Ttdfjt

What are the examples (applications) of the MIPs in which the objective function has nonzero coefficients for only continuous variables?

Add room number to postal address?

Print only the last three columns from file

The heat content of the products is more than that of the reactant in an ............. reaction

What could prevent players from leaving an island?

Did Captain America make out with his niece?

Authenticating SOAP API via UsernameToken

Purchased new computer from DELL with pre-installed Ubuntu. Won't boot. Should assume its an error from DELL?

Does the spell "Silence" affect the caster?

What is an air conditioner compressor hard start kit and how does it work?

Traveling from Germany to other countries by train?

Is Odin inconsistent about the powers of Mjolnir?

Will a paper be retracted if a flaw in released software code invalidates its central idea?

"In charge of" vs "Responsible for"

monolingual dictionary

How do I get the =LEFT function in excel, to also take the number zero as the first number?

How to switch an 80286 from protected to real mode?

How to help new students accept function notation

"To go from zero to hero"

Determine Beckett Grading Service (BGS) Final Grade

Repeated! Factorials!

Did silent film actors actually say their lines or did they simply improvise “dialogue” while being filmed?

12V lead acid charger with LM317 not charging

Unexpected route on a flight from USA to Europe

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:

$$out_n+1 = f(out_n, in_n)mathbfN + out_n.mathbfM$$

$$action_n = sigma(mathbfN cdot out_n)$$

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:

$$beginbmatrix
0&1&0&0&0&0\
0&0&1&0&0&0\
0&0&0&1&0&0\
0&0&0&0&1&0\
0&0&0&0&0&1\
0&0&0&0&0&0
endbmatrix$$

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

and $N$ would be the vector:

$$beginbmatrix
1&0&0&0&0&0&0
endbmatrix$$

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 Jul 27 at 19:52

zooby

asked Jul 27 at 18:41

zoobyzooby

67144 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:

$$out_n+1 = f(out_n, in_n)mathbfN + out_n.mathbfM$$

$$action_n = sigma(mathbfN cdot out_n)$$

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:

$$beginbmatrix
0&1&0&0&0&0\
0&0&1&0&0&0\
0&0&0&1&0&0\
0&0&0&0&1&0\
0&0&0&0&0&1\
0&0&0&0&0&0
endbmatrix$$

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

and $N$ would be the vector:

$$beginbmatrix
1&0&0&0&0&0&0
endbmatrix$$

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 Jul 27 at 19:52

zooby

asked Jul 27 at 18:41

zoobyzooby

67144 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:

$$out_n+1 = f(out_n, in_n)mathbfN + out_n.mathbfM$$

$$action_n = sigma(mathbfN cdot out_n)$$

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:

$$beginbmatrix
0&1&0&0&0&0\
0&0&1&0&0&0\
0&0&0&1&0&0\
0&0&0&0&1&0\
0&0&0&0&0&1\
0&0&0&0&0&0
endbmatrix$$

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

and $N$ would be the vector:

$$beginbmatrix
1&0&0&0&0&0&0
endbmatrix$$

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 Jul 27 at 19:52

zooby

asked Jul 27 at 18:41

zoobyzooby

67144 silver badges1212 bronze badges

$endgroup$

share|improve this question

edited Jul 27 at 19:52

zooby

asked Jul 27 at 18:41

zoobyzooby

67144 silver badges1212 bronze badges

share|improve this question

edited Jul 27 at 19:52

zooby

asked Jul 27 at 18:41

zoobyzooby

67144 silver badges1212 bronze badges

asked Jul 27 at 18:41

zoobyzooby

67144 silver badges1212 bronze badges

asked Jul 27 at 18:41

zoobyzooby

67144 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 Jul 27 at 22:33

nbro

6,16444 gold badges1616 silver badges3333 bronze badges

answered Jul 27 at 18:50

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
  Jul 27 at 19:14
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  @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
  Jul 27 at 19:40
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Why is it non-differentiable ?
  $endgroup$
  – zooby
  Jul 27 at 19:42
  
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  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
  Jul 27 at 20:14
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  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
  Jul 27 at 20:30
  

  
  
  
  

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 Jul 27 at 22:33

nbro

6,16444 gold badges1616 silver badges3333 bronze badges

answered Jul 27 at 18:50

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
  Jul 27 at 19:14
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  @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
  Jul 27 at 19:40
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Why is it non-differentiable ?
  $endgroup$
  – zooby
  Jul 27 at 19:42
  
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  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
  Jul 27 at 20:14
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  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
  Jul 27 at 20:30
  

  
  
  
  

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 Jul 27 at 22:33

nbro

6,16444 gold badges1616 silver badges3333 bronze badges

answered Jul 27 at 18:50

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
  Jul 27 at 19:14
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  @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
  Jul 27 at 19:40
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  Why is it non-differentiable ?
  $endgroup$
  – zooby
  Jul 27 at 19:42
  
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  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
  Jul 27 at 20:14
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  $begingroup$
  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
  Jul 27 at 20:30
  

  
  
  
  

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 Jul 27 at 22:33

nbro

6,16444 gold badges1616 silver badges3333 bronze badges

answered Jul 27 at 18:50

user8426627user8426627

2241111 bronze badges

$endgroup$

share|improve this answer

edited Jul 27 at 22:33

nbro

6,16444 gold badges1616 silver badges3333 bronze badges

answered Jul 27 at 18:50

user8426627user8426627

2241111 bronze badges

edited Jul 27 at 22:33

nbro

6,16444 gold badges1616 silver badges3333 bronze badges

edited Jul 27 at 22:33

nbro

6,16444 gold badges1616 silver badges3333 bronze badges

answered Jul 27 at 18:50

user8426627user8426627

2241111 bronze badges

answered Jul 27 at 18:50

user8426627user8426627

2241111 bronze badges