Deep Learning

Full Course

Class Package

Class Project(s)

1 project per half semester.

Virtual Machine (VM)

A Virtual Machine is a remote desktop that allows students to connect to it from anywhere. We provide VMs so that students use it during classes and to work on homework.

Student Progress Report

Students will get personalized progress reports and feedback from the instructor

Class Description:

Learn the most modern techniques for supervised learning, used in common applications such as facial recognition, speech recognition, and self driving cars. This course will also provide students with a linux server with GPU acceleration to run their algorithms. Topics include regression, test classification, convolutional image recognition, and more.

Features:

• This course gives students access to professional research grade hardware, including compute servers such as a 16 core server with 128GB RAM, terabytes of fast storage, and research grade graphics processors such as the Titan X Pascal GPU
• This course also teaches students how to use linux tools and the CUDA + GPU accelerated python research environment, using Tensorflow-GPU 1.4, Keras 2.1.4 - tools compiled with lubcudnn 6 and 7
• Course material draws from recent academic research published in the last 2-5 years, including deep networks, single shot detection, convolutional or vectorizated models for language, as well as (time permitting) demo projects featuring AlphaZero Go and GAN inspired sequence to sequence learning.

When students move from [AI 1] to [AI 2], their focus shifts to getting the best possible model accuracies on real world data. Half the homework involves using pre-formatted data sets, while the second half involves students finding their own data sets. Students taking [AI 2] apply neural networks to text, images, and other data. Because the course is applied, much of [AI 2] involves the practice of parsing data and preparing it for use on research servers. This includes the mastery of linux command line tools, as well gaining a familiarity with different formats for data such as comma separated files, JSON APIs, and all sorts of image file types. Although the models are written in python, the class overall is language agnostic, as students pick up different tools to most effectively deal with different data. Indeed, students are expected to become proficient in the entire research lifecycle, from coming up to a hypothesis to explaining their model results.

This course no longer uses Theano, and students will model primarily with the Keras deep learning library backed by Tensorflow.

Research from KTBYTE students and alumni

Prerequisites:

Completion of [CORE 5b] or AP CS, or permission of instructor. Also requires Algebra II math experience. [AI 1] highly recommended but not required.