Sample length influence on model accuracy

Help
#1

Hello,
my question is about how sample length can affect model accuracy. I have gathered accelometer and gyroscope data for 7 different motions. Each motion has different duration but some of them are really close to each other. I tested model with live classifications and what i noticed is that it focuses more on durations and recognizes samples with very distinct lengthes better than that have simillar durations. Also it should be considered that my dataset is really small(100 samples for each class with 400-700ms length). I guess i must increase my dataset for better results but before that i wanted to know if fixed sample size can benefit me. For example if i choose fixed size like 1000ms where actual motion is 1/4,2/4 or 3/4 of it and rest is idle(not padding value) how that can affect on accuracy. I think it may focus more on motion but i also think it might get harder for model because sample lenght is crutial information. Also number of classes might increase up to 17 in future.

Thanks in advance

#2

Hi @mikaL

To improve model accuracy, you can try the following steps:

  1. Increase Dataset: Neural networks perform better with more data. Capture more images to expand your dataset.

  2. Ensure Data Similarity: Ensure the data used for training looks similar to the data the model will encounter in real-world scenarios. If discrepancies are found, move the sample to the training set and update the label before retraining.

  3. Increase Training Cycles: Increase the number of training cycles to see if performance improves.

  4. Avoid Overfitting: If the model performs well on the training set but poorly on new data, it might be overfitting. Add more data or reduce the learning rate.

  5. Adjust Neural Network Architecture: Experiment with the number of layers and neurons to find a better fit for your data. (autoML tuning is available via EON tuner for enterprise users)

For more tips, check out this guide::

Best

Eoin