« The Self-Organized Gene (Part 2)
The Synapse Spiral »

Peltarion @ ICANN’07

We will have representatives at the 2007 International Conference on Artificial Neural Networks in Porto, Portugal between 9/9 and 14/9.

For more information see: http://www.icann2007.org/

This entry was posted on Thursday, September 6th, 2007 at 16:08 and is filed under Synapse. You can follow any responses to this entry through the RSS 2.0 feed. You can leave a response, or trackback from your own site.

4 Responses to “Peltarion @ ICANN’07”

  1. Craig Denison Says:
    September 18th, 2007 at 20:02

    Hi Luka…glad to see your announcement about the ICANN’07. How about some more details? I looked at the conference website, but found no mention of Peltarion as a paper presenter, so I presume your presence there is as an exhibitor, but not presenting any ‘papers’, right? I am very curious to here how/what you are doing in the area of time-series forecasting, and any change of mind on supporting reinforcement learning paradigm?

  2. Craig Says:
    September 18th, 2007 at 20:04

    Hi Luka…glad to see your announcement about the ICANN’07. How about some more details? I looked at the conference website, but found no mention of Peltarion as a paper presenter, so I presume your presence there is as an exhibitor, but not presenting any ‘papers’, right? I am very curious to hear how/what you are doing in the area of time-series forecasting, and any change of mind on supporting reinforcement learning paradigm?

  3. Luka (Peltarion) Says:
    September 25th, 2007 at 1:28

    Hi Craig,

    No we didn’t present a paper on this conference. We demonstrated Synapse a bit and we got a lot of good new ideas for Synapse. Regarding time series forecasting, you’ll be happy to hear that we have an LSTM component in the pipeline that we will be releasing soon.

    We did also have an interesting discussion with Schmidhuber’s team from IDSIA (http://www.idsia.ch/) on reinforcement learning using LSTMs. They had some really cool stuff to show.

    The problem with Synapse and RL remains the same - actions and states are completely problem specific and there isn’t any way of putting it in a general package. Synapse is entirely GUI based while any RL implementation requires a large amount of user written code tailored to the problem. I’d love to have RL in Synapse but it is unfortunately a question of apples and oranges. What we have discussed is to make an RL deployment template so that the system you build in Synapse is at the core of an RL wrapper. It is however of questionable value as the general wrapper part is so trivial that it is a question of a few lines of code. The real work is in implementing the state/action handling - which is entirely problem specific.

    -Luka

  4. Craig Says:
    October 1st, 2007 at 2:42

    Hi Luka!
    Good to ‘talk’ to you again. Would love to hear more about the RL deployment template idea that would allow me to use Synapse for implementing all of the neural-network related stuff that might be associated with ‘estimating’ the ’state’. The ‘action’ part is the most domain specific, and probably implemented without any need for neural-networks, right? The most often type of RL application is the ‘trading-system’ type application. The typical approach for that application is to implement a estimated ‘target’ time-series that is used in a function-approximation style implementation, but using a combination of ‘profit-cost’ and function ‘error-accuracy’ for the training. Maybe a user-forum would also be very useful to everyone to learn how to apply RL, using Synapse, for any given problem domain?

Leave a Reply