In simple words: Training learns weights; inference uses the learned weights.
Training versus inference
During training, an ANN uses past labelled records to adjust weights. During inference, the trained model accepts a new input and makes a prediction. A website visitor usually sees inference; training happened earlier.
Tuning responsibly
Try a few learning rates, batch sizes, or layer sizes and compare validation results. Then evaluate the chosen model once on a reserved test set. Record what data was used and what the score means.
See the flow
past labelled data→training→saved model→new input→inference
Simple example
A customer enters a new support request. A previously trained model predicts its topic. No new training is required for that single prediction.
Remember: Training learns weights; inference uses the learned weights.