C1 Flashcards
3 types of machine learning
- supervised learning
- reinforcement learning
- unsupervised learning
supervised learning
- learning from example + label
- database learning
- (image, label) –> correct?
reinforcement learning
- learning by interaction
- (state, action) –> reward number
unsupervised learning
- learning without examples, from inherent measures
- database learning
- clustering, data compression, dimensionality reduction
machine learning
learning a function (function approximation/reverse-engineering a function)
- iterative minimization process of error between learned function f and data labels
- regression loss: L = mse
- categorization loss: L = - sum (y log f)
high bias
the model is too simple, so it underfits
high variance
the model is too complex (too many parameters), so it overfits
end-to-end learning
feature engineering is included in the model, no need to do human preprocessing
what are RNNs used for?
They work well for sequences, such as natural language
- has state: it can remember a situation
CNN
- spatial hierarchy
- sparsely connected layers: filters share weights
- pooling layers reduce dimensionality even more
- reduce overfitting
- unfolding layers
- suffers from the vanishing gradient problem
LSTM
better version of RNN: suffers less from the vanishing gradient problem
what is intelligence?
being able to learn and memorize
what is accuracy?
the true positives plus the true negatives, divided by all observations (computed over the training data and over the test data)
what is the confusion matrix?
a depiction of how the tp’s, tn’s, fp’s and fn’s are related and the precision and recall
what is overfitting?
when the model learns too much about the training set
