About the course
Can artificial intelligence be truly creative? One of the most compelling applications of artificial intelligence and deep learning has been in computer-assisted music composition and artwork, as these applications more than any other pose this fundamental question. While deep learning has not progressed far enough to resolve this question, deep learning or AI-assisted composition has radically changed art and media in the past decade, and new advances in machine learning techniques are rapidly pushing the state of the art.
In this one day workshop, we will explore AI-generated music composition and image stylization. In the first part of the workshop, we will discuss how sequential models and ensembles of autoregressive structures can learn musical sequences for both monophonic and polyphonic composition. We will explore the use of the deep learning toolkit Magenta to compose and interpolate musical sequences. In the second part of the workshop, we will convert our use of convolutional neural networks from image classification or object detection problems to try to interpret the convolutional layers. This will pave the way for deep styles and image interpolation and an exploration of the DeepDream toolkit to generate these images.
Finally, it is important to note that unsettling applications such as deepfakes or chatbots also make use of similar AI creativity tools, making it critical for data scientists and AI professionals to understand the intricacies of these models in order to detect or prevent their harmful use. By diving into human expression outputs of the tools we use in our pragmatic pursuits, we will not only be able to understand their implementation and usage in better detail, but will also be able to avoid pitfalls and biases that routinely crop up in the widespread application of data products.
Upon successful completion of the course, students will:
Understand the use of sequential neural modeling techniques for music composition.
Be able to use the Magenta toolkit to compose music sequences.
Understand convolutional neural networks architectures and layer interpretations.
Use loss ascent (rather than descent) to generate highly stylized images.
Understand the use of generative adversarial networks for image generation.
Discuss the impact of superior AI creativity that can easily fool human observers.
Explore case studies of the impact of deep fakes and generated portraits.
Benjamin Bengfort is an experienced data scientist and software engineer who focuses on implementing data products that can learn from real-time streaming data. Benjamin is the program director of the Georgetown Data Science Certificate program where he also teaches Machine Learning. He is also ...
Because of COVID-19, many providers are cancelling or postponing in-person programs or providing online participation options.
We are happy to help you find a suitable online alternative.