WEBVTT
Kind: captions
Language: en

00:00:14.580 --> 00:00:16.240
Before deep learning,

00:00:16.240 --> 00:00:21.660
researchers are using handcrafted features for image recognition.

00:00:21.840 --> 00:00:26.680
Now the features are automatically learned by CNN.

00:00:27.360 --> 00:00:29.840
This makes us wondering:

00:00:30.140 --> 00:00:36.040
Can we automate the whole learning process and let the computers to learn the neural network architecture?

00:00:36.900 --> 00:00:44.140
Image that we only need to prepare and label the data, computer will do the rest.

00:00:44.540 --> 00:00:47.080
So everybody can be data scientist!

00:00:48.200 --> 00:00:54.040
This branch of research is called AutoML: Automated Machine Learning

00:00:54.440 --> 00:00:56.940
In terms of AutoML,

00:00:56.940 --> 00:01:00.600
Google has proposed NASNet,

00:01:01.180 --> 00:01:04.700
which can automatically search the best network architecture.

00:01:05.100 --> 00:01:11.320
NASNet has achieved best accuarcy on ImageNet dataset, surpassed all human designed models.

00:01:11.820 --> 00:01:13.820
In 2019,

00:01:14.160 --> 00:01:22.860
Google proposed a new systematic approach to balance the model scaling of network depth, width, and resolution,

00:01:22.860 --> 00:01:25.220
which is called EfficientNet.

00:01:25.520 --> 00:01:28.960
The EfficientNet outperformed previous NASNet,

00:01:28.960 --> 00:01:39.580
while being 8.4x smaller and 6.1x faster on inference than the best existing CNN model.

00:01:40.520 --> 00:01:43.980
AutoML points out a new research direction in the future.