The development of generative machine learning (ML) models in creative practices has gained significant interest among artists, practitioners, and performers. Recent improvements in the usability and availability of pre-trained models have enabled the use of these techniques in artistic domains. However, the introduction of such techniques has also revealed multiple limitations that escape current evaluation methods used by scientists. Notably, most models are still unable to generate content that lay outside of the domain defined by the training dataset. In this paper titled "Challenges in Creative Generative Models for Music: A Divergence Maximization Perspective," Axel Chemla-Romeu-Santos and Philippe Esling propose an alternative prospective framework starting from a new general formulation of ML objectives. The authors derive this framework to delineate possible implications and solutions that already exist in the ML literature, notably for the audio and musical domain. The proposed framework aims to address the lack of creativity in existing generative models by maximizing divergence between generated outputs and training data while taking into account conditioning variables that contextualize generation. The authors suggest using a CC-inspired evaluation method to promote creative output from generative models. The paper discusses existing relations between generative models and computational creativity, highlighting how their proposed framework could help overcome current limitations faced by generative ML models. By providing a more comprehensive evaluation method for generative models' creative output, this framework can facilitate better integration of these techniques into artistic domains while promoting more diverse and innovative outputs beyond those defined by training datasets. Overall, this paper presents an important contribution towards advancing creative applications of generative ML models while addressing current limitations faced by these techniques in artistic domains.
Generative machine learning models are computer programs that can create new things like music or art. People are using these models more and more to make creative things. Scientists have found some problems with how they test these models to see if they work well. One problem is that the models can only make things similar to what they were trained on, not completely new things. Some people have come up with a new way to train these models so they can be more creative and make different kinds of things. This could help artists use these models to make even cooler stuff than before!
Definitions
- Generative machine learning models: computer programs that can create new content based on patterns learned from existing data
- Usability: how easy something is to use
- Pre-trained models: generative machine learning models that have already been taught how to create certain types of content
- Evaluation methods: ways scientists test whether a model is working well or not
- Divergence maximization perspective: a new way of training generative machine learning models so they can be more creative and generate diverse outputs
- Conditioning variables: factors that influence the output of a generative model
- CC-inspired evaluation method: an evaluation method suggested by the authors of the paper, which promotes creative output from generative models beyond what was in the training dataset
Exploring the Creative Potential of Generative Machine Learning Models
The development of generative machine learning (ML) models in creative practices has gained significant interest among artists, practitioners, and performers. Recent improvements in the usability and availability of pre-trained models have enabled the use of these techniques in artistic domains. However, the introduction of such techniques has also revealed multiple limitations that escape current evaluation methods used by scientists. Notably, most models are still unable to generate content that lay outside of the domain defined by the training dataset.
In this paper titled "Challenges in Creative Generative Models for Music: A Divergence Maximization Perspective," Axel Chemla-Romeu-Santos and Philippe Esling propose an alternative prospective framework starting from a new general formulation of ML objectives. The authors derive this framework to delineate possible implications and solutions that already exist in the ML literature, notably for the audio and musical domain.
A New Framework for Evaluating Generative ML Outputs
The proposed framework aims to address the lack of creativity in existing generative models by maximizing divergence between generated outputs and training data while taking into account conditioning variables that contextualize generation. The authors suggest using a CC-inspired evaluation method to promote creative output from generative models. The paper discusses existing relations between generative models and computational creativity, highlighting how their proposed framework could help overcome current limitations faced by generative ML models.
By providing a more comprehensive evaluation method for generative models' creative output, this framework can facilitate better integration of these techniques into artistic domains while promoting more diverse and innovative outputs beyond those defined by training datasets. Overall, this paper presents an important contribution towards advancing creative applications of generative ML models while addressing current limitations faced by these techniques in artistic domains.
Conclusion
This research paper provides an insightful perspective on how to evaluate creative output from generative machine learning (ML) models with a focus on audio/music applications specifically. By proposing a new general formulation based on divergence maximization between generated outputs and training data as well as taking into account conditioning variables that contextualize generation, Chemla-Romeu-Santos & Esling’s work offers valuable insight into overcoming current limitations faced by these types of technologies when applied within artistic contexts or other creative fields where innovation is key factor for success .
The development of generative Machine Learning (ML) models in creative practices, enabled by the recent improvements in usability and availability of pre-trained models, is raising more and more interest among artists, practitioners and performers. Yet, the introduction of such techniques in artistic domains also revealed multiple limitations that escape current evaluation methods used by scientists. Notably, most models are still unable to generate content that lay outside of the domain defined by the training dataset. In this paper, we propose an alternative prospective framework, starting from a new general formulation of ML objectives, that we derive to delineate possible implications and solutions that already exist in the ML literature (notably for the audio and musical domain). We also discuss existing relations between generative models and computational creativity and how our framework could help address the lack of creativity in existing models.
