Image Segmentation Goal: to learn about what Fully Convolutional Networks (FCN) are and how they can be used for image segmentation. 🤔 Semantic Segmentation a simple way is to slide the window, creating a slow RNN? another way is to build a Equal size Fully Convolutional Network (but receptive field is limited and hard to generalize to higher resolutions) U-Net like architecture 😉 but how to UPSAMPLE the feature maps?

Recap and Conclusion Recap What is Computer Vision? building a system that can understand and process visual information Main approaches to Computer Vision semantic segmentation using Fully Convolutional Networks (FCN) 🤔 What’s next? we will discover interesting new types of deep models Neural ODE NIPS 2018 $$ h_{t+1} = h_t + f(h_t, \theta_t) \Rightarrow \frac{dh}{dt} = f(h(t), t, \theta) $$ Deep Learning will find new applications AI4S DL4CS, like using DL training a hash function to improve the hash table DL4Mathematics, auto theorem proving Deep Learning will use more data and more computional power Problems and Challenges Models are biased eg: Vector Arithmetic with Word Vectors, fit biased dataset to a model leads to poor performance; Economic Bias (usually train model on the wealthier western family?

Generative Model II for some reason, it seems like that we are backed 😆 Generative Adversarial Networks (GANs) setup: assume that we have data $x_i$ from a distribution $p_{data}(x)$, all we wanna do is to sample from $p_{data}(x)$ idea: introduce a latent variable $z$ with simple prior $p(z)$, those $z$ can be interpolated, sample $z$ from $p(z)$, and pass into a Generator $x = G(z)$, then we said that $x$ is a sample from the Generator distribution $p_{G}$,