Making Neural Networks Useful

March 6, 2020

It seems that neural networks are eating the business world for breakfast nowadays. I can hardly get through a day where someone does not strike a conversation about Neural Networks and their inevitability in changing business landscapes.

Now, do not get me wrong, neural nets are incredibly useful tools and they are an excellent choice for solving certain classes of business problems. Having said that it is also important to understand and appreciate that the current crop of neural networks have certain fundamental problems that are preventing their wide adoption across businesses.

They are (in no particular order of importance):

Lack of Uncertainty: Most Neural Networks architecture today are incapable of providing an uncertainty measure for their predictions. Now, as we all understand any measurement is meaningless if we do not give an error bound for the same. This to me is also the core reason why most neural networks today are susceptible to adversarial attacks. This is a problem that traditionally we have tried to solve using Bayesian techniques , but once again most of these are not really scalable. This needs to be solved before large enterprises can start adapting neural networks in a big way.

Explainability: Once again most neural network architectures currently function as black boxes with little insight into how they make their decisions. There has been considerable work in this area over the last few years, but much yet remains to be done. We are no where close in explaining the decision making processes of neural networks in terms of “causality” as we humans understand it.

Large Number of Parameters: The successful neural network architectures of today predominantly in the areas of vision and NLP (Natural Language Processing) have hundreds of millions of parameters. This results in huge compute times for training them, which is now a serious contributor to the global warming phenomenon as well as limits their execution on edge devices.

It is only once we have addressed the above three pivotal issues much better, that neural architectures will truly start impacting our lives and businesses in a more fundamental manner. Till then unfortunately most of them will keep existing as “Jupyter Notebooks” on data scientists laptops.