Human aspects of machine learning
MetadataShow full item record
With the widespread use of Machine Learning (ML) algorithms in everyday life, it is important to study the human aspects of these algorithms. ML algorithms are increasingly used in applications that influence our day-to-day life, both in the consumer and enterprise market. In the consumer market, machine learning is becoming the core part of most applications that we use in our life (e.g., ridesharing, social networks, etc.). Consumers expect to adopt these applications easily (human-usable) without sacrificing their privacy or security. In the enterprise market, ML-based applications are impacting consumers, for example, financial institutes use machine learning to assess the credit score of a loan application or evaluate a job application. Consumers demand visibility to how these decisions are made and expect a fair process. All these ML processes and applications, directly or indirectly, interact with humans and thus it is crucial to study their effect. Such a study include, but is not limited to, studying and analyzing (i) the human usability of an ML system, and (ii) the influence of outcomes generated by such a system on humans. These studies can result in: (1) understanding of the limitations of users when adopting an ML system, and therefore improving its usability. (2) creating fair machine-generated outcomes, where fairness is defined relative to the context, and (3) understanding the interactions between humans and the ML systems and how those interactions (e.g., human's bias) can affect the overall outcome of the system. In this thesis, we study ML paradigms from the viewpoint of human usability and fairness. For human usability, we focus on two fundamental problems in password authentication: (a) How can one generate humanly usable passwords that are secure? (b) Given a limited memorization resource, what is the highest security that a humanly usable password strategy could achieve? And can we construct such a password strategy? To answer question (a), we introduced the first usability study of humanly computable password strategies. To answer question (b), we showed that there exist humanly usable password strategies that are hard to hack; in fact, we showed that any adversary needs almost the information-theoretic number of samples to hack these strategies. For fairness, we studied representational bias in unsupervised learning settings such as the dimensionality reduction technique of principal component analysis and spectral clustering.