WEBVTT

00:00.000 --> 00:08.480
Training and testing. AI models. In this video, we will understand the process of

00:08.480 --> 00:14.200
training, validation, and testing in AI. Recognize the importance of using

00:14.200 --> 00:21.240
balanced data sets. How AI learns, like a student. Imagine teaching a student

00:21.240 --> 00:26.640
new things and then giving them a test to see how much they understood. That's

00:26.640 --> 00:32.800
exactly how AI learns. During the process of training our AI model, we divide

00:32.800 --> 00:38.040
our entire data set into two categories. Training data set and testing data

00:38.040 --> 00:44.000
set. First, we use a training data set. Let's say we want to teach an AI to

00:44.000 --> 00:50.520
recognize apples, bananas, and oranges. We start by giving it a training

00:50.520 --> 00:58.880
data set of 1,000 images, each labeled clearly as apple, banana, or orange. The

00:58.880 --> 01:04.020
AI looks at these pictures and learns to spot patterns, like color, shape, and

01:04.020 --> 01:14.160
texture. Validation and testing. While the AI is learning, we give it 200 new

01:14.160 --> 01:23.080
images to see how well it's doing. If it makes mistakes like confusing apples

01:23.080 --> 01:29.760
and oranges, we adjust the model to help it improve during training. Once training

01:29.760 --> 01:36.720
is done, we show the AI a testing data set, 300 fresh images of fruits it has

01:36.720 --> 01:42.280
never seen. Now we check how many it gets right without any help. Does it call an

01:42.280 --> 01:49.840
apple an apple? Yes, it does. Does it confuse a banana with an orange? No, it

01:49.840 --> 01:59.080
correctly tells them apart. Balanced data is important. Imagine two fruit

01:59.080 --> 02:05.320
baskets. One basket is filled with 90 apples and just 10 bananas. While the other

02:05.320 --> 02:13.000
basket has a balanced mix of 33 apples, 33 bananas, and 34 oranges. If the AI

02:13.000 --> 02:18.560
only sees one type of data, like mostly apples from basket one, it can become

02:18.560 --> 02:25.520
biased and might wrongly call a banana or orange an apple later. But if the

02:25.520 --> 02:31.280
AI learns from basket two, it sees all fruits equally and can recognize them

02:31.280 --> 02:36.800
correctly. That's why we need a balanced data set with equal and fair

02:36.800 --> 02:42.120
examples from all categories. So the AI learns correctly and treats all inputs

02:42.120 --> 02:49.840
fairly. Checking model performance. We use performance metrics to know if the

02:49.840 --> 02:56.640
model is good. Accuracy. How many answers are correct? Precision. How many

02:56.640 --> 03:03.360
chose and answers were right? Recall. Did it find all the correct answers? We

03:03.360 --> 03:07.800
can compare our model using these metrics and improve them by adjusting data

03:07.800 --> 03:14.960
or training methods. AI models learn just like students through practice, feedback

03:14.960 --> 03:21.040
and testing. With the right data and evaluation, we can help them improve and

03:21.040 --> 03:24.280
make smarter decisions.