logo

Bespoke Machine Learning Model 📂Machine Learning

Bespoke Machine Learning Model

Definition

A custom machine learning model designed and trained from scratch for a specific problem/data/domain is called a bespoke model.

Explanation

Bespoke originally refers to a tailored suit. It does not mean clothing mass-produced in fixed sizes like a ready-made product, but rather clothing crafted from scratch to fit a single person’s body. In machine learning, a bespoke model likewise refers to a custom model designed and trained from scratch for a specific problem/data/domain. The term task-specific model is also used.

A bespoke model is a concept contrasted with the universal model. A universal model is one that is trained on vast amounts of data and can be used broadly within a single domain. For example, suppose our goal is to develop a neural network that approximates the energy of materials. Here, a model trained only on data from specific environments such as exsolved nanoparticles, ternary metal oxides, quantum dots, or amorphous silicon is called a bespoke model. Another example of a bespoke model is an artificial neural network trained on the MNIST dataset. This neural network produces no meaningful output for arbitrary photographs. Because it was designed and trained solely to classify handwritten digits from $0$ to $9$, whether a cat photo or a landscape photo comes in, it will merely force it into one of the ten digits. In other words, this model is a typical bespoke model tailored to only one narrow problem: handwritten digit recognition.

On the other hand, SevenNet-Omni can be called a universal model because it was trained on a broad dataset encompassing inorganic crystals, catalysts, MOFs (metal-organic frameworks), molecules, and so on.

See Also