04 Jun Convolutional Neural Networks for IoT Time Series Data
onvoluted Neural Networks or CNNs are a type of AI typically used in computer vision to process images, but they are also applicable to process the time series data we typically get from sensors in IoT. In this episode of the IoT show I speak with Simon Crosby about how these CNNs can be used to make predictions about the future and reduce the massive amounts of data we collect to just the important stuff.
In this episode of the IoT Business Show, I speak with Simon Crosby about how these CNNs can be used to make predictions about the future and reduce the massive amounts of data we collect to just the important stuff.
Simon is the CTO of SWIM.AI after being CTO of numerous other companies and being named of one of InfoWorld’s Top 25 CTOs in 2007. Simon has also been a tenured faculty member at the University of Cambridge.
Instead of using a convolution to consider adjacent pixels for computation, we can consider adjacent time slices of sensor data. When looking sideways at time, we can use the CNN to make temporal predictions about the future. Since the future comes soon enough we can subtract the predicted data values from the actual values to come up with errors that can be promptly used to improve the CNN. By repeatedly following this “rinse and repeat” cycle, we improve our model and with it our ability to predict the future of our sensor values. Importantly, the input layer of our neural net can include other sources of data. In this podcast, we demonstrate this technique with a great example.
Here’s What We’ll Cover in this Episode
- Convolutional Neural Networks – what they are and how they are usually used.
- Traffic light use case and the topology of the CNN used.
- Pricing this type of CNN in cloud versus on prem.
- Monetizing the CNN with APIs.
- The role of the digital twin
- Using CNNs for data and complexity reduction.
Mentioned in this Episode and Other Useful Links
Support this Podcast
If you have been enjoying this podcast, there are a few ways you can support it:
Have an opinion? Join the discussion in our LinkedIn group
Do you prefer to use CNNs or Markov Chains on time series data?Click here if you have an opinion on this podcast or want to see the opinion of others