All language subtitles for 10 - K-fold Cross Validation.en

[Autogenerated] So here is there be a rat.

We know that we need to split our data

into training. Validation and test data

will produce end candidate models for

running and training and validation

processes. But just one test process to

evaluate the finding model that we have

chosen. This is referred to us singular

cross validation the splitter original

data in tow, training tests on a single

validation set. Let's visually see how we

use these three subsets off our data to

get the best possible model. We trained

the different candidate models on the

training data. Evaluate them on the

validation data. This process is called

hyper parameter tuning. Each candidate,

Morty will have different design

parameters. You're trying to figure out

which design off your model works well for

your data. And finally, after you've used

hyper parameter tuning to find the best

design for your model, you'll do our final

evaluation. The test data. So you know

this is how your model performs the EU's

off. A holdout validation set is a huge

improvement over what we were doing

earlier. However, that is still a problem.

The model's performance on the validation

sec get incorporated into the model

itself, and this may introduce bias. So

the validation set data become partof the

models designed. And that's not good.

What? We're trying to get us a model that

is as robust as we can make it, which is

why an alternative to using singular cross

validation is key. Fold cross validation.

Here. You don't have a single set off

validation data to generate each candidate

mortal. You'd repeatedly trained and

validate using different subsets off

training data. Now, this might not seem

intuitive to you at first, but we'll see

it visually and you'll understand what's

going on. Okay, full cross validation

tends to be very compute. A Shin Lee

intensive but very robust. It does not

waste. Eight are all off. The data is used

well to generate a good model. Let's

visually understand how k fold cross

validation books. You have all of the data

available to you in the real ball. You

split it into training data on dhe test

data. Test data is what you lose to

perform a final evaluation on the model.

Now instruct using the same validation

data to evaluate different candidate

models, you'll stay split your training

data into different falls. Here I have

five fold. This is fivefold. Cross

validation with five full cross validation

for each candidate model. You'll train

your model five times the first time. Fold

234 and five will be the training data.

Fold one will be the validation data.

You'll then train the same candidate of

model with a different subset of training

data full 134 and five complex. The

training data fall, too. It's a validation

data. You'll then do 1/3 round of training

for the same candidate, Marty. This time,

fold three is the validation data that the

meaning falls. Make up your training data,

and you'll continue this for split four

and split five as well. So when you use

fivefold cross validation for a single

candidate model, you've run five training

processes and five validation processes.

Training and validation is run on each

full off your training data. Once you run

these five different training and

validation processes, you average the

performance off this candidate model

across all fools, so you'll get one

average score. And for this particular

candidate, Morty, this average performance

scores what you'll use to find the best

candidate model, which candidate model has

the best average performance school across

all falls off training and validation.

Once you've trained all of her candidate

models on all of these falls, on average,

their performance score, you'll take the

best one that you phoned, evaluated on the

test data. Thus, with careful cross

validation since the validation data

changes in each fold off training, it's

impossible for the information in the

validation data to become incorporated as part of the model.