Conferences are the best way to stay up to date on the latest developments in a fast moving field like AI. Bitfactor’s AI team visited the ICML 2018 Artificial Intelligence conference in Stockholm mid July.
What is ICML?
International Conference on Machine Learning (ICML) is one of the leading AI events among the deep learning focused International Conference on Learning Representations (ICLR) and conference on Neural Information Processing Systems NIPS. For application oriented folks there are also events such as Recommender Systems (RecSys), International Society of Music Information Retrieval (ISMIR), computer vision conference (CVPR) and Intelligent Vehicles Symposium (IV). ICML has grown to a massive conference, so it is best to concentrate on a specific topic as it is practically impossible to experience it all.
ICML is divided into three parts. Tutorials on the first day helped participants learn new skills and introduced them to new areas of AI. The main conference consisted of keynote lectures and short white paper presentations divided into topics like reinforcement learning or generative models. The conference ended with workshops that were more focused and smaller in scale than the main event.
The workshops varied from the highly theoretical to the practical, like how to use deep learning in safety critical systems or in healthcare. Having done data science in the marketing and customer analytics domain, I found the healthcare workshop highly interesting. From data science perspective, many machine learning techniques used in healthcare are also applicable to customer analytics.
There were many intriguing and fascinating presentations on applications too. There is a lot of activity in the finance domain but it tends to be a bit secretive. However, JP Morgan has made a trade execution application based on reinforcement learning and they have even published a very interesting white paper on using deep learning to hedge portfolio of derivatives. Another big application area seems to be cybersecurity and how to use deep learning to detect viruses and other malware.
There were also world-class networking opportunities with academics and practitioners.
Lay of the land, AI wise
So what is the current status of AI at the moment? AI has become a large field and I will concentrate only on few specific topics here.
In reinforcement learning computer learns by performing actions. Whenever there are news that a computer is outdoing humans in computer games it is usually due to reinforcement learning.
To cut the long story short, reinforcement learning means that the agent (AI, the one making decisions) takes a look around (observes a state), picks an action and receives a reward or a punishment. If the action works, the AI is encouraged to repeat it; otherwise AI is discouraged to use the action in that particular situation. Then the cycle starts all over again.
Most reinforcement learning systems use model free method that essentially tries out all sorts of things and then converges to ones that make the most sense. The downside to this approach is that learning to act correctly takes a lot of unsuccessful actions too. In the physical world a single wrong action can be fatal, like driving a car into a ravine, and therefore model free approach is not really applicable. To employ reinforcement learning in the real world, the AI must in most cases first learn the model of the world and then use that model to predict the effect of its actions.
In addition to reinforcement learning, I would like to note many presentations and white papers on privacy and fairness in AI systems and on adversarial attacks to AI systems. Importance of privacy and fairness increases daily as AI systems become more and more common.
Adversarial attacks use weakness that exists in most deep learning based AI systems. You can tweak inputs, like some pixels in an image, to confuse the AI system. For example, using this method it’s possible to trick AI to think that a boat is a bird but it’s not yet clear how big of a practical problem this is and how to best defend against it. Also, when an attacker is using AI though API the attacker receives limited information on how to tweak the image in an adversarial manner.
AI is an art form and a science
Engineering uses theory to guide R&D. On the other hand, Mythbusters just make some back-of-the-envelope calculations and then try them out in practice to see what happens. Amazingly, sometimes what should work doesn’t and vice versa, you just can’t always know in advance. Theoretical understanding of deep learning and AI is certainly going forward but it isn’t yet clear why it works as well as it does. Currently many things are just tried out based on intuition and some theory but more robust and profound theoretical understanding is progressing.
How to create value from AI in business context
As always in data science, it is best to concentrate on very practical and concrete business cases. Here’s how to start:
- Automate tasks that are currently done manually and release your employees’ creative potential.
- Optimize logistics, improve quality or increase your company’s operational efficiency.
- Provide services that were previously prohibitively expensive or technically infeasible.
Don’t forget that often AI is just a part of a larger solution. First automate some specific high volume services using AI while utilizing people on tasks that require human judgement.
It’s important to understand what can be done today using AI and what is the best way to do it. Bitfactor’s AI team keeps itself up to date on latest developments and we’re happy to help you use AI in a responsible, profitable and sustainable manner to take your business to the next level.