December 08, 2003
Morning, 08/DEC/2003
Random news:- T.G.I. Friday's, Applebee's, and Ruby Tuesday are introducing low-carb menu items for the Atkin dieter in you. According the nice man on the excellently irreverent Daily Buzz (P.S. Ron is adorable), 20% of Friday's customers are on Atkins. This is great news for carb nazis everywhere.
- Our kitchen sink drains overnight. Last week two a day and a half had pass after a dishwasher run before the sink was usable. This is mixed news. Sure, our kitchen is closer to both "normal" and "working", but self-repair was--I thought--a property exclusive to biological systems.
- I am unable to predict the Mackey Glass timeseries (with tau=30, sampled at T=6, aka MG30) beyond a 50 step horizon, despite extensive training, heavy tweaking and evidence that it can be accurately predicted for 100 steps. Typical results:
That's 80 steps. MG30 is blue and my predictor is red.
Comments
#1 Noooooooo! So does this mean that all of their menus are going to devolve into various combinations of bacon and cabbage?
#3 Neural networks are over-prescribed! Use something that actually models the system internals :) Yeah yeah I know.
Posted by: Joel at December 8, 2003 11:20 AM#3 Your "evidence" for 100 steps (Figure 5) is not very convincing. Your 80 steps are better than Principe's. And he earlier claims that 45 steps is a good time to restart your model.
Posted by: chris at December 11, 2003 11:40 AMI skimmed his paper too quickly to catch the 45-step reset--that's good to hear. Once trained, I'll see both abysmal performance (<20 steps) and celestial performance (>100 steps) come out of the same network. I wish I had my 7 dimensional display so I could see what the attractor looks like.
Posted by: Tom at December 11, 2003 11:57 AMPost a comment