3.6 ILSVRC (ImageNet Large Scale Visual Recognition Challenge)

Loading...

From the course by Yonsei University

Deep Learning for Business

210 ratings

Yonsei University

Deep Learning for Business

210 ratings

Your smartphone, smartwatch, and automobile (if it is a newer model) have AI (Artificial Intelligence) inside serving you every day. In the near future, more advanced “self-learning” capable DL (Deep Learning) and ML (Machine Learning) technology will be used in almost every aspect of your business and industry. So now is the right time to learn what DL and ML is and how to use it in advantage of your company. This course has three parts, where the first part focuses on DL and ML technology based future business strategy including details on new state-of-the-art products/services and open source DL software, which are the future enablers. The second part focuses on the core technologies of DL and ML systems, which include NN (Neural Network), CNN (Convolutional NN), and RNN (Recurrent NN) systems. The third part focuses on four TensorFlow Playground projects, where experience on designing DL NNs can be gained using an easy and fun yet very powerful application called the TensorFlow Playground. This course was designed to help you build business strategies and enable you to conduct technical planning on new DL and ML services and products.

From the lesson

Deep Learning Computing Systems & Software

The third module “Deep Learning Computing Systems & Software” focuses on the most significant DL (Deep Learning) and ML (Machine Learning) systems and software. Except for the NVIDIA DGX-1, the introduced DL systems and software in this module are not for sale, and therefore, may not seem to be important for business at first glance. But in reality, the companies that created these systems and software are indeed the true leaders of the future DL and ML business era. Therefore, this module introduces the true state-of-the-art level of DL and ML technology. The first lecture introduces the most popular DL open source software TensorFlow, CNTK (Cognitive Toolkit), Keras, Caffe, Theano, and their characteristics. Due to their popularly, strong influence, and diverse capabilities, the following lectures introduce the details of Google TensorFlow and Microsoft CNTK. Next, NVIDIA’s supercomputer DGX-1, that has fully integrated customized DL hardware and software, is introduced. In the following lectures, the most interesting competition of human versus machine is introduced in the Google AlphaGo lecture, and in the ILSVRC (ImageNet Large Scale Visual Recognition Challenge) lecture, the results of competition between cutting edge DL systems is introduced and the winning performance for each year is compared.

3.6 ILSVRC (ImageNet Large Scale Visual Recognition Challenge)8:26

Meet the Instructors

Jong-Moon Chung
Professor, School of Electrical & Electronic Engineering
Director, Communications & Networking Laboratory

In this lecture we're going to talk about the ILSVRC.

The ImageNet Large Scale Visual Recognition Challenge.

Now, this is significant because there are very few places that you can have

these machine learning.

These Deep Learning technologies to compare and compete.

To see really which one is how much better than the other one and

this here is a representative challenge,

the representative competition that is held annually.

So, it's very interesting, it's very rare, so let's take a look into it.

We look into the ImageNet Large Scale Visual Recognition Challenge and,

for short, they call it the ImageNet Challenge.

Even shorter, they just call it ImageNet.

Now, in the ImageNet Challenge, this is an annual contest that started in 2010.

The object category is classification and detection.

Where three main challenges used to benchmark large-scale

object recognition capability.

Where there's one competition that goes for object localization,

where you pick out the top five, so therefore it's called top-5.

Then there's two competitions, two challenges for object detection.

Basically one for pictures, images, and one for motion video.

Number one, is the first one, object localization, also called top-5.

This is the original competition of the ImageNet Challenge.

The other two were added on more recently.

Training data set is 1.2 million images.

The labeled object categories are 1,000 categories.

The test image set is 150,000 photographs.

Each competing program lists its top 5 confident labels based on each test image

in decreasing order of confidence and bounding boxes for each class label.

What are these bounding boxes?

Well, on an image,

when you identify a certain object there's a box that's put around it.

This image, this is what I'm claiming it is.

And that's what the box is talking about.

Here in the top-5 competition this is evaluated based on accuracy

of the program's localization labeling results,

the test image's ground truth labels and object in the bounding boxes.

Where program with the minimum average error rate is the winner.

This is an example of the top-5.

Whereas you can see right here, top-5 selections for each image is listed

with its probability and a histogram is added to it, in this example here.

Then comes a second competition for pictures, for still images.

And this is the object detection, the second competition category.

Each program attempts identification of 200 basic-level categories, and

test image data set has fully annotated labels on each image.

The winner is the program with the highest accuracy in annotated class labels,

confidence scores, and bounding boxes.

The third category of the competition is for

motion video, and its object detection on these videos.

The contender program attempts identification of

30 basic-level categories.

This is among the subset of the 200 basic-level categories

used in the second competition for pictures, for the still images.

Each program will produce a set of annotations of frame number, class labels,

confidence scores, and

bounding boxes on the video clip in real time as the video is run.

The winner is the program with the highest accuracy on the most object categories.

Let's look at some of the participant's program requirements.

Each team is allowed 2 submissions per week,

no regulations on the number of neural network layers.

So, you can go as deep of a neural network as you need or want, and

learning scheme and parameters have to be based only on the training set.

Now in deep neural network technology, one thing to notice is that,

just because you add on more parameters, just because you add on more layers and

layers and layers does not necessarily make it more intelligent, more accurate.

It may just take you more time to train, and

actually your performance may worsen after a certain level,

after a certain number of parameters and certain number of layers that were added.

So therefore it takes a lot of knowledge and

experience to figure out the best, the optimal number of layers and

parameters to use to solve the target problem.

And that's the challenge of the design that is done by these experts,

companies, and expert organizations and

universities that submit their programs to this challenge.

It's truly significant.

And the competition results are also very interesting.

We'll soon take a look at that.

The annual results, let's just look at some of the significant ones that win.

Now, significant improvements in performance have been

seen since Deep Learning has been used in 2012.

Where for the AlexNet, which used over

there in 2012, shown a significant performance improvement.

This is once again the first time that Deep Learning was used.

Here, the 2012 winner AlexNet,

it was named after its first inventor, lead inventor and it was

developed by Geoffrey Hinton's research group at the University of Toronto.

The results show that from the 2011 winner Xerox to

the 2012 AlexNet system that we have here, you can see that

a 9.4% improvement in performance was obtained in the top-5 category.

The object localization.

Now, some of the system characteristics, what was used to make this works so well?

Well, Deep Learning was used for

the first time, a Deep Learning neural network which had 8 layers.

And 5 convolutional layers were used with 3 fully connected layers.

60 million parameters were trained for 6 days and

they used two Nvidia GTX-580s with 3 GB of memory.

Another recognizable thing to notice is the 2014 winner, the GoogleNet.

Which was the Inception-v1.

Here, this resulted in a 93.3 accuracy in

Object Localization in the top-5 localization results.

This is significant because the 90% accuracy level was exceeded for

the first time.

The system characteristics that were used in the GoogleNet is 22 layer

Deep Learning neural network with 5 million parameters that were trained for

1 week on the Google DistBelief cluster.

The 2015 winner was the Microsoft ResNet, and it resulted in a 96.5%

accuracy in the Object Localization, the top-5 competition.

This is significant because

the 94.9% human accuracy level was exceeded for the first time.

The system characteristics that were used was a 152 layer Deep Learning

neural network.

Trained for approximately 3 weeks on 4 NVIDIA Tesla K80 GPUs using

a combined processing capability of 11.3 GFLOPs.

These are the references that I used and I recommend them to you.

Thank you.

Explore our Catalog

Join for free and get personalized recommendations, updates and offers.

Coursera provides universal access to the world’s best education, partnering with top universities and organizations to offer courses online.

© 2018 Coursera Inc. All rights reserved.

Download on the App Store

Get it on Google Play