M-Step: Updating Parameters
What you'll learn: How the M-Step uses the responsibilities from the E-Step to re-estimate the means, covariances, and mixing coefficients of each Gaussian cluster in a Gaussian Mixture Model.
The Core Idea
After the E-Step computes how much each data point "belongs" to each cluster (the responsibilities), the M-Step uses those soft assignments to update our best guess for each cluster's parameters. Think of it like this: if you're organizing people into groups based on height and weight, and you've figured out how much each person belongs to each group, you now recalculate the average height and weight for each group—but weighted by how much each person belongs.
What Gets Updated
The M-Step refreshes three types of parameters:
1. Mixing Coefficients (π)
The proportion of data assigned to each cluster. If a cluster has high average responsibility across all points, it gets a larger mixing coefficient.
2. Means (μ)
The center of each Gaussian. Calculated as a weighted average of all data points, where the weights are the responsibilities from the E-Step.
3. Covariances (Σ)
The spread and shape of each Gaussian. Also computed as a weighted calculation, measuring how data points scatter around the updated mean.
Why This Matters
The M-Step is the "learning" phase where your model actually improves. By incorporating the soft assignments (responsibilities), you get better parameter estimates than hard assignments would give you. This soft updating is what makes GMMs more flexible than K-Means.
Key Takeaway: The M-Step maximizes the likelihood by updating cluster parameters (means, covariances, mixing coefficients) using weighted averages based on the responsibilities computed in the E-Step.