Harry Whalley - University of Edinburgh
Before Alpha-Go: Interpretation though improvisation and LISP
Abstract: Like Chess, Go is a non-chance combinatorial game - a strategy game with deceptively simple rules which go back over 3000 years. Only this year have human masters of the game been challenged and beaten by machine learning.
This poster will outline the process undertaken to write Go / Koan, a composition which uses a ½ point win Game between a mid-level Dan Go player and Zen19, a Go computer, which at the time (2013) was one of the strongest of its kind. Since then, Google's Alpha-GO has radically changed this landscape with wide reaching implications.
It will be in 4 parts:- An explanation the concept of gameplay in Go, and why it has been such an enormous challenge in the AI community to challenge humans within this arena.
- Mapping moves to musical structures – choice and code. A brief tour of the LISP code and how it produced musical material from which to work.
- An extract from a performance of this piece.
- What questions does the new paradigm in machine learning represent? Is Alpha-Go creative?
Andrew McLeod - University of Edinburgh
Musical Meter Detection Using Context-Free Grammars
Abstract: Meter identification is the organisation of the beats of a given musical performance into a metrical structure. The metrical structure is aligned in phase with the underlying musical performance so that the root of the tree corresponds to a single bar. We show that using a probabilistic context-free grammar (PCFG) to model the rhythmic structure of a musical piece can aid in musical meter detection. We also show that using a lexicalized PCFG (LPCFG) improves performance further, as it can model the rhythmic dependencies found in music.
James Owers - University of Edinburgh
Deep Music
Abstract: [MScR Thesis] We investigate the applicability of state of the art Convolutional Neural Network models to instrument detection, transcription, and source separation tasks. There have recently been some major breakthroughs applying Recurrent Neural Networks to speech and music data in the literature. We aim to leverage invariance in harmonic patterns introduced by preprocessing the signal with a Constant Q Transform. These 'images' have an additive property when signals are mixed unlike object detection in images so present a new set of challenges.