How we built a Pragmatic Machine Learning Pipeline to Identify Fake Flysheets as part of a Library Workflow

1 Preface

In 1979, two Swedish bands released albums. One of these albums, Voulez-Vous, was recorded by ABBA in Polar Studios, at its time, one of the most advanced — and expensive — recording studios in the world. The other album, We're Only in it for the Drugs, by the Swedish Punk band Ebba Grön was recorded on a mobile 8-channel mixer in a closed industrial office. Both these albums are great, but the resources available to create them were vastly different.

There is a prevalent perception that machine learning is only for those with deep pockets. For example, in 2022, Google created a new Language Model, PaLM, but the cost of training such a model is estimated to be between $9M to $23M. As a result, discussions of machine learning often focus on large tech companies because they are deemed to be the only ones with the finances to develop and use this technology.

Machine learning, and in particular a branch of machine learning called deep learning, has dramatically impacted a range of domains over the past ten years. There has been a growing interest in using machine learning in gallery, library, archive and museum (GLAM) institutions. A barrier to further adoption of machine learning in the GLAM sector is the perception that it requires extensive technical skill, data, computing power and other resources.

In this book, we want to show that just as Ebba Grön was able to record a seminal album without the resources available to ABBA, GLAM institutions can use machine learning for practical work without the resources of Google.