AI and Machine Learning for Oxford’s Gardens, Libraries and Museums Division (GLAM)
The test beds intend to use an incremental learning approach building on Machine Learning Common-Off-The-Shelf Tools (COTS), using tried and tested models which can then be scaled vertically or rolled out into production systems for the departments in GLAM.
Theme 1: Coins
With an early pilot in 2019, the Ashmolean Museum’s rich Roman Provisional Coins collection proved to be a successful dataset for bringing together key concepts in ML methodology. The fundamental question was whether a ML tool would be able to differentiate between ‘heads and tails’ of a coin. The pilot could not only identify the above but also went to the extent of similarity search, material identification and image enhancements. This test bed is developing a working prototype with a simple web front-end for the service with the learning extended to other themes.
Theme 2: Fossils
Tomographic techniques are used to investigate the form, function and evolution of fossils. A major focus of this work involves applying destructive serial grinding to exceptionally preserved fossils that cannot be studied using non-destructive imaging methods. Specimens are digitally photographed at regular increments as they are ground away, generating datasets consisting of hundreds of images per fossil. Currently these must be manually aligned, slice-by-slice, before they can be reconstructed as a high-quality, three-dimensional virtual fossils. Automating this work through the use of ML algorithms drastically reduces the time need to process each sample, thereby enabling the material to be studied much more quickly and efficiently.
Theme 3: Gemstones
The hundreds of different minerals that can form precious stones, or gems, can be deceptively similar. However, they can be diagnosed via a series of physical and optical properties that are well described. The Oxford Museum of Natural History has a collection of over 30,000 mineral specimens including a fine collection of cut gemstones. A subset of these, covering the most commonly found minerals, is being used as a machine learning training dataset by providing known physical and optical parameters that can be easily determined or tested such as colour, transparency, crystal form, refractive index, specific gravity, and reaction to UV. With only one or two of these parameters it is possible to narrow down the range of possible identity from hundreds to tens or fewer, and with the addition of more data it is possible to gain higher and higher degrees of precision. Simulated untrained data inputs are then used to test the efficacy of this approach.
Theme 4: Gardens
BRAHMS is an Oxford-led database management software system used to manage botanic gardens, seed banks, herbaria and other museum based natural history collections, images and literature. Built into BRAHMS is a diverse range of practical tools for collection management, all encouraging and facilitating publication of data otherwise locked up in collection archives. BRAHMS is now deployed globally with projects ranging in size from the taxonomic revisions of small genera to managing some of the world's largest herbaria, botanic gardens and seed banks. This theme is exploring the potential application of BRAHMS as a solution at South East Asia's second largest Herbarium, in Singapore, and looking at blueprints for delivering a refactored cloud optimised solution.
PIs: Haas Ezzet, Anjanesh Babu