Algorithmic mythologies
Abstract
Cultural practice in general is a slippery but deeply felt affair, where practitioners are often tasked with communicating the »underside of meaning« (Barthes, 1976). In algorithmic times, focusing on this task when working with machine learning outputs can offer an alternative to the pressure to consider quantifiable and easily machine-readable datasets as complete repositories of human knowledge. The multimodal knowledge held in sound and human experience is difficult to parse with data tools, e.g. machine learning, but the failures and messiness of these outputs add to our toolsets of constructing and reconstructing the worlds we offer in our stories. Through a presentation of some of Sinha’s discoveries navigating the worlds of theatre and machine learning-based audio research-creation, this audio paper offers examples of the possibilities of engaging with algorithmic failure as a deliberate strategy of discovery, to uplift the exploration of cultural possibilities and engage alternative models of community building and connection.
A note on »machine learning checkpoints« (referred to starting at 4m37s):
In machine learning, checkpoints refer to saved snapshots of a model's state at specific points during training. These checkpoints typically include:
- Model weights/parameters
- Optimizer state (for resuming training)
- Training metadata (e.g., epoch number, loss value)
Checkpoints allow you to:
- Resume training from the last saved state in case of interruption.
- Avoid loss of progress during long training runs.
- Evaluate performance at different stages to select the best-performing model.
- Prevent overfitting by early stopping when validation loss stops improving.
Example: Sound Model Training
Suppose you're training a neural network to generate or classify sound (e.g., speech or music). During training:
- The model learns patterns from sound waveforms or spectrograms.
- Every few epochs (e.g., every 5 or 10), the training script saves a checkpoint.
- If a validation loss decreases significantly, that particular checkpoint might be marked as the »best model«.
Disclosure
ChatGPT, 2025. Response to a question about checkpoints in machine learning with sound model example. [Chatbot] OpenAI ChatGPT, 16 May. Available at: https://chat.openai.com/ [Accessed 16 May 2025].
