Posted by Anirudh Bhardwaj | Last Updated: 28-Nov-18
It’s no surprise that Google has made a tremendous progress in the field of Machine Learning and Artificial Intelligence over the last couple of years. With so many ultra-advanced tools and technologies under its belt, Google has long been at the forefront when it comes to pioneering technologies like AI and Machine Learning. Lately, the company has made a headway with Tensorflow, a computational framework for building Machine Learning models. Developed by the Google Brain Team, Tensorflow offers a flexible architecture to deploy computation to multiple CPUs and GPUs.
As of today, Google is apparently working on eliminating the shortcomings with the existing Machine Learning systems. A paper published by Google Brain and Deep Mind explicitly states that their research team is currently addressing the shortcomings of the field and potential loopholes persisting in the Deep Learning systems. Furthermore, the team hopes to develop new techniques that would lead Machine Learning toward a human reasoning like methodology called Artificial General Intelligence. The researchers claim that the present day Deep Learning approaches lag far behind when it comes to human cognitive skills.
What Are The Shortcomings
The paper states that the existing Machine Learning systems have been designed to generalise the human experience and mostly rely on cheap data and computing resources. The present day Machine Learning systems are adroit enough to identify an object or person from hundreds of millions of images. But they can’t rate the person based on their looks and qualities since it is far beyond the illustration of the general human behavior. There are a number of reasons as to why such intelligence couldn’t be achieved after all this time. The most challenging fact is that what is cute for one human might not be the same for another. Furthermore, the lack of authentic data also adds to the complexities. To this, the paper proposes a Graphs Network to better represent human cognition. This can be done by mapping the objects and entities without mapping their relationships with each other.
What Google Proposes
The idea that Google proposes is about the use of Graph Networks that are larger than all machine learning approaches. Besides, Graphs give out the ability to generalize the structure that neural networks can’t have. Another benefit of implementing Graphs is that they are sample efficient which means they require less amount of raw data as compared to artificial neural networks.
You may also like How Machine Learning Can Help Counter Spear Phishing Attacks.
More Challenges To Address
Although the solution proposed by Google through Graph networks aims at addressing the shortcomings of the existing Machine Learning systems, it turns out that the solution itself is having a number of shortcomings. First and foremost, Deep Learning subsumes a lot of unstructured data and the data may not be associated with any entity at all. So it’s going to be quite challenging to efficiently extract discrete entities out of the sensory data. In addition to that, the graphs can be futile in some cases. For example, graphs alone, can’t express notions like recursion and control flow and it will require additional assumptions. It might also require structural forms and entities like stacks, queues, registers and memory I/O controllers.
But for what its worth, if the research teams from Google are able to find a way to address these potential challenges, it might be the next big achievement in the field of Machine Learning.