Machine Learning: Which Way to Go?

Source: Deep Learning on Medium

Machine Learning: Which Way to Go?

While the end of another decade is approaching, it’s quite interesting to witness how the field of machine learning (ML), especially deep learning (DL), has been evolving over the last years. Furthermore, while reading Neural Networks and Deep Learning, one of the observations caught my attention:

While many of us didn’t witness the structuring of many fields such as mathematics, medicine, we are quite lucky to witness a similar phenomenon in ML, and more broadly artificial intelligence (AI). While it’s is still possible to master several fields of ML, many are shifting towards specializations such as reinforcement learning, constitutional networks. Beyond this, there are currently two main trends that could be taken by anyone passionate about the field.

Historical Overview

Since the first article written by A.M. Turing in 1950 about Computing Machinery and Intelligence, ML has been gaining more and more attention. At that time, many tasks seemed to be impossible to be achieved by machines. With the progress achieved in computers allowing to store more data and process more rapidly, several breakthroughs took place leading to a great state today. Nowadays, ML has shown its potential in many tasks like classification, speech recognition, natural language processing, movies recommendations and others achieving sometimes a better performance compared to humans.

ML Buzz

Despite all these achievements, the field seems to be overvalued, especially by external observers. By getting inside the current work and challenges, the way is still quite long before the field reaches its maturity stage. From an applied research perspective, it is easy to witness cases where ML supported humans in an efficient and effective manner. In some other cases, where risk is quite high like medicine, there is still a huge reluctance from doctors as well as a significant room for improvement. From a fundamental research perspective, there are still a lot of questions related to parameters tuning heuristics as well as more mysterious questions linked to intelligence.

Imitating the Human Brain

This is probably the most exciting aspect about ML. Nowadays, there is a lot of discussions about shifting from System 1 to System 2 in the field of DL. Brain’s system 1 is fast and emotional while brain’s system 2 is slower and logical. In classification tasks for instance, a machine learns in a similar way to our brain’s system 1. All the breakthroughs in ML were inspired from system 1 of human brain. Still, many issues faced could be solved if system 2 is also incorporated in specific ways. The mission seems very complicated given the lack of information about many notions such as intelligence and consciousness but deserves a try and could lead to new discoveries on the research path.

Human Level Performance

Among humans, it is very common to see that the student surpasses his professor. It is mainly due to our ability to generalize beyond what we learn or study. Our brain, with its unique structure, allows us to be creative and innovative. When it comes to machines, it is really exciting to witness their ability to surpass humans in some cases such as classification of images for instance. In Japan, Todai Robot’s performance was better than 80% of students on the entrance exam for the University of Tokyo. This is both fascinating due to machines’ abilities to surpass human level performance, which can be very useful in many daily tasks, and disappointing because, in some cases, humans started to reduce the focus on the “why” of things highlighting many flaws in our education systems worldwide.

Prospective Ways

As a passionate about the field and through some exchanges of insights with many others, I guess two prospective ways could be the focus. The first one is the applied research. With the current achievements, many problems worldwide can be tackled using machine learning. With the available models and with suitable improvements, many projects can be tackled efficiently and effectively, especially the ones that can contribute to the general interest instead of the individual profit. The second one is the fundamental research, which is a new mysterious chapter where ML interacts with cognitive psychology and others fields in order to build more powerful models that can “think”. While the first way is relatively easier, the second way is riskier.

ML Taking the Lead

ML has in fact taken a partial lead over humans in many tasks where only a similar system 1 of our brain is involved. The more we tune models’ parameters and structures, the more we increase these machines capabilities. So far, machines are still the student that cannot surpass the professor for many reasons, especially the missing ability of generalization outside the distribution sets, which is positively correlated with our ability to incorporate an analogical system 2 into them. While this mission seems impossible to achieve, we may witness some fantastic breakthrough over the upcoming years. While we provide machines with tools to evolve, we should keep thinking about ways to tackle effectively policy, ethics, and social impact topics.

With all these achievements, the path looks more exciting and fascinating than before. While the destination is unclear and far away, we are required to keep doing our best, enjoy to the max, and look forward to the fascinating discoveries the future is preparing for us.

What are your insights on the topic? I am looking forward to hearing them :).