WEBVTT 00:00.080 --> 00:00.920 Welcome back. 00:00.960 --> 00:06.560 We succeeded in splitting the data into test and training sets. 00:06.600 --> 00:12.360 Now we're going to introduce a very important concept called feature scaling. 00:12.480 --> 00:20.160 Feature scaling is especially crucial for neural networks and many other machine learning algorithms. 00:20.520 --> 00:27.120 It standardizes the range of independent features in the data to a common scale. 00:27.160 --> 00:28.840 Back to our model. 00:28.840 --> 00:30.320 We have the input. 00:30.360 --> 00:34.920 In the input layer we have nine parameters cylinders. 00:34.920 --> 00:35.880 Displacement. 00:35.880 --> 00:42.360 Horsepower, weight, acceleration, model, year, origin one, two and three. 00:42.480 --> 00:52.080 So if we look at cylinders parameter we have numbers and the range is between 3 and 9 while the displacement 00:52.080 --> 01:02.480 is different scale which is CC and it's range from six 8 to 4 five five, while the horsepower is another 01:02.570 --> 01:12.410 Measurement and another feature with horsepower, HP and range 46 to 230, while the weight is in pounds 01:12.450 --> 01:18.490 ranging from 1613 to 5 one, four zero and so on. 01:18.490 --> 01:21.130 And the other features are the same. 01:21.130 --> 01:27.890 So those features have different scales and different units. 01:27.890 --> 01:31.090 How to measure the effect. 01:31.090 --> 01:44.290 In a one scale, we need to use one scale called feature scaling and make them as a one unit common 01:44.290 --> 01:44.970 scale. 01:44.970 --> 01:54.970 So feature scaling standardizes the range of the independent features in the data to a common scale. 01:54.970 --> 02:04.370 So here we need to create a common scale to measure the features because they are different. 02:04.370 --> 02:10.660 Why this is very important for neural network for the faster convergence. 02:10.820 --> 02:17.540 Gradient descent converges much faster when features are on similar scales. 02:17.540 --> 02:25.660 So when when dealing and when giving those data to the machine learning, it will learn faster. 02:25.820 --> 02:29.020 When the features are on similar scales. 02:29.020 --> 02:34.180 Without scaling, the algorithm zigzags towards the optimum. 02:34.180 --> 02:36.380 It will not reach the optimum. 02:36.380 --> 02:39.540 Also, we prevent dominance. 02:39.580 --> 02:43.300 Prevents large scale features like the weight. 02:43.340 --> 02:44.540 Look at the weight. 02:44.540 --> 02:57.340 We have £303,500 for our for our example, and the range is between 1613 and 2 5140. 02:57.580 --> 03:06.580 And this is a large scale compared to the cylinders displacement, horsepower, model year, origin 03:06.580 --> 03:07.470 and so on. 03:07.470 --> 03:10.510 So the dominant here is wait. 03:10.550 --> 03:21.830 We're not going to like detect and predict the MPG based only on the weight because the weight is the 03:21.870 --> 03:23.110 dominance here. 03:23.110 --> 03:32.590 So here we need to prevent large scale features like weight from dominating small scale features like 03:32.630 --> 03:39.190 acceleration, cylinders, displacement, horsepower, model year, origin, and so on. 03:39.350 --> 03:48.790 All features contribute equally to the learning process, so the MPG depends on acceleration, model 03:48.790 --> 03:54.190 year, weight, horsepower, displacement, and cylinders equally. 03:54.190 --> 04:01.830 All feature contribute equally to the learning process, reducing numerical precision. 04:01.950 --> 04:10.160 Precision issues during calculations and prevents preventing gradient, explosion or vanishing problems 04:10.160 --> 04:12.200 that we will see in the next videos. 04:12.240 --> 04:19.920 Okay, so the main concept here is to standardize all the features to one scale. 04:19.960 --> 04:24.240 To do that we need to create a scalar object. 04:24.440 --> 04:30.080 So here we have to use scalar standard scalar. 04:30.120 --> 04:32.920 What does standard scalar do. 04:33.080 --> 04:37.280 Standard scalar transforms or standard scalar. 04:37.320 --> 04:37.840 Sorry. 04:38.000 --> 04:41.160 What does standard scalar do. 04:41.360 --> 04:50.760 It transforms data to have mean equals to zero and standard deviation equals to one. 04:50.960 --> 05:01.520 With a formula x minus mean over standard deviation, x is the feature value over a minus mean over 05:01.560 --> 05:03.200 standard deviation. 05:03.360 --> 05:07.000 For example, here we have the weight feature. 05:07.000 --> 05:13.600 Before scaling we have 1000 for the first car 3,002nd card and so on. 05:13.640 --> 05:22.760 After scaling, we need to create a scale that allows us to measure the weight according to the mean 05:22.760 --> 05:24.400 and the standard deviation. 05:24.400 --> 05:27.960 So mean equals to zero and standard deviation equals to one. 05:28.000 --> 05:39.920 We use this formula to transform the 1500 to -1.2, and 3000 to 1.5, 2500 to 0.8, and so on. 05:39.960 --> 05:40.480 Okay. 05:40.760 --> 05:43.200 So this is the the scaling. 05:43.200 --> 05:46.800 This is before scaling and this is after scaling. 05:46.920 --> 05:55.040 Now we need to use the fit transform function x train and scaled. 05:55.280 --> 06:02.320 So this is the annotation I want from you to fill with scalar dot fit transform. 06:02.320 --> 06:05.240 And here we need to train x. 06:05.360 --> 06:14.290 Here we are calculating the mean and standard deviation from training data only then Transforming the 06:14.290 --> 06:15.690 training data. 06:15.690 --> 06:23.970 So here we use the training data, not the testing data, because we need to test our data in a reliable 06:23.970 --> 06:26.370 mode in a reliable way. 06:26.570 --> 06:30.450 We don't need to pass any data before for the training. 06:30.690 --> 06:37.850 Also, we used the X train in order to calculate the mean and the standard deviation from training data 06:37.850 --> 06:38.290 only. 06:38.530 --> 06:49.730 Then let's use the transform function x scaled equals to scalar dot transform x test. 06:49.770 --> 06:58.450 This will use the same mean and standard deviation from training data doesn't fit on test data. 06:58.570 --> 07:01.290 So those are very important. 07:01.290 --> 07:05.770 Here we are creating the scaler by using this formula. 07:05.810 --> 07:09.410 X minus mu over standard deviation. 07:09.410 --> 07:14.010 Having the mean equals to zero and standard deviation equals to one. 07:14.010 --> 07:21.020 So those data if we make it, we have the if we calculate, we get mean equals to zero and standard 07:21.020 --> 07:21.980 deviation one. 07:22.180 --> 07:24.780 Before scaling, the weight was this. 07:24.780 --> 07:31.860 And after scaling the weight becomes this okay without scaling problems. 07:31.900 --> 07:34.100 Neural network train very slow. 07:34.100 --> 07:41.020 So without using the trained scaled feature and calculating the mean and standard deviation from training 07:41.020 --> 07:46.540 data only, then transforming the training data using the fit, transform and transform functions. 07:46.580 --> 07:49.500 Neural networks train very slowly. 07:49.780 --> 07:57.020 Some features dominate, others model may not converge properly, and poor performance. 07:57.300 --> 08:07.220 Feature scaling is a critical pre-processing step that often makes the difference between a model that 08:07.220 --> 08:10.540 works well and one that doesn't. 08:10.540 --> 08:13.820 So this is a very, very important thing.