Artificial intelligence is the capability of a software program to perform tasks commonly associated with human cognition. Artificial intelligence is more a new reality than a passing trend. Although a current gap exists between investment in artificial intelligence and its commercial application, it is difficult to find an industry that is not researching and developing applications of artificial intelligence to deliver value to customers. For instance, the use of artificial intelligence in the U.S. education sector is forecasted to achieve a compound annual growth rate of 47.5% through 2022 through automation of administrative tasks, addition of “smart” content, deployment of “smart” tutors and personalization of instruction, and implementation of virtual lecturers and learning environments.
Machine learning is at the core of the development of most artificial intelligence systems. Machine learning focuses on learning through optimizing mathematically the relationship between a targeted performance measure and predictors of the performance. From an information theory perspective, every real-world dataset contains predictors that possess both signal (i.e., true values) and noise (i.e., error). The algorithms resulting from mathematical optimization make use of signal as well as noise to obtain the best fit between the targeted performance measure and a set of predictors.
Because algorithms developed from one set of data rarely generalize perfectly to new datasets that have their own unique signal and noise components, machine learning developers cross-validate their algorithms. The main approach in cross-validation is to split the data (one or several times) for estimating the risk of the algorithm: Part of the data (the training sample) is used for training each algorithm, and the remaining part (the validation sample) is used for estimating the risk associated with using the algorithm.
Even with cross-validation, practical applications of algorithms can lack generalizability when the optimization technique focuses on peculiar or unintended noise embedded in the features of a target variable. For example, a machine learning algorithm was estimated for differentiating images of military vehicles from civilian vehicles in an attempt to reduce military strikes on non-combatant vehicles. The algorithm remained accurate over cross-validation. However, the algorithm failed miserably in actual combat conditions because it was found to be actually focusing on noise in the training and cross-validation data. Photos of many military vehicles displayed shadows of the vehicles while civilian vehicle photos showed few vehicle shadows. When there were no consistent patterns of shadows in combat, the smart algorithm turned dumb.
In this presentation, we discuss the conditions that cause algorithms to fail “in the wild” and how decisions based on complex algorithms can be interpreted and explained (such as required in Recital 71 of the European Union General Data Protection Regulation that “[the data subject should have] the right…to obtain an explanation of the decision reached”). Such explanations are necessary if artificial intelligence can be trusted, even when it is accurate.
Topics: Learning Analytics: Research and Practice , Topics: Digital Technologies in Disciplinary Contexts