Environmental Adaptation and Differential Replication in Machine Learning

Abstract

When deployed in the wild, machine learning models are usually confronted with an environment that imposes severe constraints. As this environment evolves, so do these constraints. As a result, the feasible set of solutions for the considered need is prone to change in time. We refer to this problem as that of environmental adaptation. In this paper, we formalize environmental adaptation and discuss how it differs from other problems in the literature. We propose solutions based on differential replication, a technique where the knowledge acquired by the deployed models is reused in specific ways to train more suitable future generations. We discuss different mechanisms to implement differential replications in practice, depending on the considered level of knowledge. Finally, we present seven examples where the problem of environmental adaptation can be solved through differential replication in real-life applications.

Human-Centric AI: The Symbiosis of Human and Artificial Intelligence

“…Finally, Unceta, Nin, and Pujol [ 12 ], in their contribution entitled “Environmental Adaptation and Differential Replication in Machine Learning,” were inspired by theoretical concepts (differential replication) and framework grounded in biological evolution in order to solve what they term the environmental adaptation problem of machine learning models. In this newly introduced machine learning setting, as opposed to transfer learning and domain adaptation, a change in the conditions of a model demands the definition of a new feasible set of solutions because the solution in the source scenario is unfeasible in the target scenario.…”

Differential Replication for Credit Scoring in Regulated Environments

“…These are situations where it is not the data distributions or the problem domain that change, but a model’s environment itself. As defined in [ 14 ], a machine learning model’s environment comprises all the elements that interact with the model throughout its lifespan, including the data and their different sources, the deployment infrastructure, the governance protocol, or the regulatory framework. This environment can be mathematically formalized as a set of constraints on the hypothesis space.…”

“…These changes generally require the definition of a new model in a different hypothesis space. It is these situations that the environmental adaptation is concerned with [ 14 ].…”

“…This problem has been defined as environmental adaptation [ 14 ] and refers to situations where the learning task remains the same, but the environmental conditions change. Contrary to what one may think, changes in the environmental conditions of a model happen with a relatively high frequency and may arise from a change in the technological infrastructure, the business needs, the existing legislation, etc.…”

Desiderata for Explainable AI in Statistical Production Systems of the European Central Bank

“…We now define what is copying a machine learning classifier [41][42][43]. We use the term original dataset to refer to a set of pairs 𝑋 = (𝑥 𝑖 , 𝑡 𝑖 ), 𝑖 = 1, ..., 𝑀, where 𝑥 𝑖 ∈ 𝑅 𝑑 is a set of 𝑑-dimensional data points in the original feature space D and 𝑡 𝑖 ∈ 1, ..., 𝐾 their corresponding labels.…”