Bias Variance Tradeoff

preview_player
Показать описание
In this video, we learn about the second most important tradeoff in Machine Learning: the bias variance tradeoff. Although we are simply rephrasing the results that we got in the previous video, there are two main reasons for doing this:

1. This is the common way of looking at the approximation generalization tradeoff, even though in my opinion it should be the other way around
2. The more ways that you can independently verify your conclusion (a classic probablist/data science way of thinking) the stronger you should feel about it

Try answering these comprehension questions to further grill in the concepts covered in this video:

1. Where does the randomness come in training an ML algorithm (say linear regression)?
2. Do we ever really know the bias of a model if you only have a sample from a population?
3. Does bias depend on your data sample?
4. Why does bias + variance = test error?
5. When you have a lot of data points are low bias models better than high bias ones?
6. Why does variance always seem to work against us?
7. What is the lowest bias model?
Рекомендации по теме
Комментарии
Автор

To sum up:
- increase complexity => decrease bias / increase variance

- increase available data => bias stays the same / variance decreases

Is this right? So essentially, more data is always better and complexity of the model should depend on minimizing bias and variance.

Nixterrex
Автор

How do you think bias and variance trade off would be different if the variance of the irreducible error (sd) was higher?

janettem
Автор

Following is the code for y=ax+b

ha.fit(add_dummy_feature(x[:, None]), y)
pred_ha = ha.predict(add_dummy_feature(x[:, None]))

What will be the code for y = ax?
I am finding it difficult to set up.

KanishkJainEngineeringLovesMe
Автор

Why does this always give me an error:
>>>(X[:, None]), Y)
TypeError: '(slice(None, None, None), None)' is an invalid key

PratikGauravAironl
Автор

How do you think bias and variance trade off would be different if the variance of the irreducible error (sd) was higher?

janettem
join shbcf.ru