This video explains the fundamental difference between traditional linear regression and Bayesian linear regression. Traditional regression provides a single 'best-fit' line defined by one slope and one intercept. In contrast, Bayesian linear regression approaches the problem by asking 'which lines are believable?' It begins with a 'prior' distribution of candidate lines, assigning a probability to each based on initial beliefs before any data is observed. As data points arrive sequentially, Bayes' rule is applied: the posterior belief is updated by multiplying the likelihood of the data given a line by the prior belief in that line. Each new observation refines this belief, causing the distribution of possible lines to narrow around those that best explain the data. Consequently, predictions made using Bayesian linear regression result in a 'band' of plausible outcomes, which is narrower in regions well-supported by data and wider in areas with less data. The video concludes by stating that the true answer is not a single line, but rather a distribution of possible lines.

DataMListic

Fit a line through some data, and regression hands you back a single answer, one slope and one intercept, the best fit.

But Bayesian linear regression asks a different question, which lines are believable? It starts before the data with a whole cloud of candidate lines called the prior, where every line you might believe gets a probability.

Then Bayes' rule takes over. The posterior is proportional to the likelihood times the prior, so your new belief is just the evidence times your old belief.

And now, watch as the points arrive one by one. Every single observation reshapes that belief, and the cloud tightens around the lines that actually explain the data.

So, when it's time to predict, you get this whole band, narrow where the data lives, and honestly, wider where it doesn't. And the single best line is just the middle of it. The answer was never one line, it's a distribution over lines.

Bayesian Linear Regression: A Distribution Over Lines

Fit a line through some data, and
regression hands you back a single
answer, one slope and one intercept, the
best fit. But Bayesian linear regression
asks a different question, which lines
are believable?
It starts before the data with a whole
cloud of candidate lines called the
prior, where every line you might
believe gets a probability. Then Bayes'
rule takes over. The posterior is
proportional to the likelihood times the
prior, so your new belief is just the
evidence times your old belief. And now,
watch as the points arrive one by one.
Every single observation reshapes that
belief, and the cloud tightens around
the lines that actually explain the
data.
So, when it's time to predict, you get
this whole band, narrow where the data
lives, and honestly, wider where it
doesn't. And the single best line is
just the middle of it.
The answer was never one line, it's a
distribution over lines.

Get Our Mobile App

Bayesian Linear Regression: A Distribution Over Lines

Transcription