Description
Date depot: 1 janvier 1900
Titre: Deep symbolic learning of multiple temporal granularities for musical orchestration.
Directeur de thèse:
Geoffroy PEETERS (LTCI (EDMH))
Directeur de thèse:
Florent JACQUEMARD (CEDRIC)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
Orchestration is the subtle art of writing musical pieces for orchestras. Hence, it requires combining instrumental properties in order to reach ideas of timbre coming from the blending and combination of individual spectral properties. The paramount challenge in this field is that there has never been any systematic research in music analysis and theory on orchestration, a gap this project would start to address. Orchestration practice involves solving complex problems implicating vast knowledge of the properties of musical instruments and how they operate. In contemporary classical music where timbre has become a primary structuring force, computer-based tools are being developed in order to help composers solve problems of finding instrumental combinations that could fulfill the desired musical goals. The topic of orchestration encompasses open issues from auditory perception, music analysis and composition, signal processing, computer science and combinatorics [2]. Integration of systematic music analyses based on perceptual principles and experiments into computer-aided orchestration platforms would revolutionize creative and pedagogical practice. However, the problem of orchestration is a highly combinatorial NP-problem. Indeed, if we solely consider the range of notes, dynamics, chords and playing modes provided by a single musical instrument, we can foresee the infinite number of combinations provided by an orchestra [3].
Recently, the field of deep learning [4] has witnessed a tremendeous interest amongst the machine learning community. Its goal is to train connexionist architectures for artificial intelligence similar to those constructed in the well-known neural networks. However, the typical feed-forward perceptrons are usually bound to small depth (number of layers) because of the gradient diffusion problem that lead to inefficient learning in deeper networks. By relying on greedy layer-wise training algorithms [5], it is possible to train each layer of the networks independently and in an unsupervised manner. This leads to various types of architectures such as deep belief networks that can display an extensive number of layers [6]. This leads to deep neural architectures of representations similar to those found in the brain for perceptual task such as visual or auditory processing. This approach appears as a valid candidate to learn models of AI that could bypass the problem of the curse of dimensionnality, but also to learn automatically higher-level abstractions for both symbolic and temporal datas. Even though these principles are undergoing intensive research in signal processing communities, its implications on symbolic representations of music are yet to be explored, a research direction that we call deep symbolic learning. A first step towards such a system would be to study the learning behavior of connexionnist architectures with regards to orchestral composition. These researches could be enhanced by current knowledge in perceptual effects of orchestration led by a partner laboratory in McGill university. Furthermore, orchestration lies at the crossroads of signal (sound) and symbolism (musical score) but also thrives between micro-temporal evolution and macro-temporal articulations.
Hence, this project is aimed to explore neuroscientifically-informed machine intelligence systems for symbolic data with the non-determinism related to the exploration of a highly combinatorial space (because of the quasi-infinite possibilities offered by the orchestral colors) but also centered around temporal problematics and their inherent variable granularities [1]. The goal of this PhD project is to provide an approach that could help in translating the intent of a composer in the process of orchestration. Hence, the main idea is to first learn the inherent structures that co-exist between different musical elements (relationships inside the symbolic knowledge of musical scores, between different signals but also between the signal and the score). Then, based on the learned connexionnist architectures of representation, the system could propose some re-orchestration and original improvisations. That way, this system could provide an interaction between the spectrum of sounds and their evolution in time.
Doctorant.e: Crestel Leopold