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?
How to Create an Image for Cantor's *Diagonal Argument* with a Diagonal Oval
Iterate over non-const variables in C++
What does コテッと mean?
Is it legal for private citizens to "impound" e-scooters?
Can the 2019 UA Artificer's Returning Weapon and Radiant Weapon infusions stack on the same weapon?
Send a single HTML email from Thunderbird, overriding the default "plain text" setting
Why/when is AC-DC-AC conversion superior to direct AC-Ac conversion?
Why are all my history books dividing Chinese history after the Han dynasty?
What to do when you reach a conclusion and find out later on that someone else already did?
Why are so many countries still in the Commonwealth?
Explain why watch 'jobs' does not work but watch 'ps' work?
Basic Questions on Wiener Filtering
Can I make a matrix from just a parts of the cells?
Decreasing star size
Weed in Massachusetts: underground roots, skunky smell when bruised
Why are off grid solar setups only 12, 24, 48 VDC?
How can I receive packages while in France?
Marrying a second woman behind your wife's back: is it wrong and can Quran/Hadith prove this?
Request for a Latin phrase as motto "God is highest/supreme"
High income, sudden windfall
Where to place an artificial gland in the human body?
What do I do when a student working in my lab "ghosts" me?
TSA asking to see cell phone
Why can't my huge trees be chopped down?
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;
$begingroup$
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
asked 9 hours ago
zoobyzooby
6564 silver badges12 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
asked 9 hours ago
zoobyzooby
6564 silver badges12 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
asked 9 hours ago
zoobyzooby
6564 silver badges12 bronze badges
$endgroup$
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
neural-networks long-short-term-memory
asked 9 hours ago
zoobyzooby
6564 silver badges12 bronze badges
asked 9 hours ago
zoobyzooby
6564 silver badges12 bronze badges
edited 8 hours ago
zooby
asked 9 hours ago
zoobyzooby
6564 silver badges12 bronze badges
asked 9 hours ago
zoobyzooby
6564 silver badges12 bronze badges
asked 9 hours ago
zoobyzooby
6564 silver badges12 bronze badges
6564 silver badges12 bronze badges
add a comment |
add a comment |
1 Answer
1
active
oldest
votes
$begingroup$
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.
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
answered 9 hours ago
user8426627user8426627
22411 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
8 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
8 hours ago
$begingroup$
Why is it non-differentiable ?
$endgroup$
– zooby
8 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
7 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
7 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
);
);
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
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
Required, but never shown
1 Answer
1
active
oldest
votes
1 Answer
1
active
oldest
votes
active
oldest
votes
active
oldest
votes
$begingroup$
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.
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
answered 9 hours ago
user8426627user8426627
22411 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
8 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
8 hours ago
$begingroup$
Why is it non-differentiable ?
$endgroup$
– zooby
8 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
7 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
7 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.
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
answered 9 hours ago
user8426627user8426627
22411 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
8 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
8 hours ago
$begingroup$
Why is it non-differentiable ?
$endgroup$
– zooby
8 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
7 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
7 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.
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
answered 9 hours ago
user8426627user8426627
22411 bronze badges
$endgroup$
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.
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
answered 9 hours ago
user8426627user8426627
22411 bronze badges
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
edited 5 hours ago
nbro
5,6604 gold badges15 silver badges32 bronze badges
5,6604 gold badges15 silver badges32 bronze badges
answered 9 hours ago
user8426627user8426627
22411 bronze badges
answered 9 hours ago
user8426627user8426627
22411 bronze badges
answered 9 hours ago
user8426627user8426627
22411 bronze badges
22411 bronze badges
$begingroup$
Well that's good! Glad I'm on the right track! (Not sure what you mean at the end about variable lengths).
$endgroup$
– zooby
8 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
8 hours ago
$begingroup$
Why is it non-differentiable ?
$endgroup$
– zooby
8 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
7 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
7 hours ago
add a comment |
$begingroup$
Well that's good! Glad I'm on the right track! (Not sure what you mean at the end about variable lengths).
$endgroup$
– zooby
8 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
8 hours ago
$begingroup$
Why is it non-differentiable ?
$endgroup$
– zooby
8 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
7 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
7 hours ago
$begingroup$
Well that's good! Glad I'm on the right track! (Not sure what you mean at the end about variable lengths).
$endgroup$
– zooby
8 hours ago
$begingroup$
Well that's good! Glad I'm on the right track! (Not sure what you mean at the end about variable lengths).
$endgroup$
– zooby
8 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
8 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
8 hours ago
$begingroup$
Why is it non-differentiable ?
$endgroup$
– zooby
8 hours ago
$begingroup$
Why is it non-differentiable ?
$endgroup$
– zooby
8 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
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
7 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
7 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
7 hours ago
add a comment |
Thanks for contributing an answer to Artificial Intelligence Stack Exchange!
- Please be sure to answer the question. Provide details and share your research!
But avoid …
- Asking for help, clarification, or responding to other answers.
- Making statements based on opinion; back them up with references or personal experience.
Use MathJax to format equations. MathJax reference.
To learn more, see our tips on writing great answers.
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
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
Required, but never shown
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
Required, but never shown
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
Required, but never shown
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
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
