For the next interview in our series speaking to technology and IT leaders around the world, we’ve welcomed experienced CIO and lead AI expert at BagsID, Erik Van Breusegem to share his thoughts on the current state of AI.
Tell us about the business you represent, what is their vision & goals?
BagsID is bringing state of the art AI vision technology to baggage handling. Using photo recognition and machine learning will eventually remove the need for (and waste caused by) adding hardware like traditional barcode scan tags or RFID to baggage.
The goal would be less lost luggage, faster handling times and a better option for the environment, as well as more convenient travel.
Can you share a little bit about yourself and your journey into artificial intelligence?
I have a background in Computer Science and technology consulting, so that gave me a solid grounding on learning to distinguish breakthroughs from small innovations, and hype from fundamental trends.
Back then neural networks (the technique behind the latest AI push in vision) were an afterthought, literally we spent one hour on the subject in a university course on Expert Systems, with the firm declaration that “the idea is nice, but it doesn’t work” from the professor there (which was true in 1998!). On the other hand, the explosion of the internet and web was in full swing.
So, around 2012 when deep learning jumped onto the world stage I saw more or less the same thing that I saw with the web, the main difference being that speed of development and capital employed was even faster than what we saw then. Having moved to entrepreneurship by then, at that stage I started closely tracking the technology, but from my perspective it was a bit too early: the tooling was very immature, libraries and techniques were just coming out and in general it was still quite capital and people intensive.
By 2016 that had changed, which is when I decided it was possible for smaller companies and smaller teams (vs. Google for example) to start building out applied AI solutions, at least in proof of concept. That’s when I started the current initiative which led to BagsID.
What does your day to day responsibilities look like at your organisation?
In the abstract sense, I’m responsible for the AI team and related topics. In practice, this currently means recruiting, AI system design, spreading the good word in conferences and publications, talking to (prospective) customers and keeping an eye on the development of the field.
What are the key differences between computer science, machine learning and AI?
Starting with last one: AI as a term is “polluted”, as it means several different things to different people in different contexts. In the context, I operate (technically building AI systems), AI is roughly equal to deep learning (a form of neural network), and this one technique started working spectacularly well and was responsible for all the big breakthroughs of the last years.
Working our way back from there, it becomes quite simple: Computer Science is the whole field of computing, Machine learning is a subfield of Computer Science that focuses on self-learning systems (typically based on statistic or stochastic techniques), and AI/Deep learning is again a specific type of machine learning technique.
What are some misconceptions that you believe the average person has about AI?
Quite a bit, not in the least because of said term confusion. A few bullets: Deep learning (Machine Learning) isn’t magic. It’s based on math, and very powerful, but at the same time very stupid and brittle. The amount of data needed to get ‘simple’ AIs to talk compared to humans is still quite large.
AI software systems are not robotics. I blame Hollywood for this one: Most AI systems deployed today don’t have a physical component to them, but end up in e.g. in web portals, accounting software, mapping, and cameras.
The field is developing rapidly and exponentially (faster even than Moore’s Law), and we humans are very prone to wildly over and underestimating short- and long-term impact.
The work of an AI researcher involves tons of data processing, and very little actual AI design. In analogy with teaching: the teaching part is easy, getting the course materials and setup right is 90% of the work.
What advice would you give to someone wishing to start their career in artificial intelligence?
Depending on where exactly you want to go, beyond the basics (do a study) Learn how to program and build robust data pipelines. Getting a toy AI model to talk isn’t hard, getting it deployed and maintained is. Learn how to communicate: every AI system needs context, and every data scientist or machine learning engineer likewise needs to understand what the goal is of what you are trying. Do the right things, not just do things right.
Would you like to share any artificial intelligence forecasts or predictions of your own with our readers?
AI (as we define it) is here to stay and is now in the middle of the stage of maturing. This means more focus on maintainability, repeatability, explainability, and more platforms and tools that makes this easier to do (e.g. like what WordPress did for website building).
New breakthroughs and major steps in performance will occur (e.g. relatively recently with GPT-3), but they are rather unpredictable. In parallel, even without new breakthroughs raw performance will continue to grow but typically plateau in a few years, and over time it will become cheaper.
What is your experience of using AI-backed data analysis or log management tools?
AI in logging (like the anomaly detection capabilities of the Elastic Stack) can be useful as an additional tool for detecting deviant behaviour without having to set manual alarms.
In that sense, it is a useful addition to the existing toolset, with as always the challenge being a balance between sensitivity (getting all interesting behaviour out) vs. specificity (making sure you don’t get overwhelmed with false signals).
Especially for organizations that are scaling, it can be a huge cost avoidance tool, and that’s also how we use it.
If you enjoyed this article then why not check out our previous guide on everything you need to know about Gitlab vs Github or our post on why would I make a text field in Elasticseach and Kibana aggregatable?