actually up to the task of fixing. (The main project is beyond me, as I'm not fluent in any programming languages.) Attached is a mockup of the symbol at different rarities. The MSE symbol file can be located here. If this doesn't suit your fancy, I'm happy to take another crack at it.
Is that an adinkra?
AFAIK, it's a representation of the network. The top and bottom represent input/output layers of neurons, and the middle represents a hidden layer of neurons. The lines between them represent the connections (analogous to synapses in the brain). But an adinkra is another good interpretation, lol, although it's not bipartite.
EDIT: At first I thought you meant an actual adinkra symbol of the Akan people, and I was going to say "Yes, this symbol represents madness or, alternatively, divine inspiration." lol
That is so amazing. Wow! Input layer, hidden layer, output layer. Genius!! Text cannot describe how big of a grin appeared upon my face when I saw this. I mean, if you want to change it, that's fine too, the output will be equally interesting I'm sure. How does it look at the size it'll actually be rendered at? That'd be one concern. It might be a bit cluttered. You may consider moving things about like Zaratustra suggested.
I was worried about clutter, too. Taking your concern, and Zaratustra's idea of upping the pentagonal factor, I've created this:
MSE's standard export looks like this:
The uncommon rarity is the worst case for visibility, and its still fairly clean. The rest are as so:
The MSE Symbol file can be found here. Forgive me if any of the links or embeds fail. Dropbox is not my forte.
First off, thank you all for the work you've put into this, it's bee more than thoroughly entertaining. (I spent most of the weekend reading this thread and mocking up cards to surprise my group with a draft.)
However, I had to make an account just to to share a concerning card that Croxis's page spit out while I was mucking about with it.
1R
Punciment
Sorcery
Destroy target non-land permanent. If that player doesn't have flesh, that player discards a card.
I'm a bit troubled at the idea of the AI taking an interest in our physical status. Or, perhaps more interestingly, does it expect to be playing the game itself, and this is an odd way to check for the 'human-status' of the player? Which makes this an instance of meta-game 'robot hate' maybe?
Hm. Which is better: the highest croxis epoch, or the highest better_croxis epoch, and what is the difference? And is there a way with the online thingy to do the things where we tell it to generate threshold or such?
...You re-wrote the entirety of char-rnn in a different language? Which one? And how??
Haha, well, I didn't rewrite it so much as modified the neural network tools I already wrote for work to be able to use recurrent layers. It's written on top of Theano, which is a python library meant for neural networks and such.
Since I haven't needed recurrence for work, my toolset was structured in such a way that it only could train standard NNs. I just modified the model construction and training codes so that it works properly with RNNs. So now I can build arbitrary models with recurrent layers and do some experiments that I didn't have the know-how to do with lua
Okay, trying to post this for the fourth time after the first three were eaten, but at least I think I know why.
Anyway, I wanna know who "Someone" is. Let's find out!
I split up all the unique, named people's names into their syllables (that is, "Pia" and "Kiran" were counted separately, but two creatures representing the same person weren't). I then put them into groups based on where they appeared in the name, and whether or not they were a terminal syllable.
I think the reason it's disappearing is something in the list, so I'm not going to copy it out here. Sorry!
I used this to generate some random two-syllable names:
[...]
When you say "generate," what do you mean? Are you picking two syllables at random from the "first syllable" and "last syllable" categories?
And personally, I like "Tsaikiwon." It would fit fairly well in a Chinese-themed set!
@Fistborn, I wonder if there's an easily accessible dump of that... if there is, I might train a flavour network on that tonight.
@talcos, I found a paper today examining the performance of LTSM networks and long story short, it "recommends adding a bias of 1 to the forget gate of every LSTM in every application" to increase the performance. Not sure if that's already been done in our network, but since I have no idea of how the code is laid out, I can't really check for myself.
First off, thank you all for the work you've put into this, it's bee more than thoroughly entertaining. (I spent most of the weekend reading this thread and mocking up cards to surprise my group with a draft.)
However, I had to make an account just to to share a concerning card that Croxis's page spit out while I was mucking about with it.
1R
Punciment
Sorcery
Destroy target non-land permanent. If that player doesn't have flesh, that player discards a card.
I'm a bit troubled at the idea of the AI taking an interest in our physical status. Or, perhaps more interestingly, does it expect to be playing the game itself, and this is an odd way to check for the 'human-status' of the player? Which makes this an instance of meta-game 'robot hate' maybe?
We need a ruling on whether flaying your opponent is instant speed.
@Fistborn, I wonder if there's an easily accessible dump of that... if there is, I might train a flavour network on that tonight.
@talcos, I found a paper today examining the performance of LTSM networks and long story short, it "recommends adding a bias of 1 to the forget gate of every LSTM in every application" to increase the performance. Not sure if that's already been done in our network, but since I have no idea of how the code is laid out, I can't really check for myself.
That is an extremely useful paper for me, thank you for bringing it to my attention. I'll have to read it later.
So what you're saying is that we should add a bias so that it obsessively hoards interesting information?
If my understanding is correct, if the forget gate outputs a value of zero and the input gate outputs a one, it causes the cell to dump the previous values it was holding and to adopt new ones. If the input gate outputs a zero and the forget gate outputs a one, then the contents of memory stay the same. Any other values should cause the contents of memory to become a weighted combination of the old and new.
So by adding the bias, we'd be encouraging it to color all new experiences with the knowledge of the old. On the other hand, during training, we're also encouraging it to dump the contents of its memory at the end of a card since they won't be useful in predicting the next card. I'm wondering how that tension would play out.
But no, I don't think there's a bias in the code currently.. but, I mean, we could add one. The forget gate is just a sigmoid module, and we could append an add module onto it. Yeah, that's definitely do-able. I could try that tonight and see if it makes any significant difference for us.
@Talcos, I'm looking forward to you trying that, I'm fascinated to see if it'll have any noticeable improvement on our cards. I'd try it myself... but this RNN code is still a big black mystery box to me
Tonight I think I'll try to find some good fantasy-book quotes and train an alternate flavour network with that. I don't suppose anyone has a convenient text file full of MtG flavour text that I can join with fantasy-book stuff, by chance?
Hm. Which is better: the highest croxis epoch, or the highest better_croxis epoch, and what is the difference? And is there a way with the online thingy to do the things where we tell it to generate threshold or such?
Yesterday, the former was better. There's one new better_croxis epoch produced since then, but I doubt it's significantly better. Perhaps in a few days. The diference is the training parameters. The better_croxis parameters should lead to a better result eventually, but hasn't yet.
You could probably put threshold in 'append body text'. I havent tested that yet.
Don't use capital letters in any fields. It won't work but it won't explain why. (It's because it's trained mostly on lowercase so some capitals aren't present in the vocabulary. I think.)
@Talcos, I'm looking forward to you trying that, I'm fascinated to see if it'll have any noticeable improvement on our cards. I'd try it myself... but this RNN code is still a big black mystery box to me
Tonight I think I'll try to find some good fantasy-book quotes and train an alternate flavour network with that. I don't suppose anyone has a convenient text file full of MtG flavour text that I can join with fantasy-book stuff, by chance?
HOLY CRAP YOU GUYS. I found out why the network keeps making you name cards to no effect! It's all to enable this guy!
So, surely, if you were to prime it with something like "Truenamer ? Whenever you name a card," couldn't you generate a bunch of these guys?
Indeed! And yes, we have done so in the past.
EDIT(2): If I Lua'd correctly, I think I succeeded in adding in a bias of 1 to every forget gate of every cell, as maplesmall suggested after reading that very recent work by Jozefowicz, Zaremba, and Sutskever (it came out mid-July, so it was extremely recent). I started training again with the parameters that I used last time. We'll see how it does. It could go well or very poorly. I honestly have no idea. But whatever happens, it'll be informative! Part of me fears that the change will encourage obsessions and fixations, but who knows? We may get lucky.
So by adding the bias, we'd be encouraging it to color all new experiences with the knowledge of the old.
That sounds an awful lot like it would be even 'better' at remembering the slightly funky colour pie of pre-Modern cards, hence making any generated cards even more colour-confused. Possibly this experiment should be done with the full corpus (because why not) as well as a corpus with only Modern cards?
So by adding the bias, we'd be encouraging it to color all new experiences with the knowledge of the old.
That sounds an awful lot like it would be even 'better' at remembering the slightly funky colour pie of pre-Modern cards, hence making any generated cards even more colour-confused. Possibly this experiment should be done with the full corpus (because why not) as well as a corpus with only Modern cards?
Well, I already started the training before shutting down my terminal for the day, so we'll burn that bridge when we get to it.
For big, long-term stuff, trends, etc. I'm less concerned. I'm more concerned about what that'll do for short-term memory, because forgetting is crucial for topic shifts (i.e. one card to the next). Now it may compensate and do just fine, of course. But for you and me, adding in the bias would be like setting orange juice in your fridge at home and getting to work and having a gut feeling that the fridge in the office contains orange juice.
EDIT: Or, for another analogy, I recall my work at one of our nation's government research labs, and I read about work on a neural repeater for patients with dementia symptoms. It merely echoed back signals it picked up, which meant that ideas couldn't slip away as easily. On the other hand, I reasoned that if the repeater were more powerful and operated on larger timescales, then the patient's past would start to collide with their present, sort of like a real life Paradox Haze. Fortunately, I only experiment with machine simulations of neural networks, so I don't have to worry about horrendous ethical consequences.
EDIT(2): Oh, and in case any of you get the opportunity, I loved interning at a government research lab. However, I will say this: all the security just puts me on edge, not very conducive to my work. Too many men with guns. And I wanted to take a picture of some ducks by the lake in the compound, but I couldn't. They were government ducks. Photography was forbidden.
Flavour-wise he is terrific; kill off his guildmates, and the more he kills the more powerful he gets. Even the cardname is perfect. The only pity here is the colour (and the rarity and subtype, but those don't matter as much). This is pretty much the epitome of a black card, though white has some sacrifice-and-benefit effects (though not many, like Caregiver, and none of them do buffs). In fact, this is more an Orzhov card, I feel; somewhat like Maw of the Obzedat. Or alternately, a more sacrificial and personal version of Kjeldoran War Cry.
Still, this is one of my favourite cards to emerge from the network.
Also; I'm analysing 16,000 cards or so, and about 50 have been generated with the text 'choose a colour', and not one of them has mentioned that colour ever again. Clearly, though this network is powerful, it has much to learn.
I dunno, it's a two card combo with Doubling Season, Hardened Scales or Enduring Scalelord...
Draft Link
Just a casual Legacy/Vintage/No-Bans player who mostly plays EDH.
Sorry for the mistake.
I was pretty interested in your old work, but your new breakthrough teaser has me truly excited! *Waits patiently for a few weeks.*
Yaay! Quality renders are a go!
I've noticed the NN lacks a proper set symbol, which is something I'm actually up to the task of fixing. (The main project is beyond me, as I'm not fluent in any programming languages.) Attached is a mockup of the symbol at different rarities. The MSE symbol file can be located here. If this doesn't suit your fancy, I'm happy to take another crack at it.
If I may make a suggestion? Add a fourth circle at the top and separate the middle ones so you have a Magic-style pentagon in the middle.
I liked your post last night to indicate that I had read it. And yes, I agree with your assessment. Good points.
Not a bad idea, actually. Lots of descriptive text in there that we could split up into brief blurbs for the sake of training.
That is so amazing. Wow! Input layer, hidden layer, output layer. Genius!! Text cannot describe how big of a grin appeared upon my face when I saw this. I mean, if you want to change it, that's fine too, the output will be equally interesting I'm sure. How does it look at the size it'll actually be rendered at? That'd be one concern. It might be a bit cluttered. You may consider moving things about like Zaratustra suggested.
Well, I meant for limited purposes, haha. But no, I know, there are potential uses if you have a card pool involving such cards.
First, I love Murderous Guildmage. Second, the network is very unclear about the meaning of "choose a color". The problem is that cards with that text have no regularity to them whatsoever, so it's not apparent to the network just where that clause joins up with the next.
That being said, it is possible to get cards that make use of the chosen color... just in very messy, ill-defined ways. Here are some examples that I forced the network to make by compelling it to include "ETB, choose a color"/"choose a color"
The problem is exacerbated by the fact that "choose a color" and other such choices aren't always phrased the same way. Sometimes it's the "color of your choice" instead.
Also, sometimes "choose a color" is separated by a newline character (Alloy Golem) and sometimes it is not (Addle).
Altogether it's a very messy situation for the network to learn from. If this were a natural language corpus, we'd likely have more data to go on to figure out how to use the construct, and if this was a more regular formal language corpus, we wouldn't have to worry about such varied templating. Magic is somewhere in between. The templating needs to be consistent enough so players have no problem understanding the cards, yet not so strict that the wording comes off as extremely stilted and unnatural. Given the small size of the corpus, we can have situations like "choose a color" where the phrasing is difficult for the network to master.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
Is that an adinkra?
AFAIK, it's a representation of the network. The top and bottom represent input/output layers of neurons, and the middle represents a hidden layer of neurons. The lines between them represent the connections (analogous to synapses in the brain). But an adinkra is another good interpretation, lol, although it's not bipartite.
EDIT: At first I thought you meant an actual adinkra symbol of the Akan people, and I was going to say "Yes, this symbol represents madness or, alternatively, divine inspiration." lol
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
I was worried about clutter, too. Taking your concern, and Zaratustra's idea of upping the pentagonal factor, I've created this:
MSE's standard export looks like this:
The uncommon rarity is the worst case for visibility, and its still fairly clean. The rest are as so:
The MSE Symbol file can be found here. Forgive me if any of the links or embeds fail. Dropbox is not my forte.
Edit: Moved to Imgur.
However, I had to make an account just to to share a concerning card that Croxis's page spit out while I was mucking about with it.
1R
Punciment
Sorcery
Destroy target non-land permanent. If that player doesn't have flesh, that player discards a card.
I'm a bit troubled at the idea of the AI taking an interest in our physical status. Or, perhaps more interestingly, does it expect to be playing the game itself, and this is an odd way to check for the 'human-status' of the player? Which makes this an instance of meta-game 'robot hate' maybe?
@Jormengand, that's awesome. Some of those are terrific fantasy names (Kalidor in particular).
When you say "generate," what do you mean? Are you picking two syllables at random from the "first syllable" and "last syllable" categories?
And personally, I like "Tsaikiwon." It would fit fairly well in a Chinese-themed set!
EDIT:
Arnold agrees
I think a dump from wikiquote (perhaps focusing on fantasy novels) is probably better.
@talcos, I found a paper today examining the performance of LTSM networks and long story short, it "recommends adding a bias of 1 to the forget gate of every LSTM in every application" to increase the performance. Not sure if that's already been done in our network, but since I have no idea of how the code is laid out, I can't really check for myself.
We need a ruling on whether flaying your opponent is instant speed.
That is an extremely useful paper for me, thank you for bringing it to my attention. I'll have to read it later.
So what you're saying is that we should add a bias so that it obsessively hoards interesting information?
If my understanding is correct, if the forget gate outputs a value of zero and the input gate outputs a one, it causes the cell to dump the previous values it was holding and to adopt new ones. If the input gate outputs a zero and the forget gate outputs a one, then the contents of memory stay the same. Any other values should cause the contents of memory to become a weighted combination of the old and new.
So by adding the bias, we'd be encouraging it to color all new experiences with the knowledge of the old. On the other hand, during training, we're also encouraging it to dump the contents of its memory at the end of a card since they won't be useful in predicting the next card. I'm wondering how that tension would play out.
But no, I don't think there's a bias in the code currently.. but, I mean, we could add one. The forget gate is just a sigmoid module, and we could append an add module onto it. Yeah, that's definitely do-able. I could try that tonight and see if it makes any significant difference for us.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
Tonight I think I'll try to find some good fantasy-book quotes and train an alternate flavour network with that. I don't suppose anyone has a convenient text file full of MtG flavour text that I can join with fantasy-book stuff, by chance?
Yesterday, the former was better. There's one new better_croxis epoch produced since then, but I doubt it's significantly better. Perhaps in a few days. The diference is the training parameters. The better_croxis parameters should lead to a better result eventually, but hasn't yet.
You could probably put threshold in 'append body text'. I havent tested that yet.
Don't use capital letters in any fields. It won't work but it won't explain why. (It's because it's trained mostly on lowercase so some capitals aren't present in the vocabulary. I think.)
Sure, I can give it a try
And yes, actually I do. Here:
https://drive.google.com/file/d/0BxF7G2b8kigCek10T1RUZUQzYkU/view?usp=sharing
You might want to take out the "|" symbols so it meshes better with the fantasy-book stuff, unless you're chunking up the books in the same way.
EDIT:
Indeed! And yes, we have done so in the past.
EDIT(2): If I Lua'd correctly, I think I succeeded in adding in a bias of 1 to every forget gate of every cell, as maplesmall suggested after reading that very recent work by Jozefowicz, Zaremba, and Sutskever (it came out mid-July, so it was extremely recent). I started training again with the parameters that I used last time. We'll see how it does. It could go well or very poorly. I honestly have no idea. But whatever happens, it'll be informative! Part of me fears that the change will encourage obsessions and fixations, but who knows? We may get lucky.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.
Regarding your explanation of the bias, you said,
That sounds an awful lot like it would be even 'better' at remembering the slightly funky colour pie of pre-Modern cards, hence making any generated cards even more colour-confused. Possibly this experiment should be done with the full corpus (because why not) as well as a corpus with only Modern cards?
Well, I already started the training before shutting down my terminal for the day, so we'll burn that bridge when we get to it.
For big, long-term stuff, trends, etc. I'm less concerned. I'm more concerned about what that'll do for short-term memory, because forgetting is crucial for topic shifts (i.e. one card to the next). Now it may compensate and do just fine, of course. But for you and me, adding in the bias would be like setting orange juice in your fridge at home and getting to work and having a gut feeling that the fridge in the office contains orange juice.
EDIT: Or, for another analogy, I recall my work at one of our nation's government research labs, and I read about work on a neural repeater for patients with dementia symptoms. It merely echoed back signals it picked up, which meant that ideas couldn't slip away as easily. On the other hand, I reasoned that if the repeater were more powerful and operated on larger timescales, then the patient's past would start to collide with their present, sort of like a real life Paradox Haze. Fortunately, I only experiment with machine simulations of neural networks, so I don't have to worry about horrendous ethical consequences.
EDIT(2): Oh, and in case any of you get the opportunity, I loved interning at a government research lab. However, I will say this: all the security just puts me on edge, not very conducive to my work. Too many men with guns. And I wanted to take a picture of some ducks by the lake in the compound, but I couldn't. They were government ducks. Photography was forbidden.
My LinkedIn profile... thing (I have one of those now!).
My research team's webpage.
The mtg-rnn repo and the mtg-encode repo.