In today’s analytics-oriented environment, data is being wielded as an indispensable instrument for efficient decision making backed by concrete insights. Especially intriguing is the way internet firms are analysing voluminous consumer-centric data that is being generated at an estimated rate of 2.5 quintillion bytes per day. Businesses have moved on from the traditional approach centered around intuition and/or guesswork, to a data-driven decision-making mechanism backed by quantitative rigour. But that is just the tip of the iceberg.
The scope of Data Science, which is the larger umbrella term in this domain has expanded beyond prescriptive and predictive modelling for business purposes. It has rather ventured into areas where the possibility of ‘humanising’ machines is being explored through machine learning and its more sophisticated form – deep learning. We commonly call this developing area of technology, Artificial Intelligence (AI).
Of course, much has been achieved here, especially in the last few years with applications like voice-recognition, image-processing, semi-self-driving vehicles, etc. But without undermining the technological revolution it has brought about; it will be conservatively prudent to not over sensationalise AI.
So, let’s take a step backward. The real struggle when it comes to building machine learning models has been with respect to the interpretability of the results. In fact, in most cases, the higher the predictive accuracy of a model, the opaquer seems to get in terms of knowing why the model did what it did. So, while the accuracy of the ‘what’ has been achieved at near-perfect levels, the transparency of the ‘why’ is not even close to where we would want it to be for it to become reliably and realistically autonomous.
At this juncture, I would like to simplify things through a real-life business scenario I had faced. One of the e-commerce clients at the analytics-consultancy firm I worked at in the past, had its profits and operational efficiency plagued by a huge number of online product returns. With the aim of combatting this issue, my team and I tried to build a predictive model, that could estimate the likelihood of a purchase ending up in a return. This would help them in planning for such transactions, identifying the problematic customers, locational characteristics, etc.
After setting up the business case for this data science project, by identifying various factors and features relating to customers, their location, order size and value, product category, payment method and others we were able to develop an array of predictive statistical and machine learning models for the client. And the pattern was also very much visible, i.e., for models with comparatively fewer levels of predictive accuracy, there was enhanced transparency, and vice-versa (for example, the logit model vs neural networks).
Finally, in the model adoption stage, we were not shocked when the client picked the one with the slightly lower accuracy, for it was far more crucial for them to find the root cause of online returns and fight it at the source, than to just have information on what the binary outcome was going to be.
One of the more complicated machine learning models, neural network, is designed to mimic the human brain. Experts in medical sciences have not still been able to dig deep enough so as to get to the roots of how our brain functions, so it is very difficult to replicate what we don’t know 100% about. Therefore, it remains a very real possibility that AI could be too smart for us.
So, it’s important to note that the motive here is not to downplay the efficacy of an algorithm that can predict an outcome with, say, 95% accuracy but cannot elucidate the same outcome. Indeed, there are numerous applications of these techniques, and leveraging them has only resulted in reducing the amount of uncertainty in decision-making and mitigating potential losses in a host of industries. The intention here is to rather create an understanding around the repercussions of something that is not entirely explicable, and thus uncontrollable.
In furthering that understanding, raising the stakes might help a bit. AI has made its way into areas such as health-care and the automobile industry which are prime examples of those where human lives are involved. Leaving decision making up to an algorithm that performs countless calculations without completely letting in on the logical base behind it, in such a context is almost suicidal. The ‘black box’ of high-complexity machine learning models evade human understanding at a level that it puts a massive roadblock in deploying autonomous technologies at a full-fledged scale.
Imagine a self-driving car taking the wrong turn resulting in a catastrophic outcome due to the hidden workings and iterations of an algorithm that we might never be able to decode. But stop right there, as thinking beyond this about expanding the reach of AI without coming up with a solid understanding is only detrimental for the world at large.
Man is born to be free. But this freedom that we have earnt for ourselves has come with a caveat, that is the shear ability to reason. And when we as humans cease to reason, we become the very thing we have strived against not to be- slaves. In light of everything that has been discussed above, it will be safe to say that machines; are even though very close to be humanised, should be far from operating with full freedom.
Data science has been attracting a lot of attention primarily owed to its potential to take the element of guess work out of the process of making decisions with the ultimate target of minimising obscurity. The algorithms presenting the so-called black-box are hence against the very fundamentals at which this innovative field has been built on. It is therefore the duty of the data science and machine learning communities to take responsible action and lead the way for humanity into surely a more convenient tomorrow, but a safer one all the more.
On this note let’s replace the question, ‘Are you ready for AI?’ with ‘Is AI ready for you?’.