Intro: Technical & Business Corner
Hi all, my name is Christian Siagian. I am your host for the “Technical & Business Corner.” I will be writing about how discoveries, both academically and commercially, have and continue to have an affect on the products and services we enjoy.
Traditionally, academics try to solve a core problem with the most impact in the most general setting, while companies apply the solution to specific use cases. However, talent composition between the two constructs, as well as long term thinking, has changed considerably these days as many fundamental research is also done in commercial labs.
We will primarily focus on human-centric topics, such as media and entertainment in its various forms (music, art, sports, etc. ), but will also venture into related sectors such as retail, e-commerce, financial technology, and more. The coverage will be for both start-ups and established companies. This will involve a building up of knowledge to allow the audience to appreciate the difficulties that researchers are facing at the cutting edge of breakthrough research.
My background in Artificial Intelligence (AI) is mainly on the vision side: robotic, computer, and biological vision. In all three areas, we are trying understand what is in visual data (image or video) using techniques that include deep learning, the current dominant approach. My dissertation topic is in vision-based robot navigation (related to self-driving), where I build a wheelchair-based robot name Beobot 2.0. Because of that, I am also trained in utilizing complex algorithms in real-time settings, a requirement for many products and applications.
Aside from academic work at Cornell, USC, and Caltech, I also worked in a start-up called AIO Robotics. There, my team and I produced a number of patent applications in 3D printing, scanning, and copying. My knowledge of designing, implementing, supporting hardware and software products, as well as learning how the business world works, is due to experience at AIO Robotics.
Even though there will be technical information being disseminated, I will not be teaching fundamentals of AI or machine learning (ML). I will add as much grounding to ensure accessibility of each article and closely discuss the complexity and nuances of the issues faced. From this perspective, the focus of conversation will be on what is possible and what is not, what is reliable and what is not, and whether a solution is still a proof of concept or already commercially viable.
Given that most AI/ML or robotics technology are not 100 percent perfect, we have to characterize under what conditions will a product become less effective . Safety, reliability, and consequences need to be built into the business model.
Let us examine image classification as an example. That is, given an image of an object, can an algorithm guess what the object is. It is not completely solved as image classification algorithms still make baffling mistakes, such as confusing a table for a giraffe. However, for many applications it is good enough. Let us look at three applications: automatic-tagging for social media posts, automated removal of blacklisted symbols (e.g. swastikas in Germany) in a movie, and road sign recognition for self-driving cars. Note, for the second application, we think of a movie as a collection of images.
Social media tagging is not particularly dangerous if the algorithm missed, and is already deployed in various similar settings. We can rely on users to correct them. The second application has to be done perfectly as a legal matter, but the algorithm itself does not have to be perfect. We can devise a system to bring operators to double check the results (termed human in the loop), but automate most of the burden. Self-driving cars, on the other hand, are the most demanding when it comes to accuracy requirements for obvious reasons.
And so, I hope from these discussions the audience will become more educated in evaluating new discoveries and on how to apply them to their problems. We can project how to build and evaluate cost of a service, or focus on specific applications where failures are not dangerous and imperfections are acceptable.
If you are in Los Angeles, please come out too one of our upcoming activities: https://AI_la.eventbrite.com