Problem in Classification

Question/Issue:
The problem is I am classifying between sanitizer, vicks and air classification model is giving 100 percent accuracy but when doing live classification i am doing it in air so it should give me air in result but there is some problem the classified data is showing in the range of air in the cluster graph but in result it is showing vicks every time dont know why even though the data is shown in air cluster area

Project ID:
project id = 252295

Hi @Intern,

Based on my initial look at your data, you only have around 101 samples for training, which is not a lot for this kind of application, especially if you are working with 2400 input dimensions. I recommend a couple of approaches to training a more robust classifier:

  • Collect more data
  • Try other processing blocks (if you are using the Syntiant hardware, this may not be feasible unless you pre-process the data prior to uploading to Edge Impulse)

I tried using two raw data processing block it worked but the main problem is it increases the size of the library.
And i will definitely try using method by adding more samples but isnt my cluster forming good there seperating prety good so why result is wrong?

Hi @Intern,

Yes, the cluster forming looks good, but keep in mind it is based on a relatively small dataset. What you’re likely seeing is the curse of dimensionality, as your inputs have 2400 dimensions. Generally, as you increase the dimensions, you need more samples to make up for the new increased spacing between each sample (in 2400-dimensional Euclidean space). With 2400 dimensions, I would expect you would need thousands or tens of thousands of samples to start getting a decent classifier.

The flip side is to find a processing block that would reduce the number of input dimensions (e.g. computing the mean and variance of each sensor channel). From what I can tell, your data does not change quickly along the time axis. As a result, I might recommend trying the flatten processing block to see if that helps you train a more robust classifier. This is, of course, assuming you are not tied to the Syntiant hardware.

1 Like

thank you for your reply I used raw data and spectral analysis it just spectrat analysis needed some tweaking and it worked like a charm.

2 Likes