This is the implementation of the principle of practical not perfect. When will your model be usable? When should you stop improving it? If this step is missing, you can have runaway cost, poor performance, or a model that doesn't work sufficiently and is misleading. Note that after we calculate error on batch, we can either keep going or we can evaluate the model. Evaluating the model needs to happen on full dataset, not just a small batch. If you have one pool of data, then you'll need training data and validation data. You can't use it all in both places or you won't get measurable error. Training and evaluating an MO model is an experiment with finding the right generalizable model that fits your training dataset but doesn't memorize it. As you see here we have an overly simplistic linear model that doesn't fit the relationships in the data. You'll be able to see how bad this is immediately by looking at your lost metric during training and visually on this graph here as there are quite a few points outside the shape of the trim-line, this is called under-fitting. On the opposite end of the spectrum is over fitting, as shown on the right extreme. Here, we greatly increased the complexity of our linear model and turned it into an nth order polynomial, which seems to model the training dataset really well, almost too well. Well this is where the evaluation data set comes in. You can use the evaluation data set to determine if the model parameters are leading to over fitting. Over fitting or memorizing your training data set can be far worse than having a model that only adequately fits your data. If someone said they had a machine learning model that recognizes new instances and categorizes the correctly 100% of the time, it would be an indicator that the validation data somehow got mixed up with the training data, and that the data is no longer a good measure of how well the model's working. Read these slides backwards. If the question says data is scarce, then you should be thinking independent test data or cross validate our candidate answers. Be familiar with the various methods of cross-validation including training validation and test and cross-validation. Expect to know the basics of TensorFlow and key methods. They're covered in the data engineering courses. So to recap, you need to know regression and classification, labels, features, you need to know the progression of train, evaluate, and predict, and you need to understand some basic TensorFlow API calls.
