Machine Learning uses many specific terminologies used for describing conditions or entities. This article aims to briefly explain common machine learning jargons that are important to fully understand machine learning.
We use a case for explaining these terms: a machine learning algorithm has the task to predict the likelihood of a basketball team winning a matchup given a sufficient amount of data.
Hypothesis Space is the set of all possible output from the current dataset. The Machine Learning algorithm would determine which best hypothesis available from the hypotheses space to predict the output. For the case, the hypothesis space is a real number between 0 and 1.
Instances are the objects of data used for learning the algorithm. The instance may include multiple variables /attributes that are relevant for predicting the target output. For predicting win likelihood, each instance could be data of each game from last season containing final scores, point difference in each quarter, 3-point shots made, team rebounds, team assists, etc.
Loss Functions are the functions in machine learning to measure the model's performance. If the loss function’s value is high, then the difference (error) between predictions and target output is too high; Hence, adjustments are necessary to reduce the error. Loss Functions such as Mean Squared Error, Mean Average Error, or Mean Bias Error is suitable for measuring likelihood.
Inductive Bias are assumptions used by the model to predict the outputs from inputs or instance outside the training dataset. For instance, the linear regression model assumes that relationships between attributes and output are linear. If this model is used in the case example to represent the connection between points scored and winning likelihood, then the mapping functions from points to likelihood are linear.
Underfitting and Overfitting refer to the model's poor performance in finding patterns from the data. Underfitting means that a model is incapable of finding the pattern from training data or making predictions for the data.
On the other hand, Overfitting means that the model captured patterns within the data to detail. Random fluctuations and noises in the training data are also considered part of the patterns.
This could lead to high performance in training data but low performance in new data since the model ‘memorize’ the training data rather than generalize it. Underfitting is easily recognizable when using suitable performance metrics; however, more advanced techniques such as cross-validation are necessary for avoiding overfitting.
Bình luận