Some thoughts on the state & future of AI
Prof. Nikos Paragios
Distinguished Professor of Mathematics at Université Paris-Saclay/CEO at TheraPanacea
Our planet – despite some recent quite ambiguous decisions regarding the impact of climate changes - has entered a new development cycle - the fourth industrial revolution, or the digital/big data/artificial intelligence era -. This revolution is the natural outcome of the digital revolution and is driven from breakthrough usages/technologies in three different areas : (i) increasing computing capabilities of sensors, (ii) growing availability of data through the internet of things and change of the data ownership model, & (iii) progress in machine learning and artificial intelligence. Let me briefly go through these aspects:
- The Moore’s law, a great computational scientist predicted that since sixties computing capabilities will keep doubling every two years. This has been observed in practice associated with constant compaction of sensors. Most likely all of us have a smart phone that is as powerful as the NASA supercomputer of sixties and seventies that was used to optimize the Apollo space missions to the moon. Furthermore, we keep observing novel computing means emerging, like for example the case of graphics processing units (GPUs) that were introduced for gaming and now becoming a transformation component in artificial intelligence due to their adequacy with deep learning methods or even quantum computing that is constantly progressing.
- The second component of this revolution is the proliferation of sensors, the amazing development of wireless data sharing/communication networks (we are moving towards 5G) and the shifting to a new era of the data creation/generation/ownership model. In the last decade of the previous century data was collected/produced/controlled from media or governmental organizations or states. This has change thanks to a number of things, like proliferation of sensors (every phone has now two or three cameras) and sensors to the internet being able to continue measure and transfer information (relevant or irrelevant). This information is produced by the user and is owned from the user who is willing to share it in the presence of some policy for its own benefit. Regarding data, soon – if it is not the case already – the photos uploaded daily to Facebook if printed in paper and stack together they should suffice to create a bridge connecting earth to the moon. The application Waze – bought from Google - is another example where we do share our driving habits/trajectories/etc. in exchange of a more accurate driving recommendation option.
- The last component of this revolution is machine learning and artificial intelligence that consists of expressing a mathematical logic into a computer generated solution that is able to reproduce this logic. This is the result of advances in mathematical and computing knowledge, coupled with accessibility on data and computing resources. In late 50s Alan Turing – an exceptional computing and cognitive scientist- predicted that in 2000s an average human will not be able (with 70% chances) to distinguish whether he is communicating with another human or an AI computed generated interlocutor. Personal assistants from GAFA are not there yet but constantly progressing and becoming reasonably efficient. Other scientists predicted that algorithms and their computer generator alter egos will outperform humans in intelligence games like chess. It happened for chess (deep blue versus Gary Kasparov in 1997), Go (deep mind versus Lee Sedol in 2016). In both cases humans were against machines in the context of a deterministic action system where at each step one seeks simulate further and deeper in time as many solutions as possible and make the best possible choice. In this setting, where the objective is to provide a deterministic action from a set of options, it is certain that given the increasing availability of data and computing power humans will not be able to compete with AI…Deep and reinforcement learning are methods that once fed with a lot of annotated data are able to reproduce the underlying logic and therefore for specific tasks to go from individual to collective intelligence. However, it is important to note that most of the AI algorithms do not really understand the problem that they are solving but simply reproduce the behavior of the data. An automatic translation system does not understand neither the origin or the destination language or the text on which they perform translation.
The rise of these systems is present all domains of our lives and raises numerous of questions:
Ethics: several concerns are risen from the use of data and the personalization of decisions based on them; one can cite numerous domains like for example health providers which should in theory being able to choose subscription plans depending the on life expectancy or even predictive health conditions? Several issues should be appropriately addressed like for example how does legislation imposes rules on black-box algorithms that will eliminate bias and discrimination? How does legislation protect and guarantee that personal data is used appropriately?
Accountability: algorithms soon will be able to make life threatening/saving decisions in different fields like for example in automotive industry, or in the health sector. Despite impressive performance on learning and testing data, their generation and their behavior on unseen examples/situations are unknowns. These systems target to reproduce human intelligence without understanding it. The human decision model is based on what we call expert systems, AI algorithms – at least the ones largely adopted - not! To whom the responsibility belongs? To the designer of the algorithm (scientist)? To the provider of the algorithm (company that commercializes the solution)? Or to the integrator of the AI solution to the final product? Or to the final user of this technology? Autonomous driving is a great example where these questions remain unanswered.
Small Data: well most of the existing top-performing solutions require massive annotated data for training. Furthermore, they are based on highly complex, non-humanly understandable mathematical models…In numerous domains, these data will never be available (like for example rare diseases or accidents). How do we train algorithms to solve such problems? Two possible options, the first consists of the so-called transfer-learning which simply transfers knowledge from one domain to another. The second comes from research on what we call deep architectures. It appears that once the model is trained, we are able to maintain 100% of each performance while throwing out more than 90% of its parameters! In simple wording, we do not need that complex models! Simpler models are can do the same job which require much smaller data sets to train!
Knowledge discovery and Integration: Last, but not least artificial intelligence solutions target to answer a specific question and can outperform humans on this question. Humans do employ global reasoning since quite often there are multiple phenomena that can explain a single observation. Detecting a tumor helps a lot, but the treatment is not the same if it a primary, metastatic or a secondary tumor. Humans do reason globally and integrate information coming from various sources. Computers cannot do that yet…since state of the art solutions seek to reproduce the underlying behavior of the data without understanding it, their ability to reason is limited and quite compromised with the current artificial intelligence trends.
Once these questions/concerns have been answered, artificial intelligence could deliver fantastic / successful solutions under certain conditions
- addressing full scale problems! We cannot reason on a small part of the problem and leave the rest to others. Successful researchers/entrepreneurs/companies should be able to sole large-scale problems and their solutions should target integrated knowledge discovery!
- with unique domain expertise and knowledge: it takes 10 years to become a radiologist, and a radiologist is not able to reason/answer a single specific question but is able to interpret/integrate initial knowledge with life-learning mechanisms that are driven from anatomical and physiological evidence. Assuming that in the future black-boxes will be able to do the same while being unable to understand the domain is laughable. Algorithms should directly integrate part of this knowledge in their conception and should be able to go beyond reproduction of observations.
- with algorithmic supremacy: knowledge is not on the observations but on the mathematical logic behind the interpretation model. The scientific approach is critical and will be always critical. Surprising enough we have observed 10 years back the rise of millions of self-proclaimed experts on big-data that convert themselves to data sciences five years later and now to artificial intelligence ones. Invest on knowledge, invest on science, this makes the difference when seeking to reproduce complex intelligence human behavior.
- and with unique access to data: well, the best models need to be trained! That is what humans do as well. The training data play a tremendous role on the performance of the algorithm, and aspects relative to their quality, their ability to represent the entire set of scenarios that the algorithm should be able to answer are fundamental. Unless access to such data sets are granted, solutions will be limited, most likely not generalizable and with unpredictable behavior.
Let me conclude by stating that our world is under an amazing transformation where data, artificial intelligence and computing would/could make a huge (positive) impact. I think that making predictions regarding the future is quite challenging but personally I think that we should moderate our expectations and address before a number of issues that are byproduct of this revolution! Ethics, legislation, algorithmic transparency, privacy and even more important tremendous transformation of our societal model are to be addressed. It is not that hard to predict that technology progress could lead to a complete collapse of the job market for low and middle level jobs in a number of sectors through automation and artificial intelligence while one can imagine also that this progress could potentially further increase of the gap between the privileged and the unprivileged society parts.
follow me on twitter: @agonigrammi
Great article, there should be more of these to tell everyone what AI is not: it is not a magic black box, it doesn't solve every problem. I especially appreciate your mentioning the case of very rare events where by definition there's hardly any data to train the algorithm on. One should remember that intelligence is made of intuition, analysis, imagination and reason. AI to me is mostly the analytical part. It is a condition to intelligence, but by no way sufficient.