Continuing its Rethink Webinar Series, Endeavor Greece invites Constantinos Daskalakis, Professor at MIT, Serafim Batzoglou, Chief of Computation at Seer & Co-Founder at DNAnexus, Marios Stavropoulos, Partner, General Manager at Microsoft and Founder of Softomotive, to a discussion concerning the limitations that Machine Learning and Big Data pose towards the attainment and understanding of true, general AI, and the sectors in which these advances in technology will have the greatest impact. Moderator, Konstantine Arkoudas Senior Manager of Applied Science at Amazon’s Alexa AI, brought many thought-provoking topics to the table and got the panelists to express their true opinions on the future of AI.
When asked what Machine Learning does best, Mr. Constantinos Daskalakis tells us that though not the most interesting form of Machine Learning, the most transformative application of it in the internet economy is the way it’s programmed to understand how to sell things to people and how to present them with engaging content. Mr. Serafim Batzoglou acknowledges the advancements in voice recognition, fraud detecting in trading, machine translation, and the impressive developments in search engines. If he were to predict the future, the healthcare sector with the help of the networks opening their data would become a priority in terms of quality and economics. Access to data such as patient records, imaging, lab results, and treatment tracking, from a few million people is needed for the sector to open its doors to what AI has to offer. Additionally, Mr. Batzoglou believes the military would be another sector impacted by Machine Learning in the near future as both himself and Mr. Daskalakis predict that over the next 10 years we’re going to see the first war that is won primarily by AI.
In regards to killer applications of Machine Learning, Mr. Stavropoulos highlights that email and the World Wide Web have driven the adoption of certain technologies. Due to this, we have seen Machine Learning applied very widely in computer vision, in addition to image, face, speech, and pattern recognition.
Mr. Daskalakis expresses that Deep Learning has driven a lot of the recent successes in Machine Learning and Artificial Intelligence, but to not forget that neural networks have been around for many decades and the differentiation between the terms and power they hold, is the amount of data and the access to data. This access to data became apparent through the accessibility of the internet and the availability of cheaper storage. Another powerful aspect that helped accelerate the Deep Tech world, is graphical processing units as they allowed computers to perform fast linear algebra operations, which is the heart of what is needed to train neural networks on vast amounts of data. Lastly, though very involved in the making, production, and understanding of Deep Tech and AI, he claims that Machine Learning technology is powerful, but has made the world a bit more dangerous as nobody really understands the interactions from a scientific standpoint.
He uses the analogy of little “blackboxes”. We only know the input and the output, but what happens inside the “blackbox” is unknown. The results it produces are evidently a byproduct of the operations, yet the operations are unknown and vary in different scenarios. Mr. Batzoglou adds to Mr. Daskalakis’ point by describing that the differentiation between Deep Tech and Machine Learning is the convergence of data, GPUs, computing power, and technical advances that have been able to train the multilayer networks. Mr. Stavropoulos chimes in with an interesting point; Quantity matters. The abundance of data and processing power available today is a prime example as the neural network has more capabilities through its processing units to make them as powerful as possible.
“Pattern recognition is not intelligence”
– Constantinos Daskalakis
Mr. Daskalakis agrees with the statement Mr. Arkoudas made, being that there is a domination of Machine Learning and Deep Learning in AI, to the point that the conflation of the two sciences are not healthy to the ecosystem and are inaccurate. He expresses that we cannot attribute the success of AI purely to Deep Learning. Even though Deep Learning is an important component in its evolution, it is undesirable for the field to have attention focused on one specific aspect. In doing so, one misses all the mathematical grounding and connections with the other domains that are equally as important in reaching the ultimate goal; to develop truly intelligent algorithms.
Mr. Batzoglou attributes the attention to the terms, to the huge advances we saw in a short amount of time. Mr. Stavropoulos claims that the dominance of Deep Tech in comparison to others is expected and not to worry about the overload of press because what matters is that those in the field know the distinction and the other capabilities the technology has to offer. The attention it’s gaining, regardless of the type, is important as it ignites conversations that need to start happening due to the revolutionary advancements approaching in the next few years.
AI Solutionism is the belief that Machine Learning is a silver bullet that can solve any problem, along with a general eagerness to use Machine Learning to solve the problem. Mr. Arkoudas asks if the panelists have any words of advice on how practitioners or business people can use Machine Learning, and on the flip side, the main pitfalls to navigate in the decision of whether or not to implement Machine Learning in the first place.
Mr. Daskalakis’s answer is twofold. The first step would be to understand the domain in which you’re operating. Before attempting to implement Machine Learning, he recommends mastering the easy methods and logistics, and understanding the baseline techniques before moving on to improve on those practices. The second step is comprehending that prediction is not understanding. One needs to make counterfactual predictions because oftentimes we’re not interested in making predictions within the same environment in which the data was acquired.
Adding on to Mr. Daskalakis point in fully understanding the domain one is operating in, Mr. Batzoglou suggests that at least one leader of the company needs to be technically savvy in order to be able to decide whether Machine Learning can be used, and to what extent it should be used that could benefit the team and the business.
“Business leaders in the company need to be familiar with Machine Learning.”
– Serafim Batzoglou
The protective nature of healthcare networks with their data alongside the fear of individuals in providing it is what according to Mr. Batzoglou, attributes to the lack of genomic medicine and Machine Learning in healthcare. The lack of information isn’t entirely dependent on the patient, but more so on the vested interest of the health networks being cautious in trying to do the right thing, resulting in them hiding behind patient privacy. Mr. Batzoglou believes that just like patients have the right to privacy, they should also have the equal right to transparency.
Mr. Arkoudas brings to light how Machine Learning is used in situations that are “delicate” and could impact society at large. Such applications could deliver negative repercussions on entire groups of people and it has been recognized that if the training data reflects existing biases, it will inevitably carry over to the predictions.
“Do you think that avoiding bias and unfairness is a problem that is itself amenable to technological solutions?”
– Konstantine Arkoudas
Mr. Daskalakis expresses that the issues at large are the paradigms used to consume data and the inability to comprehend that predictions are not understanding. If one doesn’t understand causal relationships between the underlying variables of the problem, one won’t understand what the important factors are through which you should be basing your decisions on versus the coincidental patterns that can relate to decisions made in the past by somebody else. Create a system in which evolves with a true understanding of the world and not a system based on pattern matching to the past. Oftentimes, people make decisions assuming that the data consumed is independent of one another, but we interact in society and consume and produce data in a social network. The dependencies in the data and on each other make it reactive and strategic; giving it agency.
“The machine is always right.”
– Marios Stavropoulos
Mr. Stavropoulos claims that the level of authority and trust we attribute as a society to a computer system is immense and we have all been raised with the notion that the computer will always be right. It’s a danger even more so when we don’t exactly know how these results come. One of the most important areas in AI right now is the attention people are paying to the grounds for the outcome that the “blackbox” produces, known as “right to explanation” in Europe.
Mr. Arkoudas poses a question to the panelists in regards to a potential dystopian scenario that makes people buzz when one discusses the future of superhuman AI, and the potential existential threat it can pose to humanity. Mr. Daskalakis believes it’s possible, but highlights the potential cooperation and integration between the human species and the technology developed by them. Given the fear we have, he sees a future where humans, their algorithms, and their technology are put together. On the flip side, Mr. Stavropoulos sees many consequences associated with having AI that is so capable, whether it be more or less than that of a human. He reiterates the point of its potential use in the military, and how if we go down this path, the balance of power globally can change and bring unimaginable consequences. The common solution that many people have is “unplugging it,” which he expresses is wrong because by then, there will be so many dependencies on these systems, that much like the internet, we won’t be able to live without it.
Mr. Daskalakis answers the question of the hour by stating that “connectionists,” can achieve general AI; a prime example being the human brain. Given that we don’t understand it well, such as how we don’t understand the mechanisms the “blackbox” undergoes to produce an output, Mr. Daskalakis says that the paradigm used to train such connections is problematic and we must give the system agency before we achieve general intelligence. Mr. Batzoglou expresses that a number of components are missing such as reasoning, multitasking, learning by doing, but most of all; emotions and consciousness. Emotions are a powerful tool and heuristic that humans use in learning. He believes that we will fall under the category of AI reaching superhuman level abilities and claim consciousness. Mr. Stavropoulos agrees in regards to self-awareness with the theory of mind, and the difficulties associated with measuring consciousness. By surpassing those two hurdles, he believes that we can reach the limit of Machine Learning and intelligence.
Mr. Stavropoulos foresees a lot of disruption in the job market attributable to the automation of tasks. The emergence of a “useless class,” though not his favorite term, is true in the sense that goods can be produced with fewer people in the production process. He believes that this is something that can only be solved through policy. In agreement, Mr. Batzoglou notes that some human jobs will be saved due to the fact that we like human contact, yet other more redundant jobs will see a massive displacement of the workforce.
“Adapt to perturbations of your job.”
– Constantinos Daskalakis
Mr. Daskalakis connects this topic to that of the internet economy, noticing that in the past 20 years, there has been an undeniable impact on the labor force. Every new technology is going to bring a shift in the labor force, and the important lesson is to adapt skill-wise with the help of changes in policy.
An “AI Winter” is a term used to describe a drought in funding for research, to which all participants agreed that there will not be. The reason being, according to Mr. Batzoglou, that there is sufficient AI deployed today with an economic footprint big enough that if the funding were to stop, a competitor would take their business. Taking that thought further, Mr. Stavropoulos states that AI has become so central to technology and the economy, in terms of the deployment of money and capital, that it can’t compare to the situation in the past. Mr. Arkoudas mentions that a fundamental difference between now and the last AI Winter, (being in the 80s) is that the funding back then came from the government, and now, most comes from private investments from big tech companies. Last but not least, Mr. Daskalakis says that although we discussed many limitations to Machine Learning, at the end of the day, there is a value proposition. With plenty of data and hardware, there will be immense progress on the scientific front.
There is evidently more to be said about the impact and limitations AI has, and will have on society. The discussion surrounding what sector Machine Learning will impact next and the silver bullet mentality of always using it as a solution will continue to foster many thought-provoking conversations. As mentioned, biases in AI, along with the superhuman ability that it could potentially possess in the future will become a prominent topic in many fields. For the full format of the webinar, read the transcript available on our website.