A while back I trained a small LSTM based neural net to classify the power phases of a device I work on based on their current consumption over time.
The model worked seemingly great and it took a while for me to notice that it did not catch every phase perfectly.
Yesterday I created a larger and more complex CNN based model on the recommendation of my coworkers which I trained over night since I had to use my work laptop. When applying it to my real data I ran out of RAM. After fixing this issue and getting it to run, it misclassified far too many samples.
I spent the rest of the day building an algorithmic solution that has yet to mislabel a single sample.
This isn’t really all that relevant to the post I guess but I found it a nice reminder to myself to actually think about a problem instead of throwing brute force at it and hoping it will solve it. As a side benefit, I can now actually explain why my data is classified the way it is instead of pointing at a black box. There are definitely usecases for AI but you should know enough to recognize when an algorithmic approach is better suited.
A while back I trained a small LSTM based neural net to classify the power phases of a device I work on based on their current consumption over time.
The model worked seemingly great and it took a while for me to notice that it did not catch every phase perfectly.
Yesterday I created a larger and more complex CNN based model on the recommendation of my coworkers which I trained over night since I had to use my work laptop. When applying it to my real data I ran out of RAM. After fixing this issue and getting it to run, it misclassified far too many samples.
I spent the rest of the day building an algorithmic solution that has yet to mislabel a single sample.
This isn’t really all that relevant to the post I guess but I found it a nice reminder to myself to actually think about a problem instead of throwing brute force at it and hoping it will solve it. As a side benefit, I can now actually explain why my data is classified the way it is instead of pointing at a black box. There are definitely usecases for AI but you should know enough to recognize when an algorithmic approach is better suited.