Back to Main Agenda
Back to Post-Meeting Feedback
10/26/19 Gary An: Recognizing the limitations of current approaches to ML and AI (and finding a role for multi-scale modeling)
10/28/19 Misha Pavel: Rigorous modeling of human behavior
Comment
Your name
Jacob Barhak
Table sorting checkbox
Off
A few issues that were discussed at the meeting and in side talks outside and perhaps do not receive sufficient attention are:
- Machine Learning techniques include many computational techniques - some are not that new. For example k-means is a an unsupervised machine learning technique that has been used by many for a long time without giving it the title "machine learning" - so it is important to be specific about techniques used and their benefits.
- The integration of PDE/ODE and machine learning was not the best idea. It is possible today to merge solution techniques. However, this is not the major challenge. It is more important to know when to use each technique to solve problems. In other words, there was too much theory that should of been merged into the theory session on the second day - deep learning /PDE/ODE techniques are mature enough to focus more on applications.
- There was some emphasis on published papers in some cases, please note that many papers in the machine learning field are not fully reproducible, especially since the tools used are evolving fast. Moreover, due to recent hype, there should be more scrutiny on details - at the same time, many papers provide nice ideas - so its is important to talk about these as sources for ideas rather than absolute truths. The narrative on how the work is presented is important to avoid misleading the observer.
- Survano De in his pre-meeting Webinar gave a key thumb rule that defines the ballpark of accuracy of machine learning technique - those are generaly 60%-80% accurate. Sometime accuracy can be increased, yet It is important that this ballpark range be advertised and repeated in narrative to avoid giving the wrong impression of what these techniques can do. Generally speaking, those techniques automate a lot of work that would have taken much effort - they are not a "magic" solution. It is important this message will pass forward.
- One important idea that needs to be adopted into the MSM culture - many times it is more important that the machine comprehends things than the human comprehends them. This means that humans need to learn to "speak" in machine terminology and by this give away old paradigms of how things "should be done". If a machine is able to solve a problem much better than a human, it is less important if the human "understands" how this is done - human comprehension has its place and humans should accept machine comprehension as a powerful assistant rather than dismiss it if it cannot be explained. There should be bounds and tests on machine comprehension, yet the current internal psychological need of the scientists to understand everything that is going on, should not stop progress of development of better machine learning methods.