Using linear parameterizations can be seen to fail even when the model is linear if we don’t have enough data. In this post, we will see that neural networks (NN) can success in learning non-linear models, but this is only true if we have sufficient data. In this post we will work with the simplest … Read more
Download in pdf format We consider a learning problem with input observations $\bbx\in\reals^n$ and output information $\bby\in\reals^m$. We use a linear learning parametrization that we want to train to predict outputs as $\hby=\bbH\bbx$ that are close to the real $\bby$. The comparison metric between $\bby$ and $\hby$ is the squared Euclidean error $\ell(\bby, \hby) = … Read more
We can define convolutions for time signals, convolutions for images, and convolutions on graphs. These three operations have deep connections that we explore in this post.
Read this post if you want to find out about the papers on which this course is based.