Mdthbs

Can I make a matrix from just a parts of the cells?

Does academia have a lazy work culture?

Why doesn't current cancel in a generator?

Why are so many countries still in the Commonwealth?

How do we explain the E major chord in this progression?

Is there a reason why I should not use the HaveIBeenPwned API to warn users about exposed passwords?

How to deal with a player who makes bad characters and kills them?

How can I prevent corporations from growing their own workforce?

Why is drive/partition number still used?

Keeping an "hot eyeball planet" wet

How do I run a game when my PCs have different approaches to combat?

Iterate over non-const variables in C++

Decreasing star size

AC contactor 1 pole or 2?

Spoken encryption

Examples of simultaneous independent breakthroughs

Why/when is AC-DC-AC conversion superior to direct AC-Ac conversion?

Does the Intel 8086 CPU have user mode and kernel mode?

Basic Questions on Wiener Filtering

Why was Sauron preparing for war instead of trying to find the ring?

Problem in styling a monochrome plot

How do I stop my characters falling in love?

Can two figures have the same area, perimeter, and same number of segments have different shape?

What's the difference between 2a and 10a charging options?

Would this neural network have short term memory?

Arbitrarily big neural networkWhy would neural networks be a particularly good framework for “embodied AI”?If a neural network approach becomes widely used within a real-world situation, how would one debug/understand/fix the outcome if in one case poor?Detect street and sidewalk surface in aerial imagery (neural network)When do you back-propagate errors through a Neural Network when using TD LambdaHow to create a task-graph based neural network?Why not teach to a NN not only what is true, but also what is not true?Neural Network for Optical Mark Recognition?Using an 'operation ID' as a neural network inputWould this NN for my chip outputs work?

.everyoneloves__top-leaderboard:empty,.everyoneloves__mid-leaderboard:empty,.everyoneloves__bot-mid-leaderboard:empty{ margin-bottom:0;
}

1

$begingroup$

I want to design a NN that can remember it's last 7 actions and use them as inputs. So for example it would be able to store words in it's memory. Therefore if it had a choice of 10 different actions, the number of words it could store is $10^7$.

Here is my design:

$$out_{n+1} = f(out_n, in_n)mathbf{N} + out_n.mathbf{M}$$

$$action_n = sigma(mathbf{N} cdot out_n)$$

Where $f$ represents some layered neural network. Some of the actions would be physical actions and some might be internal (such as thinking of the letter 'C').

Basically I want $out_n$ to be an array that keeps the last 6 action values and puts them back in. So $M$ will be the matrix:

$$begin{bmatrix}
0&1&0&0&0&0\
0&0&1&0&0&0\
0&0&0&1&0&0\
0&0&0&0&1&0\
0&0&0&0&0&1\
0&0&0&0&0&0
end{bmatrix}$$

i.e. it would drop the 6th item from it's memory.

and $N$ would be the vector:

$$begin{bmatrix}
1&0&0&0&0&0&0
end{bmatrix}$$

I think this would be equivalent to an equation of the form:

$$out_{n+1}=F(in_n,out_n,out_{n-1},out_{n-2},...,out_{n-6})$$

So I think this would be an advantage over an RNN since this model remembers precisely it's last 6 actions. But would this be better than an RNN or worse? One could increase it's memory to more than 7 quite easily.

I think it's basically the same archececture as an RNN except elinimating a lot of the connections. Is this a new design or a common design?

One problem with this design is that you might also want a memory that is over longer time periods (e.g. for actions that take more than one tick.) But that might be solved by enhancing the archecture.

neural-networks long-short-term-memory

share|improve this question

edited 7 hours ago

zooby

asked 8 hours ago

zoobyzooby

65644 silver badges1212 bronze badges

$endgroup$

add a comment | 

1

$begingroup$

I want to design a NN that can remember it's last 7 actions and use them as inputs. So for example it would be able to store words in it's memory. Therefore if it had a choice of 10 different actions, the number of words it could store is $10^7$.

Here is my design:

$$out_{n+1} = f(out_n, in_n)mathbf{N} + out_n.mathbf{M}$$

$$action_n = sigma(mathbf{N} cdot out_n)$$

Where $f$ represents some layered neural network. Some of the actions would be physical actions and some might be internal (such as thinking of the letter 'C').

Basically I want $out_n$ to be an array that keeps the last 6 action values and puts them back in. So $M$ will be the matrix:

$$begin{bmatrix}
0&1&0&0&0&0\
0&0&1&0&0&0\
0&0&0&1&0&0\
0&0&0&0&1&0\
0&0&0&0&0&1\
0&0&0&0&0&0
end{bmatrix}$$

i.e. it would drop the 6th item from it's memory.

and $N$ would be the vector:

$$begin{bmatrix}
1&0&0&0&0&0&0
end{bmatrix}$$

I think this would be equivalent to an equation of the form:

$$out_{n+1}=F(in_n,out_n,out_{n-1},out_{n-2},...,out_{n-6})$$

So I think this would be an advantage over an RNN since this model remembers precisely it's last 6 actions. But would this be better than an RNN or worse? One could increase it's memory to more than 7 quite easily.

I think it's basically the same archececture as an RNN except elinimating a lot of the connections. Is this a new design or a common design?

One problem with this design is that you might also want a memory that is over longer time periods (e.g. for actions that take more than one tick.) But that might be solved by enhancing the archecture.

neural-networks long-short-term-memory

share|improve this question

edited 7 hours ago

zooby

asked 8 hours ago

zoobyzooby

65644 silver badges1212 bronze badges

$endgroup$

add a comment | 

$begingroup$

I want to design a NN that can remember it's last 7 actions and use them as inputs. So for example it would be able to store words in it's memory. Therefore if it had a choice of 10 different actions, the number of words it could store is $10^7$.

Here is my design:

$$out_{n+1} = f(out_n, in_n)mathbf{N} + out_n.mathbf{M}$$

$$action_n = sigma(mathbf{N} cdot out_n)$$

Where $f$ represents some layered neural network. Some of the actions would be physical actions and some might be internal (such as thinking of the letter 'C').

Basically I want $out_n$ to be an array that keeps the last 6 action values and puts them back in. So $M$ will be the matrix:

$$begin{bmatrix}
0&1&0&0&0&0\
0&0&1&0&0&0\
0&0&0&1&0&0\
0&0&0&0&1&0\
0&0&0&0&0&1\
0&0&0&0&0&0
end{bmatrix}$$

i.e. it would drop the 6th item from it's memory.

and $N$ would be the vector:

$$begin{bmatrix}
1&0&0&0&0&0&0
end{bmatrix}$$

I think this would be equivalent to an equation of the form:

$$out_{n+1}=F(in_n,out_n,out_{n-1},out_{n-2},...,out_{n-6})$$

So I think this would be an advantage over an RNN since this model remembers precisely it's last 6 actions. But would this be better than an RNN or worse? One could increase it's memory to more than 7 quite easily.

I think it's basically the same archececture as an RNN except elinimating a lot of the connections. Is this a new design or a common design?

One problem with this design is that you might also want a memory that is over longer time periods (e.g. for actions that take more than one tick.) But that might be solved by enhancing the archecture.

neural-networks long-short-term-memory

share|improve this question

edited 7 hours ago

zooby

asked 8 hours ago

zoobyzooby

65644 silver badges1212 bronze badges

$endgroup$

share|improve this question

edited 7 hours ago

zooby

asked 8 hours ago

zoobyzooby

65644 silver badges1212 bronze badges

share|improve this question

edited 7 hours ago

zooby

asked 8 hours ago

zoobyzooby

65644 silver badges1212 bronze badges

asked 8 hours ago

zoobyzooby

65644 silver badges1212 bronze badges

asked 8 hours ago

zoobyzooby

65644 silver badges1212 bronze badges

1 Answer
1

active

oldest

votes

2

$begingroup$

Congrats, you have invented 1d convolution. Convolution combined with RNN would have some advantage over just RNN. Think about the perception field.
In this layer, you do aggregate $6$ values to one. Imagine two of them - it will be $36$ already, etc. But, in the end, you still need RNN at the end to aggregate a variable length to constant length.

share|improve this answer

edited 4 hours ago

nbro

5,66044 gold badges1515 silver badges3232 bronze badges

answered 8 hours ago

user8426627user8426627

2241111 bronze badges

$endgroup$

• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Well that's good! Glad I'm on the right track! (Not sure what you mean at the end about variable lengths).
  $endgroup$
  – zooby
  7 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  @zooby This is not a 1D CNN, its a non differentiable RNN. (actions must be sampled under some categorical distribution based on whats described). The only similarity to a 1d cnn is the sliding window
  $endgroup$
  – mshlis
  7 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Why is it non-differentiable ?
  $endgroup$
  – zooby
  7 hours ago
  
  
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  do you train with sequences of different lenght, right? also if you put output as input think about output may be wrong so you can consider to force-feeding ( expected data instead of output)
  $endgroup$
  – user8426627
  6 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  I could be wrong but generally actions are drawn from a distribution (that’s why you show one hot encodingns) and you can’t differentiate through a categorical distrib
  $endgroup$
  – mshlis
  6 hours ago
  

  
  
  
  

add a comment | 

Your Answer

StackExchange.ready(function() {
var channelOptions = {
tags: "".split(" "),
id: "658"
};
initTagRenderer("".split(" "), "".split(" "), channelOptions);

StackExchange.using("externalEditor", function() {
// Have to fire editor after snippets, if snippets enabled
if (StackExchange.settings.snippets.snippetsEnabled) {
StackExchange.using("snippets", function() {
createEditor();
});
}
else {
createEditor();
}
});

function createEditor() {
StackExchange.prepareEditor({
heartbeatType: 'answer',
autoActivateHeartbeat: false,
convertImagesToLinks: false,
noModals: true,
showLowRepImageUploadWarning: true,
reputationToPostImages: null,
bindNavPrevention: true,
postfix: "",
imageUploader: {
brandingHtml: "Powered by u003ca class="icon-imgur-white" href="https://imgur.com/"u003eu003c/au003e",
contentPolicyHtml: "User contributions licensed under u003ca href="https://creativecommons.org/licenses/by-sa/3.0/"u003ecc by-sa 3.0 with attribution requiredu003c/au003e u003ca href="https://stackoverflow.com/legal/content-policy"u003e(content policy)u003c/au003e",
allowUrls: true
},
noCode: true, onDemand: true,
discardSelector: ".discard-answer"
,immediatelyShowMarkdownHelp:true
});


}
});

draft saved

draft discarded

Sign up or log in

StackExchange.ready(function () {
StackExchange.helpers.onClickDraftSave('#login-link');
});

Sign up using Google

Sign up using Facebook

Sign up using Email and Password

Post as a guest

Name

Email

Required, but never shown

StackExchange.ready(
function () {
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fai.stackexchange.com%2fquestions%2f13622%2fwould-this-neural-network-have-short-term-memory%23new-answer', 'question_page');
}
);

Post as a guest

Name

Email

Required, but never shown

2

$begingroup$

Congrats, you have invented 1d convolution. Convolution combined with RNN would have some advantage over just RNN. Think about the perception field.
In this layer, you do aggregate $6$ values to one. Imagine two of them - it will be $36$ already, etc. But, in the end, you still need RNN at the end to aggregate a variable length to constant length.

share|improve this answer

edited 4 hours ago

nbro

5,66044 gold badges1515 silver badges3232 bronze badges

answered 8 hours ago

user8426627user8426627

2241111 bronze badges

$endgroup$

• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Well that's good! Glad I'm on the right track! (Not sure what you mean at the end about variable lengths).
  $endgroup$
  – zooby
  7 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  @zooby This is not a 1D CNN, its a non differentiable RNN. (actions must be sampled under some categorical distribution based on whats described). The only similarity to a 1d cnn is the sliding window
  $endgroup$
  – mshlis
  7 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Why is it non-differentiable ?
  $endgroup$
  – zooby
  7 hours ago
  
  
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  do you train with sequences of different lenght, right? also if you put output as input think about output may be wrong so you can consider to force-feeding ( expected data instead of output)
  $endgroup$
  – user8426627
  6 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  I could be wrong but generally actions are drawn from a distribution (that’s why you show one hot encodingns) and you can’t differentiate through a categorical distrib
  $endgroup$
  – mshlis
  6 hours ago
  

  
  
  
  

add a comment | 

2

$begingroup$

Congrats, you have invented 1d convolution. Convolution combined with RNN would have some advantage over just RNN. Think about the perception field.
In this layer, you do aggregate $6$ values to one. Imagine two of them - it will be $36$ already, etc. But, in the end, you still need RNN at the end to aggregate a variable length to constant length.

share|improve this answer

edited 4 hours ago

nbro

5,66044 gold badges1515 silver badges3232 bronze badges

answered 8 hours ago

user8426627user8426627

2241111 bronze badges

$endgroup$

• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Well that's good! Glad I'm on the right track! (Not sure what you mean at the end about variable lengths).
  $endgroup$
  – zooby
  7 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  @zooby This is not a 1D CNN, its a non differentiable RNN. (actions must be sampled under some categorical distribution based on whats described). The only similarity to a 1d cnn is the sliding window
  $endgroup$
  – mshlis
  7 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Why is it non-differentiable ?
  $endgroup$
  – zooby
  7 hours ago
  
  
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  do you train with sequences of different lenght, right? also if you put output as input think about output may be wrong so you can consider to force-feeding ( expected data instead of output)
  $endgroup$
  – user8426627
  6 hours ago
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  I could be wrong but generally actions are drawn from a distribution (that’s why you show one hot encodingns) and you can’t differentiate through a categorical distrib
  $endgroup$
  – mshlis
  6 hours ago
  

  
  
  
  

add a comment | 

$begingroup$

Congrats, you have invented 1d convolution. Convolution combined with RNN would have some advantage over just RNN. Think about the perception field.
In this layer, you do aggregate $6$ values to one. Imagine two of them - it will be $36$ already, etc. But, in the end, you still need RNN at the end to aggregate a variable length to constant length.

share|improve this answer

edited 4 hours ago

nbro

5,66044 gold badges1515 silver badges3232 bronze badges

answered 8 hours ago

user8426627user8426627

2241111 bronze badges

$endgroup$

share|improve this answer

edited 4 hours ago

nbro

5,66044 gold badges1515 silver badges3232 bronze badges

answered 8 hours ago

user8426627user8426627

2241111 bronze badges

edited 4 hours ago

nbro

5,66044 gold badges1515 silver badges3232 bronze badges

edited 4 hours ago

nbro

5,66044 gold badges1515 silver badges3232 bronze badges

answered 8 hours ago

user8426627user8426627

2241111 bronze badges

answered 8 hours ago

user8426627user8426627

2241111 bronze badges