After seeing my models get 100% accuracy on the training data but struggle to pass 80% on the test data I found that suspicious, especially after trying several regularization methods and several (dozens of models and NN architecture changes).
I began visualizing the data (it's 500 features, highly correlated) and noticed that some classes in the test set looked very different compared to the training set.
I asked my manager about how this data was generated. He said the embedded systems took measures over 3 full days and then a week later they took a full day. The first 3 days were the training data and the 1 day (a week) later was the test day.
I told him, you can't expect a model that is training on a different time period to be tested on one random day.
I told him we probably need several days sampled at various times, across the year even to get a good representative balanced dataset.
He said it was impossible and that a competing client got 97.5% accuracy on the test data.
I explained my findings when shuffling the data, but he said we can't do that.
I'm a tad frustrated now, is there something I'm overlooking here, or am I fundamentally right?
Most of the datasets I work with are created by my team and I, I rarely ever get a random dataset like this.
I've tried so many models, from simple to complex NNs, tried a bunch of regularisation methods, tried PCA (made things much worse at 60s% test accuracy), tried data augmentation techniques, clustering (which is how I found the differences between Class A in one dataset vs Class A in the other).
I also suspect some classes in the test data and mistakenly labeled, but he said the client assured him they are not.
Thoughts?
Can I add potentiometers (or some similar volume control device) to speaker set B to even it's volume out compared to set A?
Speakers A are in a decent sized living room, and are nominally 6 ohm floor standing speakers (2).
Speakers B are in a small kitchen, are (2) 8 ohm bookshelf speakers (2) up on top of a cabinet.
Walking from one room to the other with both sets activated, it's a heck of a lot louder in the small kitchen than it is in the living room. Can I install an inline volume control, like a simple potentiometer, on the Set B lines to reduce their volume compared to Set A? Would this introduce any strain on the receiver for some reason by changing the impedance it encounters, or is that totally unrelated?
If I can just add some volume control/pots to the line, what kind do I need?
Join us as we try our hand at a LIVE version of the show, chatbox very much included. This is uncharted waters for us, so we will see how it goes! Watch along for what will be either an awesome show or a complete debacle!
Either way, we will be recording, so if the live stream doesn't work, there will still be an episode today!
We will be breaking down the game last night between the Leafs and Jets, taking a walk down Byfuglien memory lane, looking into the playoff picture, and more.
Come be part of The LZ Sports Show!
https://preview.redd.it/ls752l3ovxu61.png?width=1280&format=png&auto=webp&s=df1411eb3ef1d30b9205d5e8b406e917725be33d
Long time reader first post.. I set up a train line like this (Pinker pink line), is there anyway to set up a timetable goes A-B-C-B-A:
Cheers,
https://preview.redd.it/ss6dkbtmkd061.jpg?width=1220&format=pjpg&auto=webp&s=12c8b348884d4859c7bbbdc904b128bfeae3d633
