horn-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Edward J. Yoon" <edward.y...@samsung.com>
Subject RE: Ready for supporting RNN and LSTM ?
Date Mon, 19 Sep 2016 00:49:59 GMT
As far as I know, there's not much difference between an RNN and feed-forward network, so I
think you can try to implement it.

 

Before implementing complex nets, Elman nets can be a good start: http://mnemstudio.org/neural-networks-elman.htm

 

Elman net have an additional context layer that connected to the hidden layer with one-to-one
connections. The state of the hidden units is 'copied back' to the context units, so that
it becomes available as part of the input to the network on the next time step.

 

+ One thing we need to think carefully is the user-side programming interface for managing
the states of neurons' value. I'm not sure but I roughly guess, ... the context layer can
be exist logically. Each hidden neuron unit object can keep the state and get the all the
states of context layer. 

 

--

Best Regards, Edward J. Yoon

 

 

-----Original Message-----
From: Yeonhee Lee [mailto:ssallys0130@gmail.com] 
Sent: Monday, September 12, 2016 6:28 PM
To: dev@horn.incubator.apache.org
Subject: Ready for supporting RNN and LSTM ?

 

Dear edward and team,

 

Suddenly, a straight question came across my mind.

Does our platform is ready to implement LSTM or RNN stuffs ?

 

Since Horn runs on a bigdata platform, our system can handle a large-volume

of data easier compared to other deep learning platform such as tensorflow

and caffe.

And thus, in my intuition, online and incremental learning can be a kind of

proper application.

So, I wonder whether our system is stable and has basic elements to

implement it.

And I also want to receive your opinion on implementing LSTM.

 

Best regards,

Yeounhee


Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message