Read the evidence your model leaves behind.
Use this compact guide before changing a parameter. One controlled change tells you more than five random ones.
01
Training data
Your model can only learn patterns represented in the examples you choose. A varied set helps it separate class structure from lighting, color, or background shortcuts.
02
Transfer learning
A fixed MobileNet backbone has already converted each image into a compact numeric embedding. You train only the small decision-making head, which is fast enough for the browser.
03
Overfitting
A model overfits when it becomes excellent on familiar examples but brittle on new conditions. Hidden validation measures that gap without revealing the answers.
04
Activation functions
ReLU keeps positive signals, sigmoid compresses values between 0 and 1, and tanh spans −1 to 1. No activation is included as a deliberately weak experiment.
05
Softmax
The final three outputs are converted into probabilities for Nova, Orbit, and Pulse that add to 1. The largest probability becomes the prediction.
06
Learning rate
This controls update size. Too small may learn slowly; too large may jump past useful solutions. Compare curves, not just final accuracy.
07
Epochs and batch size
An epoch is one pass through the 60 selected examples. Batch size controls how many examples contribute to each weight update.
08
Validation
Validation estimates generalization. You receive only aggregate metrics so the hidden set remains a trustworthy instrument across all ten attempts.