Thank you for posting this great video, Professor! This technique seems quite simple, yet extremely powerful. I am curious what types of otherwise intractable systems might become tractable if one uses this technique with more powerful modern NNs.
Hi Professor Kutz, I was wondering what the intuition was behind choosing your three activation functions (lines 25-27). I've seen logsig and tanh functions used a lot for the hidden layers of regression networks but I'm not sure what the radial basis function brings to the table and why it is placed in the middle. In addition, I thought the output layer for matlab's regression networks is 'purelin' by default so is the third hidden layer a bit redundant? Or is there a reason you chose to have the last hidden layer to have a purelin activation function?
What might be the advantage of training a NN to solve an IVP? It seems that to train a NN to solve for the IVP one must already have a way to generate trajectories.
I was searching around for a neural network architecture that works by taking in an input multi-modal "image" at time T, and predicts multi-modal output at tiime T+1. But, how would you represent actuators in this scheme? The position of a motor as an input and predicted output (y_a1 - x_a1) ... would that diff trigger a motor to actuate the world? since LLMs are all based on next-frame prediction, it seems that you could make a generic box and hook up some inputs to pixels, other inputs to positions of actuators; but it's mysterious how the output would control the actuators. ie: if a neural network is trying to minimize surprise, then it can minimize surprise by actuating the world on its own.
One of the best professors in system analysis.
Pro tip : watch movies on Flixzone. Me and my gf have been using them for watching loads of movies these days.
@Yahir Rohan definitely, I've been watching on flixzone for months myself :D
@Yahir Rohan Definitely, have been using flixzone for since december myself :)
Beauty of System Identification!
Excited to see how this can be used for IK in robotics, thanks for your time.
the best demonstrating way i have even seen
Thank you for posting this great video, Professor!
Thank you Nathan
Thank you for posting this great video, Professor! This technique seems quite simple, yet extremely powerful. I am curious what types of otherwise intractable systems might become tractable if one uses this technique with more powerful modern NNs.
Hi Professor Kutz,
I was wondering what the intuition was behind choosing your three activation functions (lines 25-27). I've seen logsig and tanh functions used a lot for the hidden layers of regression networks but I'm not sure what the radial basis function brings to the table and why it is placed in the middle.
In addition, I thought the output layer for matlab's regression networks is 'purelin' by default so is the third hidden layer a bit redundant? Or is there a reason you chose to have the last hidden layer to have a purelin activation function?
Lovely work!
All my respect to you, master!!
Great video Nathan, I recently got your other book "Data-Driven modeling &Scienfitif computation" I'm enjoying
What might be the advantage of training a NN to solve an IVP? It seems that to train a NN to solve for the IVP one must already have a way to generate trajectories.
Really thank you so much Sir.
Great way to teach.
Thank you. Really interesting
Great video!
Hi, thanks for the interesting vedio. Is there a method that we could include the "b; sig; r;" as input paramaters for nerual network?
I was searching around for a neural network architecture that works by taking in an input multi-modal "image" at time T, and predicts multi-modal output at tiime T+1. But, how would you represent actuators in this scheme? The position of a motor as an input and predicted output (y_a1 - x_a1) ... would that diff trigger a motor to actuate the world? since LLMs are all based on next-frame prediction, it seems that you could make a generic box and hook up some inputs to pixels, other inputs to positions of actuators; but it's mysterious how the output would control the actuators. ie: if a neural network is trying to minimize surprise, then it can minimize surprise by actuating the world on its own.
this as opposed to encoding actuators in weights somewhere; and letting backprop modify them.
Does someone know how this approach changes when you also consider an input, u, in your system?
did you get the answer
Most complicated example I’ve seen in my entire life. And I’m very old.
2:08