Faglige nøgleord: Machine learning, embedded systems, embedded machine learning, concept drift, internet of things
Oplæg tilgængeligt på: Engelsk
About:
My PhD is about AI and machine learning for very small computers, known as embedded machine learning (EML). More concretely, I work on either downscaling existing models, or developing my own models such that we can use machine learning on internet of things (IoT) devices. IoT devices are small computers capable of gathering data and communicating over a network, for example bluetooth or WiFi.
Description
In the presentation I will start with my own journey. As someone who actually was not a big science nerd coming into high school, and who was very tired of the schooling system when I left, I hope that my story can be relatable to the students regardless of where they find themselves. I will talk about both myself in high school as well as the political activism I did during my gap year afterwards that inadvertently led me further into science.
During university I was presented with a lot of cool opportunities to partake in student projects, from launching Norway's first functional student-built satellite into orbit to working with drone navigation. Despite being Norwegian, I actually was at DTU on exchange for 1,5 years so I will incorporate that time into my presentation as well.
Once I feel like they know my path I will be moving into my field. I will start with giving a general intuition about AI and machine learning, introduce the nuance and also pull the curtain back to show what goes on behind the scenes. However, I will not be going down into the math of (partial) derivatives unless either the teacher or students specifically ask me to. Instead I will focus on the simpler multiplicative and additive nature of how AI actually makes decisions.
With the background knowledge down I will move onto my project more specifically, I will explain some of the problems with current AI, i.e. resource and infrastructure requirements, and how going "tiny" can help solve this. For this section I can bring props to what I am working with, and you can expect several intuitive figures and examples to highlight both the downsides of traditional AI and what I mean when I say "tiny."
Finally I plan on rounding it off with where we are now, what are the current obstacles to implementation of tiny machine learning. This is where I will mostly focus on my own research, which currently will involve intuition about how we can make self-learning machines.
If wanted I am also open to talking about how I use AI in my workflows, and I can also talk more general about when and how to utilise AI to support your learning rather than as a simple "shortcut." This will not by default be a part of the presentation but I am happy to include it.