There is a popular saying that “If you can’t explain it simply, you don’t understand it well”. It’s 4 months already into my PhD. I’ve submitted my first research form and I’m already trying my hands on few ideas. My mum already sees my PhD as a big thing and she is looking forward to the day she’ll be called “mama doctor” (from me as she already has 2 medical doctors). But let’s see if I can get my mum to understand and get interested in what I’m actually doing in my PhD. So I’ll explain the ideas a bit technically and I’ll explain same thing to my mum.
Basically the project is around 3 main concepts.
1. Moving towards edge analytics
Conventionally all the data stream generated by devices are being sent to the cloud where the data is processed and analysed for specific application. Centralised learning architectures are deployed such that the learning algorithms are trained and managed in the cloud. However, this research would be focused on learning closer to the data source. A possible direction is that low-level analytics is done at the edge as a pre-filtering technique and subsequently pushed to the cloud for data fusion and possible integration with external data sources.
To mum:
Instead of us going to Daddy (the cloud) any time we have a need, how about he gives it to you so that we can come to you (the edge), which means our needs can be solved faster (real-time) and we don’t have to wait for him to come back from his work trip (latency and transmission delays).
2. Anomaly detection
Anomalies are a deviation from the expected behaviour of a system. It has a wide range of interpretation and application in various domain. In my research I would be focusing on finding anomalies in video streams to help understand how users in a scene interact with each other and with the spaces.
To mum:
You know how I can do 100 things right and you don’t even act like you notice but when I get one thing wrong, you remember it for ages. I would like to build that kind of system that is very alert to abnormal behaviour and can easily recognise and flag it.
Traditional anomaly detection systems for video survelliance
3. Online learning
To facilitate the real-time process, online learning algorithms are trained on data generated dynamically from a continuous flow of data. In a camera network, for instance, you have a continuous stream of data and you want to make some decisions on the fly or optimise the system based on the specific scene the camera is observing.
To mum:
It’s like dealing with issues as they happen without having to put them off until a later time. Wouldn’t it have been better to correct a mistake I made at the point I made it than having to store them up until I make another mistake 6 months later and you bring it up as a reference?
Now, your turn!
I’m hoping that the concepts are simple enough to grasp. But if you have any questions or you would like to share your own simple words for my project, I would be more than happy to see them in the comment box.
oh wow, this is very simplified.
You literally taught me what data science is all about.