While not traditionally a HPC workload, we increasingly see domain scientists across the board to incorporate machine learning (ML) into their analysis, the same is true for users from industry.
Even if the deployed machine learning algorithms are not deployed on HPC environments, increasingly they are being trained on supercomputers. As such we as HPC practitioners have a particular responsibility to be aware of potential social impact that results of advancing these technologies as well as potential misuse.
A brief overview of standing challenges with ML & AI in relation to HPC. Many of the ethical challenges originate not in HPC but change their dynamic with HPC available as a tool.
Bias Amplification & Fairness
Data collected in many situations picks up bias already embedded in our society. How can we prevent undesired bias from getting amplified when using ML. How to quantify seemingly objective algorithms and indicate if they really are not?
How to avoid discrimination by gender, age or other attributes? How to measure fairness? And even if removing sensitive variables, how to avoid proxy variables still introduce bias. Disentanglement? But how when these factors are correlated.
ML allows automating logistically challenging data association for example from CC tv and other data source. This can be used to circumvent privacy.
Even voluntarly shared data often might hold information which can be reconstructued with machine learning to which when indiviudally asked about it any people might object.
Machine learning increasingly allows to automate demanding cognitive tasks traditionally performed by humans. How can we limit the impact this has on society?
When workers are displaced, be sure to participate in transition strategies to limit and mitigate impact.
ML can be used to automatically generate highly convicing fake data. Examples for this includes images, video and sound. While this has many interesting applications, it is also
Machine learning approaches can be very, sometimes surprisingly, effective. Unfortunately, many methods offer limited introspection mechanisms to understand why some neural networks provides one answer over another. This should impose limits on where it is adequate and risk-free to apply machine learning.
Luckily, there is active research with many promising approaches to change this. Unfortunately, often these approaches are more complicated of very expensive, which can encourage users to opt for an ML approach without good understanding of failure modes.
Machine learning requires substantial computational resources for training besides access to skilled practitioners to leverage ML. As such organizations with access to those resources can have an advantage.
At the same time, various efforts exist to make machine learning accessible. And in comparison to the past, even computational resources have become more accessible then they have ever been before. although not always at a cost-competitive level
It seems fair to assume, that HPC will often be invovled when it comes to the most capable AI solutions.
Related to bias, fairness and explainable AI. When deployed in these context.
Similarly, have to see how to consolidate.