Page 145 - Proceedings of the 2018 ITU Kaleidoscope

Machine learning for a 5G future

compound the problem. But trying to solve this issue agencies to invest in more research and development on this
through heavy-handed regulation and strict ex-ante controls front.
would present its own set of challenges. Such interventions
may come at the cost of stifling efficiency and innovation. Finally, we must remember that the datasets being used for
This also presumes a certain level of state capacity to training machine learning algorithms are created in the real-
effectuate the regulation, which is often not available in world, i.e. outside the AI ecosystem. Therefore, while
reality. How then can we strike a balance between these building reactive use-case based solutions (NITI Aayog,
positions to make sure that AI research evolves in a socially 2018[26]) may solve some of our immediate needs, the
and ethically responsible direction? We propose a three step larger agenda must be to correct the training dataset itself.
approach towards this goal. To take an example, the outcomes of natural language
processing can be made more inclusive if the persons
The first step would be to embed the concept of “fairness generating the underlying text (writers, researchers,
by design” in AI frameworks (Abbasi et al, 2018[28]). This policymakers, journalists, publishers and other creators of
draws from the concept of “privacy by design” that has digital content) work towards the feminization (using words
evolved in the context of data protection debates like she and her) and neutralization (chairperson instead of
(Cavoukian, 2011[29]). Fairness by design should compel chairman) of the language that they use (Sczesny et al,
developers to ensure that the very conception and design of 2016[33]). Here again, there is a role for the State to use
AI systems is done in a manner that prioritizes fairness. awareness, education and, if required, other policy tools to
Abbasi et al, 2018[28] propose that the components of such promote the use of gender fair language. Similar solutions
a framework would include: need to considered for other fields of AI research,
(i) creating cross-disciplinary teams of data scientists accompanied by the identification of the persons and
and social scientists; processes needed to effectuate the desired changes.
(ii) identifying and addressing the biases brought in by
human annotators; 5. CONCLUSION
(iii) building fairness measures into the assessment
metrics of the program; From its very inception, the field of AI has largely remained
(iv) ensuring that there is a critical mass of training the domain of men. This paper illustrates how the gender of
samples so as to meet fairness measures; and its founders and subsequent researchers has played a role in
(v) adopting debiasing techniques. determining the course of AI research. While efforts are
now being made to fill this gap, including by promoting
A fair amount of research has been done on building more women in STEM, the gender problem of AI is not just
solutions for gender biases in natural language processing. about the representation of women. It is also about
For instance, Bolukbasi et al, 2016[20] use debiased word understanding whose agendas are being pursued in AI
embeddings for removing negative gender associations research and what is the process through which that
from word embeddings generated from a dataset. Another knowledge is being created.
strategy is to use gender swap techniques to remove any
correlation between gender and the classification decision Research has shown that AI’s reliance on real-world data,
made by an algorithm (Park et al, 2018[30]). A variation to which is fraught with gender stereotypes and biases, can
this would be to conduct “stress tests” where certain parts result in solutions that end up reinforcing or even
of the data (such as the gender of some candidates in a exacerbating existing biases. While fairness and non-
selection process) can be randomly altered to check discrimination are well recognized principles in the human
whether the randomization has an effect on the final rights discourse, these principles often fail to translate into
outcome that is generated, i.e. the number of women being practice, often on account of the conscious and unconscious
shortlisted (Economist, 2018)[31]. biases. The challenge therefore is to find ways to bundle the
technological progress of AI with the objectives of pursuing
While encouraging further research of this nature, a lot greater fairness in society -- for machines to eliminate
more needs to be done in terms of mainstreaming these rather than reinforce human biases.
solutions and making them readily available to smaller
developers. Google’s “What-If” tool offers a useful We propose a three step process towards this end. First, we
example. It is an open source tool that allows users to need to develop a set of publicly developed AI ethics that
analyze machine learning models against different embed the concept of “fairness by design”. To travel the
parameters of fairness. For instance, the data can be sorted distance from formulating ethical principles to their actual
to make it “group unaware” or to ensure “demographic implementation is another challenge. We find that
parity” (Weinberger, 2018[32]). Given the many positive “fairness” as a concept is prone to diverse interpretations,
externalities to be gained from the creation and opening up which can result in its under-production in the system.
of such fairness enhancing tools, the second step of the re-
envisioning AI project would be for governments and other The second step would therefore be to invest in research
and development in formulating technological tools to

– 129 –