Use-case: Why We Need Parameter Estimation
Parameter estimation is essential in pattern recognition because:
- In Bayesian classifiers, we must estimate the likelihood and prior probabilities to calculate posterior probabilities .
- Real-world data distributions have unknown parameters that must be learned from training data.
- Proper parameter estimation allows our model to generalize to unseen data.
- The parameters define the specific instance of our model (e.g., which Gaussian from all possible Gaussians).
Maximum Likelihood (ML) Parameter Estimation
For a dataset of observation vectors , assuming they are independent and identically distributed (i.i.d.), the likelihood of the training material is:
The Maximum Likelihood (ML) Parameter Estimation finds the parameter vector that maximizes this likelihood:
ML Estimate of the Mean
ML Estimate of the Covariance Matrix
where:
- is the total number of samples.
- is the -th sample.
- ^ represents the fact that the parameters are estimated.