There are many ways to answer this important question: financial, intellectual, societal, as well as political. ML has become such a major force in so many areas, from business growth and economic security to scientific development and political impact, that it is hard to overestimate its impact in the decades to come.
Let’s begin with the obvious, its economic and social impact. Let’s take just one example. Just yesterday, the co-founder of Facebook wrote an impassioned op-ed column in the New York Times calling for the US Government to dismantle Facebook, because he said, and I quote, “Mark’s influence is staggering, far beyond that of anyone else in the private sector or in government”.
Facebook and its associated products now reach 6 billion people on the planet. That’s a staggering number, by any standards, and without taking a personal stance one way or another on this issue (I don’t currently have a Facebook account, and never had one!), in large part what makes Facebook powerful is the ability to digest vast quantities of personal information about people through the use of machine learning.
It’s not restricted, of course, to Facebook. One can include Google, Amazon, Tesla, Uber, and so on. Virtually every major people-facing company is now employing the tools of machine learning to gather information on its customers and seeking to exploit their data for commercial advantage. There is nothing inherently wrong or sinister about this process, but what alarms many observers is how the tools of machine learning now make it possible for individuals to have enormous power, far more than any king or queen ever had in human history. Machine learning makes this possible.
Every technology that affords such great power, throughout human history, including nuclear weapons, naval armadas, chariots, and now machine learning-enabled software, needs to be used with caution. So, as the power of machine learning grows, its influence will become ever more dominant, and as the European Union did with GDPR, countries will slowly begin to rein in machine learning by exerting some amount of control. How far that happens will affect the extent to which machine learning continues to be the dominant technology in business, but as far as I can foresee, I don’t see anything happening in the near term that is going to affect the increasing deployment of machine learning and data mining tools.
A significant concern with using machine learning on global scales are issues, not just with privacy, but also of fairness. Machine learned models will reflect biases inherent in the data, and there are well-documented cases of gender and racial biases in machine learning systems that are causing alarm. Machine learning is used to learn face recognition models, and there’s plenty of reason to be cautious in overly relying on such models.
In terms of scientific impact, machine learning is rapidly entering widespread use throughout all of science. In my own experience, in the last few years, I actively participated in an exciting project involving the application of machine learning to analyze data coming from the Mars rover, Curiosity. I saw first hand the impact of machine learning methods on planetary scientists, chemists, and astronomers who were quick to see the enormous benefits of being able to derive complex mappings from data.
For example, Curiosity uses a LIBS (laser-induced breakdown spectroscopy) instrument to zap rocks and beams back the emission spectra to Earth. The resulting spectra (a 6000-dimensional vector) tells planetary scientists something about the composition of rocks on Mars. Traditionally, planetary scientists use simple rules based on examining the peaks in the spectra: if you see this peak, then it probably reveals the presence of silicon dioxide (SiO2) etc. Well, now, with the modern tools of machine learning, we were able to show how based on a set of labeled data involving rocks on earth, one can quickly within a matter of milliseconds provide far more accurate analyses of Martian rocks, even to the point of discerning which laser readings to ignore, because they were contaminated by Martian dust. This is just one tiny example of the power of machine learning to influence many scientific studies, from biology to sociology.
Leaving aside these economic issues and application domains, one can ask what the major challenges are within machine learning itself. To me, if I had to try to summarize the major challenges that face ML, I would characterize the major open problems as follows:
- Large datasets do not always lead to success: nowhere is this more obvious than language. ML-enabled language models can now be run on text datasets that are many orders of magnitude larger than anything a single human can read over their lifetime. Yet, text-based learning methods have yet to come remotely close to the ability of humans in understanding language with all its richness. Humans use language in many ways, from giving commands to asking questions, to creative uses like poetry and metaphor. Take this beautiful poem by Emily Dickinson. She was a world-renowned poet who lived in Amherst, a small town in Western Massachusetts where I spent the last 16 years. Who can deny the beauty and depth of her verse? Can any language learning system understand, let alone produce, such verse?
“Because I could not stop for Death –
He kindly stopped for me –
The Carriage held but just Ourselves –
2. Explaining how brains produce minds: current explanations of how neural activity result in intelligent action are highly inadequate and utterly unrealistic. Humans do not need millions of examples of dogs to tell them apart from cats. It is utterly implausible that some mysterious gradient descent procedure is tweaking the synaptic strength of hundreds of billions of neural connections, for every image, sound, or activity we perceive. This explanation becomes even more implausible when we see that individual neurons are largely communicating with each other using a kind of Morse code, sending spike trains of signals. How could this ever be reconciled with gradient methods? It remains a mystery.
3. Integrating knowledge and learning: current statistical learning systems are woefully inadequate in being able to represent the relational richness of the world. Currently, everything is converted to “vectors”, and this does a grave injustice to representing the richness of how social and political networks represent the world, not to mention scientific networks. Children go through a series of intellectual jumps in their early childhood, beautifully documented by the work of Jean Piaget and other developmental psychologists. Piaget showed how children gradually develop the abstract concept of an “object”. No machine learning system has ever been developed that can learn the abstract concept of an object. No deep learning system is capable of looking at a scene and counting the number of objects in the scene, something a three-year-old child can do with ease.
All of these challenges will be addressed by future generations of researchers, and as they do, the power of machine learning will inevitably grow. The ultimate power of ML is hard to gauge, as it will depend to a great amount on societal and political issues, but for the foreseeable future, it is safe to say that ML’s influence is only going to skyrocket.
If you wish to learn machine Learning courses visit