WEBVTT 00:00.080 --> 00:04.160 Hey everyone, welcome to our Machine Learning course. 00:04.160 --> 00:11.480 Today we're going to build something really exciting, a car mileage predictor app that uses artificial 00:11.480 --> 00:17.680 intelligence to estimate a car's fuel efficiency based on its specifications. 00:17.680 --> 00:20.920 We're going to build not a simple model. 00:20.920 --> 00:29.240 We're going to build a moderate to complex model using TensorFlow, and we're going to integrate it 00:29.240 --> 00:32.000 with our Android applications. 00:32.000 --> 00:37.080 Let's start with an overview of our model. 00:37.080 --> 00:42.440 We'll have an input layer that takes nine parameters. 00:42.560 --> 00:45.360 Did you hear it right nine numbers. 00:45.360 --> 00:48.360 We're going to work with nine parameters. 00:48.360 --> 00:50.360 Let's start with the cylinders. 00:50.480 --> 00:53.720 We're going to make a range of cylinders 3 to 8. 00:53.960 --> 00:59.400 Our example will take six cylinders just just for testing. 00:59.400 --> 01:06.800 So you can put any number but just for testing and making a reference for our predictions. 01:06.840 --> 01:11.600 A displacement from 68 to 455 cc. 01:11.800 --> 01:20.720 Our example takes 256 horsepower from 46 to 230 horsepower. 01:20.880 --> 01:23.760 Our example takes 100 hp. 01:24.040 --> 01:29.960 The weight from 161 3 to 5 one four £0. 01:30.000 --> 01:38.080 Our example takes £3,500 acceleration from 8 to 25 seconds. 01:38.280 --> 01:39.360 15 seconds. 01:39.400 --> 01:45.120 Our example the model year from 1970 to 1982. 01:45.480 --> 01:47.840 Our example 1976. 01:48.000 --> 01:58.320 The parameters origin one uh refers one for American, zero for non-American origin to the the manufacturer 01:58.320 --> 02:01.960 or the country of the manufacturer. 02:02.080 --> 02:05.330 Origin two if we put one so it's European. 02:05.370 --> 02:06.970 European zero. 02:07.010 --> 02:07.970 Our example. 02:08.290 --> 02:09.410 And the third? 02:09.610 --> 02:11.690 The third origin is Japanese. 02:11.730 --> 02:17.210 Put zero to make it zero for Japanese, zero for European and one for American. 02:17.210 --> 02:22.010 We consider that our car is an American origin. 02:22.210 --> 02:22.730 Okay. 02:22.930 --> 02:33.130 So we have nine input features cylinders, displacement, horsepower, weight, acceleration, model 02:33.130 --> 02:35.490 year, and the car's origin. 02:35.730 --> 02:40.610 Our target is the mpg, the mile per gallon. 02:40.770 --> 02:43.690 That's what we want to predict. 02:43.730 --> 02:44.610 We have. 02:44.970 --> 02:47.890 The first layer is the hidden layer. 02:48.090 --> 02:51.290 Number of neurons is 64. 02:51.490 --> 02:56.130 This layer learns basic feature combination. 02:56.170 --> 03:05.170 Different neurons specialize in recognizing different patterns like height, weight, plus low horsepower 03:05.290 --> 03:09.970 or many cylinders plus old model year. 03:10.170 --> 03:15.850 Then we're going to move to the second hidden layer, which is number two. 03:15.890 --> 03:22.730 Here we have two architectures 264 two by 64 architecture. 03:23.010 --> 03:32.050 This is a well tested starting point for problems with 10 to 20 input features and hundreds to thousands 03:32.050 --> 03:33.290 of samples. 03:33.530 --> 03:36.050 So why 64 neurons? 03:36.050 --> 03:37.530 It's a sweet spot. 03:37.770 --> 03:43.810 Enough capacity to learn complex relationships, but not so much. 03:43.810 --> 03:48.970 That is easily overfit our 400 sample data set. 03:49.130 --> 03:53.290 So our car prediction problem is moderately complex. 03:53.330 --> 04:00.170 Nine input features with nonlinear relationship between the input features. 04:00.330 --> 04:07.250 Need to learn patterns like, uh, heavier cars with more cylinders equals to lower mpg. 04:07.570 --> 04:08.250 For example. 04:08.330 --> 04:10.450 another, uh, another pattern. 04:10.490 --> 04:15.970 Newer models with same specs equals to better mpg. 04:16.250 --> 04:24.730 So the first layer, if we make our application only containing one layer and 16 neurons, it would 04:24.730 --> 04:31.250 be a it might give us 3.5 Ma, which is too simple. 04:31.450 --> 04:32.850 Did you remember the Ma? 04:33.290 --> 04:40.770 And if we make it with two layers exactly like our application, we have only two hidden layers, each 04:40.810 --> 04:43.050 with 64 neurons. 04:43.290 --> 04:43.890 It will. 04:44.290 --> 04:47.610 It will make, uh, gives us a good balance. 04:47.610 --> 04:55.130 If we make four layers, we have 256 neurons and might get 2.1 Ma. 04:55.170 --> 04:57.850 It's closer to two layers. 04:57.850 --> 05:04.490 So we're going to use two layers for simplicity and making good results. 05:04.650 --> 05:10.730 And by the way this is this is a general like general formula. 05:10.730 --> 05:18.850 If you have 10 to 20 input features and hundreds to thousands of samples, go with two layers. 05:18.890 --> 05:24.130 Go and create two layers, each with 64 neurons. 05:24.330 --> 05:28.650 Then we have the output layer, which gives us the mileage. 05:28.850 --> 05:42.090 And if we put six lenders 200 250 600 hp, 3500 LPs, 15 seconds, acceleration 1976 and an American 05:42.090 --> 05:49.890 origin will get an closer to 22.5 miles per gallon. 05:50.050 --> 05:55.330 And by the way, guys, those are real examples and real data. 05:55.610 --> 06:00.650 In this application we're going to work with real data. 06:00.650 --> 06:09.850 We're going to cover the principles data pre-processing, neural network design, model conversion and 06:09.850 --> 06:15.620 mobile integration Applied to countless other machine learning applications. 06:15.620 --> 06:20.900 So in the previous section, we learned about simple linear regression. 06:21.100 --> 06:30.700 In this application, we're going to learn about very complex and moderate complex applications. 06:30.700 --> 06:36.580 That applies to countless other machine learning applications. 06:36.700 --> 06:38.020 Thanks for watching. 06:38.260 --> 06:48.540 In our next video we're going to start in start exploring and creating our model and adding new features 06:48.740 --> 06:52.580 like the nine features that we specified here. 06:52.900 --> 07:01.180 Then we're going to implement our model into Android Studio and see if our application works fine with 07:01.180 --> 07:05.060 the flight models and test them. 07:05.100 --> 07:09.980 Let's start building our car mileage prediction application.