it's the 10-15% of the CPU part that worries me, that is a significant chunk of proccessing power for just the input device and means that if the developer wants NATAL suport in a game they will have to tone down the complexity of the AI, physics or reduce the number of objects on screen at any point. That means that it's probably not going to be used in games that are not designed spacifically for NATAL, So ideas like having hand gestures in Halo Reach won't be happening unless Bungie can reduce the CPU requirements by using NATAL for just that specific task and will probably mean that 80% of the games that support NATAL will be mini game collections and Milo. and will probably mean that NATAL is doomed.

Well, we don't know yet how much comprehensive the "training" has been for Natal's software. We have to hope that they gave it a very good coverage of the "human poses" alphabet. But as I said, there's an extra ace up the sleeve, that is knowledge about kinetic limitations of human body parts.

Even if you have to strike a really weird pose, you will have to get there from a normal one through intermediate, less weird poses. The software might be able to provide a good interpretation of the sensor's input just by extrapolating previous data of body parts' positions and speeds. I suppose that's what it will do not only when a limb is occluded from sight, but also when it has a very low matching score for what it sees.

It's an interesting problem, I guess we'll have to wait and see. I'm sure it will have plenty of quirks, my previous post was just meant to expand on the a priori knowledge base topic.

And you guys are hitting on a potential positive about having the software running outside of the hardware. It is easier for updates to the knowledge base of the network. MS will most likely work with developers on the software and provide updates when necessary since each game disc will most likely have the data on it.

A neural network is basically a series of weighted nodes that take a set of inputs and output a set of output value(s). Training the network alters the weights of the nodes until an acceptable accuracy is achieved for all input sets. I did a neural network in college where the inputs(actually, the input names are irrelavent since it was just float values, the network could handle any number of inputs and any number of expected outputs) were enemy distance, ammo, health, and weapon and the outputs were attack, defend, or stay. For training purposes you provide the inputs with the expected outputs and let the system keep iterating over the training sets and altering node weights until each set is met with an acceptable outcome.

A funny story that I read once about neural networks... Some military asked for a neural network to distinguish between images of tanks and cars. It worked perfectly during the training process, and totally failed when used in reality. Why? All the pictures they took of tanks for the training were taken on a rainy day, and the pictures of cars were taken on a sunny day. So actually the neural network learned to recognize the weather, not cars and tanks as they wanted it to :D

Of course this is very unlikely happen in systems which are actually tested during development... but it's still a funny story.

^^That is pretty funny.