Essay · 11 February 2020 · 4 min read
Contextual Explainability
Artificial Intelligence (AI) has been steadily moving to the forefront of technological innovation, firmly embedding itself as the main differentiator for services and products. AI methodologies such as neural networks and deep learning showing increasingly accurate performances which in some cases exceed that of humans. However, recently, and despite this seemingly unchallenged progression, there have been growing calls to condition AI adoption with AI explainability. Advocating that the inability to explain should in turn prohibit, or at the very least, limit, the use of AI. These calls have spurred valiant efforts by data scientists across the globe to make the inner workings of AI algorithms explainable. All wanting to avoid the dreaded so-called black box.
Firstly (and potentially, most importantly), what does explainability actually mean? “Explainability” or “explain-ability” can be inferred from Merriam-Webster as the quality or state of being able to make something *plain *or understandable. Thus the necessity to be able to understand appears obvious. This leads us to contemplate further - *what *and by *whom is understanding being sought. *
Perspectives
In order to take the first step in addressing these questions there is a need to view AI from three separate perspectives; (1) AI as just a technological engine; (2) AI as a different form of engine; and (3) AI as an engine potentially impacting social norms.
The third perspective is hinged on the nature of AI to uncover patterns, and as such it is potentially the most challenging and emotionally charged of the three. These patterns, whether due to bias in historical data or indeed truly being a statistical fact, may touch on an exposed social nerve(s) when showing outcomes that appear to favour or disfavour certain individuals or groups of individuals.
However this is not a question of *AI *explainabilitiy as much as it is a consideration of social acceptability. For example, in a situation where it is possible to coherently explain in a plainly understood manner that a particular gender should be favoured or disfavoured in order to achieve a more efficient outcome, would it make it any more or less socially acceptable?
The second perspective is focused on how AI solves problems. As with many common scientific disciplines, AI may solve complex problems using nonlinear solutions. That is, a solution where the change in the output is not proportional to the change of the input. In other words, it may not be possible to directly understand *how *the output is related to the input but only that they are related. Nonlinear problems are common in many scientific disciplines as most real-world systems are inherently nonlinear in nature.
The challenge with AI, somewhat ironic as it is concurrently the value proposition, arises from the fact that AI is powered by big data. In the past the number of inputs used would be in the range of tens to hundreds, now AI algorithms can easily consume hundreds of thousands, and more, of inputs. Thus the difficulty in the ability to understand the input-output relatedness only grows with the volume of data consumed.
Understanding how models work is important in advancing our understanding of AI, potential implications as well as providing a better understanding to the systems and behaviour being modelled. Reliance on black-boxes may be limited due to the dynamic nature of a business.
Nonetheless, it is concurrently important to establish the trade-off between understanding input-output relatedness and business applicability. In other words, to what extent is it required to understand the exact relatedness between input and output for an application to have business viability.
Contextualisation
The aforementioned third and second perspectives are of a social and technical lens. Addressed, somewhat, through social acceptance and scientific feasibility, whereas the first perspective anchors on business considerations. Viewing AI as ‘just’ a technological engine to be embedded within a business context. We are posed with the deliberation of how and should AI explainability differ from any explainability currently required in the creation of a business product or service. Is AI held to a higher standard and thus subjected to a more onerous explainability requirement throughout the business process?
Explainability is a spectrum and as such it is important to lay a framework that identifies what explanations are required by which individuals, contextually aligned to their role and responsibilities. Otherwise there is genuine risk that everyone would require either too much or too little information needed for decision making.
Explainability at the business layer needs to be anchored on a level that is independent from the AI models used. Thus, any explainability methodology, framework or tool that is dependent on the model would fail on wide spread adoption and, quite frankly, relevance.
In summary, it is important that AI explainability is addressed across all levels in a manner that is contextually relevant. Furthermore, it is equally critical that we plainly understand what is it that we are trying to explain.
References;
Follow on LinkedIn Subscribe on LinkedIn