AI@UCF: Fall 2018

Welcome back to SIGAI! We'll be re-introducing SIGAI for newcomers and refreshing it for veterans. Following that, we'll cover some basics of generating graphs (a very common task for data science and research). If you're enticed, we'll also get you setup on the university's supercomputer, as all following meetings will stream from there! :smiley:

Data is arguably more important than the algorithms we'll be learning this semester - and that data almost always needs to be curated and finagled to really develop an understanding of what the data is trying to tell you.

Neural Networks are all the rage, nowadays, but simpler models are always great places to start! We'll cover how to do Linear/Logistic Regression as well as preparing data for such a function to work.

With some basic ideas in mind about how one might tackle a task, we'll now go and explore a Tensor framework (PyTorch) and build a Neural Network which can accurately classify handwritten digits, as well as articles of clothing.

Convolutional Neural Networks are at the forefront of Deep Learning and they enable machines to "see" much more effectively than they used to. So well, in fact, that they can tell what's in an image, or even place points onto them.

Fully Connect Neural Networks and Convolutional Neural Networks are absolutely wonderful, but they miss out on one key component of our world, time. This week we'll look at Networks which also model time as part of their inputs – because of this, they'll be able to write nearly as well as Shakespeare! 😃

We're filling this out!

The world is complex, making it difficult for algorithms to come to solutions in reasonable amounts of time. To speed them along, we can employ Heuristics to get us significantly closer, faster. Today, we’ll try to approximate the Traveling Salesman Problem by using Simulated Annealing and Particle Swarm Optimization – two Heuristics which move us towards finding the shortest path we can use to visit all the destinations.

Sometimes the algorithms we use to predict the future can be difficult to interpret and trust. A Decision Tree is a learning algorithm that does a half decent job at prediction, but, more importantly, is very easy to understand and interpret. No black boxes here... until we start talking about Random Forests.

Support Vector Machines are a simple and powerful classification algorithm that perform well in nearly every situation. They're commonly used in image recognition, face detection, bioinformatics, handwriting recognition, and text categorization. The math behind it is pretty cool, as it relies upon embedding data into higher dimensional space to create linear divisions between categories. SVMs are a great resource to add to your data science toolkit, as they're relatively simple to understand and are also one of the best classification algorithms that do not involve neural networks.

We're filling this out!