Comments on: Log for last few weeks: November 27, 2009 http://www.fsavard.com/flow/2009/11/log-for-last-few-weeks-november-27-2009/ Programming, personal knowledge management. Topics unstable. Thu, 01 Jul 2010 12:30:19 +0000 hourly 1 http://wordpress.org/?v=3.0 By: Francois http://www.fsavard.com/flow/2009/11/log-for-last-few-weeks-november-27-2009/comment-page-1/#comment-7337 Francois Sat, 28 Nov 2009 14:04:05 +0000 http://www.fsavard.com/flow/?p=542#comment-7337 Hi Hossein, I'm surprised you found out about my blog so quick! :-) The perceptron/really scaled down problem is another good idea, and I think I'll try it too when I get time to resume working on this. There were two factors motivating using a CNN for the toy problem. The first is that I wasn't really sure how simple the architecture could get until the embedding cost became irrelevant/inapplicable. (Yet if it's supposed to work at such a small scale, then why not? My goal is also to boil it down to its simplest element, to understand what's going on.) (The second factor is that, at any rate, I needed some toy problem to demonstrate the approach works for the term project I'm working on this for.) Thanks again for the suggestions. Hi Hossein, I’m surprised you found out about my blog so quick! :-)

The perceptron/really scaled down problem is another good idea, and I think I’ll try it too when I get time to resume working on this.

There were two factors motivating using a CNN for the toy problem. The first is that I wasn’t really sure how simple the architecture could get until the embedding cost became irrelevant/inapplicable. (Yet if it’s supposed to work at such a small scale, then why not? My goal is also to boil it down to its simplest element, to understand what’s going on.)

(The second factor is that, at any rate, I needed some toy problem to demonstrate the approach works for the term project I’m working on this for.)

Thanks again for the suggestions.

]]> By: Hossein Mobahi http://www.fsavard.com/flow/2009/11/log-for-last-few-weeks-november-27-2009/comment-page-1/#comment-7336 Hossein Mobahi Sat, 28 Nov 2009 12:42:51 +0000 http://www.fsavard.com/flow/?p=542#comment-7336 By linearly separable, I did not mean perfect separation... I meant that by minimizing the squared error you should end up getting a good solution. By linearly separable, I did not mean perfect separation… I meant that by minimizing the squared error you should end up getting a good solution.

]]> By: Hossein Mobahi http://www.fsavard.com/flow/2009/11/log-for-last-few-weeks-november-27-2009/comment-page-1/#comment-7335 Hossein Mobahi Sat, 28 Nov 2009 08:33:04 +0000 http://www.fsavard.com/flow/?p=542#comment-7335 I am sorry that bug is not resolved yet. Maybe you want to use a single perceptron instead of a CNN for the toy problem. The toy problem is linearly separable, so using a single perceptron neuron is enough and it should simplify monitoring the weights and hopefully point you to the problem. I am sorry that bug is not resolved yet. Maybe you want to use a single perceptron instead of a CNN for the toy problem. The toy problem is linearly separable, so using a single perceptron neuron is enough and it should simplify monitoring the weights and hopefully point you to the problem.

]]>