Click the above
“Gu Mu Pai Li Mo Chou” Follow
*Friendly reminder:If you want to part-time join various high-tech startup teams, please add Li Mo Chou’s WeChat number: 13575329911 Say JoinTeam.
Li Mo Chou said
Today, Mo Chou wants to talk to everyone about deep learning, which is a great thing. It is a major breakthrough in the field of machine learning in recent years, with very broad application prospects. With Google unveiling the Google Brain project, the heroes of the scientific community have become enamored with deep learning.
Mo Chou gives an example: the power of AlphaGo lies in the fact that it replaced two of its components with CNNs, and with CNNs, it gained two benefits: 1) it has powerful nonlinear mapping capabilities, and the moves calculated are impressive; 2) the deep learning capabilities of CNNs mean that playing Go no longer requires the same extensive study as in the past. One move with CNN can solve countless problems across various fields. Once upon a time, when Mo Chou was writing evaluation functions for games, one had to have a deep understanding of a particular game, essentially needing to be one of the best amateurs.
Deep learning has made groundbreaking progress in fields such as speech recognition and image recognition. If you want to impress your girlfriend with deep learning: the most useful thing so far is image processing, which can be used to make early judgments about cancer or to find Li Mo Chou in a crowd. However, if Mo Chou asks if you can write a small program to impress your girlfriend, you might not be able to believe it. This technique is called artistic style transfer, where with just a click, you can transform your girlfriend's portrait into a postmodern Picasso artwork, thus successfully turning your girlfriend into a dog or a wife. Artistic style transfer is an ancient yet modern theme; many artists have bent over backward to depict the works of others.
Today, Mo Chou shares this idea with you, hoping everyone can go home and cuddle with their girlfriends.
The First Level: Image Recognition
If you start to understand image processing with deep learning, the first task you encounter will definitely be image recognition:
For example, input your beloved cat into a regular CNN network and see if it’s a cat or a dog.
A very ordinary CNN, like this few-layer CNN ancestor Lenet. If you have a good dataset (like Kaggle’s cat vs. dog competition), you can get a fairly reliable classification result (over 80% accuracy), although it’s not too high.
Of course, if you add some knowledge about specific problems, you can also identify personal faces and start a startup called face something:
Those who play like Mo Chou can also identify pig faces (Mo Chou thinks they all look the same), thus identifying each pig’s identity and then selling high-quality pork, making it hard not to profit.
I don’t know my identity; I only know I’m a pigOr check what diseases different plants have, like these different lesions that people are too lazy to look at; it can show you. People in plant protection can take their phones to the fields.
Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. “U-net: Convolutional networks for biomedical image segmentation.” International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2015.
Although plant protection is really useful, classification can get quite boring after a while. The direction we evolve is to use more advanced network structures to achieve better accuracy, like the residual networks shown below (which can already achieve over 99.5% accuracy on the cat vs. dog dataset). Once you master classification, you’ll feel like a deep learning master, with a hammer and nails in your eyes.
He, Kaiming, et al. “Identity mappings in deep residual networks.” European
Conference on Computer Vision. Springer International Publishing, 2016.
The Second Level: Object Detection
Pretty soon, you realize that classification skills are not very useful in most real-life situations. Because real tasks often look like this:
Or like this:
With so many things together, your classification network trained with cat and dog portraits can easily get confused. Even if there’s a cat and a dog in one image, adding some noise to the cat can completely throw your classification network off.
In reality, there aren’t that many images; a picture usually contains one big cat or beautiful woman; more often, the items in an image are chaotic, and you need to draw a box around the target you want to see, then check what those things are.
Thus, you arrive at the next challenge in machine vision – object detection (bounding the target object in a large image and recognizing it), which brings about a new network architecture known as R-CNN. This network can not only tell you the classification but also the coordinates of the target object, even identifying multiple target objects in one image. In practical applications, this is very powerful, as it can be paired with city cameras for “Heavenly Eyes.”
Ren, Shaoqing, et al. “Faster R-CNN: Towards real-time object detection with region proposal networks.” Advances in neural information processing systems. 2015.
Identifying suspects among so many onlookers is a piece of cake; security personnel can’t help but get excited.
The newly emerging YOLO algorithm achieves fast real-time object detection. It’s as if Mo Chou is walking down the street, and it immediately tells her where all the people and objects in her field of view are, which is a godsend for autonomous driving.
YOLO fast detection method Redmon, Joseph, et al. “You only look once: Unified, real-time object detection.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016.
Of course, even here, you will still find it boring. Even if the network can be quite complex, it’s merely a CNN network (region proposal) with another layer of CNN for classification and regression. Can we do something different?
Firstly, Mo Chou will briefly talk about deep learning, which is a framework of machine learning and one of the foundational technologies of artificial intelligence. Its concept can be traced back to control theory in the last century, later evolving into connectionism, and by 2006, it developed from artificial neural networks into the currently popular deep learning.
There are many types of neural networks in deep learning, and today we mainly discuss CNNs and GANs mentioned in the article.
Convolutional Neural Networks (Mo Chou: this is CNN, also known as DCNN) primarily deal with image data but can also handle voice data. It has a scanning layer, and this form of processing does not attempt to analyze the entire training data at the beginning. Mo Chou gives an analogy: for an image of size 200X200 pixels, you wouldn’t want to build a layer of 40,000 nodes. Instead, you would construct a 20X20 pixel input scanning layer and input the first 20X20 pixel image from the original image (usually starting from the top left). Once this part of the image is processed (which may be used for training the convolutional neural network), you would then process the next 20X20 pixel image, gradually (usually moving one pixel at a time, but the step size can be set) moving the scanning layer to process the original data.
Generative Adversarial Networks (this is GAN) consist of a pair of “twins” that can be made up of any two networks working together (but usually FF and CNN), where one is used to generate content and the other to discriminate the generated content. The discriminating network receives both the training data and the data generated by the generative network. The accuracy of the discriminating network is used as part of the error for the generative network. This creates a competition: the discriminating network becomes increasingly skilled at distinguishing real data from generated data, while the generative network becomes better at producing unpredictable data, both refusing to back down, fighting fiercely. This method is very effective, partly because even quite complex class noise patterns are ultimately predictable, but generated data with features very similar to the input data is hard to distinguish.
The Third Level: Image Segmentation
Aha, this brings us to the third level, where you not only need to detect the objects in the edges and corners of the image, but you also have to perform the daunting task of cutting them out from the image. You should know that infants can’t distinguish object boundaries; for example, if there’s an apple on Mo Chou’s table, what is the table, and what is the apple? Why is the apple not sitting on the table? So, whether the network can cut objects out of an image represents whether it can mimic the essence of human vision, which is also a small test. Simplifying this problem, we are merely generating a “mask” of the original image, similar to the mask in Photoshop.
This is cutting out an image
Drozdzal, Michal, et al. “The importance of skip connections in biomedical image segmentation.” International Workshop on Large-Scale Annotation of Biomedical Data and Expert Label Synthesis. Springer International Publishing, 2016.
Note that in this task, we need to obtain another image from one image, which is the “mask.” At this point, the so-called U-Net comes into play, and this is our first generative model. Its components are still convolutional, but it adds the reverse process of max pooling for up-sampling.
This segmentation task should not be underestimated, especially for friends in the scientific community. For example, with the rise of private satellites and drones, if you want to check the terrain around your neighborhood to see if there’s a hidden vault, input the information, and the satellite image will display it clearly. Where there are trees, where there is water, where there are military bases, no human is needed; it will all be cut out for you.
If you want to count cells, that’s quite easy; just convert them into such outlines.
In the field of medicine, it’s not just about counting cells; deep learning has long been a part of medical imaging applications. As we all know, medical imaging technology plays an important role in early diagnosis, staging, and efficacy evaluation of diseases and has become an indispensable clinical technology in disease diagnosis and efficacy evaluation. Currently, deep learning is widely used in medical image segmentation, image registration, computer-aided diagnosis, etc., especially in intelligent decision-making. Although deep learning still requires a lot of work in optimizing image biomarkers and improving predictive models, Mo Chou believes that deep learning will accelerate and promote the development of medical image analysis technology and its application in clinical practice.
The Fourth Level:
We start to get fashionable. If you are a small clothing store owner on Taobao and want customers to input a picture of clothing and then get a set of recommended clothing, how do you implement a reverse image search function? Note that I can crawl a large number of images from the internet, but these data are unlabelled. What should we do? Mo Chou tells you there’s a way, and that method is clustering.
Mo Chou teaches you the simplest clustering method, which is to put all images into a convolutional network, but we don’t extract classification; we only extract some features from intermediate layers of the network. These features are like a visual QR code for each image, and then we perform k-means clustering on these QR codes, which will yield unexpected results. Why deep? Because deeply extracted features are unique.
Then, reverse image searching is simply finding other images in the same cluster, and the rest is mindless operation.
Based on clustering, you can perform a search.
The Fifth Level:
We begin to aspire to the stars. The previous classification applications for making money are too boring. Can machine vision engage in science? Mo Chou tells you, absolutely, as someone who observes cells while gazing at the stars, I often find that the noise in the astronomical or cellular images I obtain is unbearable. Then, deep learning provides you with a set of denoising and image restoration methods. A tool called auto-encoder plays a significant role, and just like that, the image becomes clear.
This is not the most powerful; the application of adversarial learning using game theory can also help you eliminate noise. If you understand adversarial networks, GANs are also a tool for image generation, allowing the network to remove noisy images, and it can’t even recognize the natural image without noise, not even its “mother.”
Schawinski, Kevin, et al. “Generative adversarial networks recover features in astrophysical images of galaxies beyond the deconvolution limit.” Monthly Notices of the Royal Astronomical Society: Letters 467.1 (2017): L110-L114.
The Sixth Level:
Having made enough money in the industry, science is too nerdy, Mo Chou comes to play with artistic thinking and philosophy. The first move is image style transfer.
This is the cat that Mo Chou raises; this cat has very strong genes
No matter how you combine it, I am still a dog
You might think he is a dog, but he could also be your girlfriend
Those who can truly play with this are still the adversarial learning GANs mentioned earlier, such as the renowned CycleGAN, which can almost achieve a custom “image translation” function without requiring annotations. Just take two sets of images from winter and summer, and it will automatically find the correspondence between the two sets.
Zhu, Jun-Yan, et al. “Unpaired image-to-image translation using cycle-consistent adversarial networks.” arXiv preprint arXiv:1703.10593 (2017).
Operations in the art field are indeed very interesting. Mo Chou spent two years watching two million unlabeled videos, forming concepts about objects and backgrounds. Input a screenshot from a video and predict how it will move later… this is also a rhythm that can scare you awake at night.
Teaching machines to draw simple sketches can easily roll through this process using Jupyter Notebook, Mo Chou: “Help me draw a sheep.”
This understanding and learning process is very interesting, and the results achieved can take many forms.
The following should be the most observed by everyone, Mo Chou thinks her drawing is truly vivid and expressive.
The Seventh Level:
Having lost interest in image translation, neural networks claim to understand images, so let’s generate an image from noise and see.
Still using GANs, and the most basic convolutional GAN (DCGAN) can do this for you.
Take a look at the hotel scene imagined by GAN; can you believe it was created by a computer?
Goodfellow, Ian, et al. “Generative adversarial nets.” Advances in neural information processing systems. 2014.
As I write this, Mo Chou feels that GANs are very promising, promising, promising. Previously, Mo Chou thought it was just fun; she didn’t expect this thing to be so powerful.
Today, the seven-tier pagoda shared by Mo Chou is just the tip of the iceberg of deep learning discovered by humanity,Mo Chou gives everyone a slightly clearer outline:
Mo Chou also tells everyone an interesting artificial intelligence medical conference:
Event Announcement
“Artificial Intelligence and Medicine Xixi Forum” will be held on January 27, 2018, in Hangzhou.This conference is co-hosted by the Zhejiang Overseas High-level Talent Association (Biomedicine Branch) and the Shao Yifu Hospital affiliated with Zhejiang University, with Roche Pharmaceuticals, Alpha Cat (Hangzhou) Artificial Intelligence Biotechnology Co., Ltd., and Haining Huaitai Kangming Biotechnology Co., Ltd. as co-organizers.
The conference will focus on the theme of “Artificial Intelligence and Precision Medicine,”drawing on the TED conference’s presentation model, and will invite many technical leaders and “Thousand Talents Plan” experts as well as clinical doctors to discuss the application of AI in medicine, share new research results, clinical needs, and related projects. If you wish to attend this conference, please search and add the WeChat number 13867470693, and be sure to note “AI1000” while scanning the QR code below to fill out the registration information.
Science is magic, as Arthur C. Clarke said, any sufficiently advanced technology is indistinguishable from magic.
The distance between science and industry is getting shorter and shorter. For example, after AlphaGo played Go, it began to venture into the medical health field. The DeepMind team recently launched a new research collaboration plan to use artificial intelligence in partnership with medical institutions to combat breast cancer.
Biomedicine has entered the “big data era,” and the application of artificial intelligence in medicine is an inevitable trend.
The significance and value of artificial intelligence for biomedicine are: 1) AI can greatly enhance human ability to conquer diseases; 2) AI can significantly enhance human capacity for researching the essence of life; 3) AI promotes the development of the pharmaceutical and medical device industries, such as discovering new drug targets and designing new drugs, significantly speeding up and improving the accuracy of big data analysis in clinical trials.
*This conference’s official release platform is “Thousand Talents Intelligence Valley”
Time
January 26-28, 2018
Location
Hangzhou Xixi Hotel
Theme
Artificial Intelligence and Medicine
Agenda
January 26, 2018 (Friday)
Before 17:30
Participants arrive and check in at the hotel
18:30-20:00
Dinner
January 27 Morning (Saturday)
8:00-12:00
Artificial Intelligence and Precision Medicine
Sub-forum One: Applications of Artificial Intelligence in Medicine
Sub-forum Two: “For the Stomach” Jiangsu, Zhejiang, and Shanghai Young and Middle-aged Gastrointestinal Tumors
12:00-13:00
Lunch
15:00-17:30
Visit and Inspection
18:00-19:30
January 28, 2018 (Sunday)
9:00-12:00
Theme Reports and Free Discussions
12:00-14:00
Lunch
14:00-17:30
International Summit Forum on Artificial Intelligence and Medicine, New Drug Development
Forum One: Artificial Intelligence and Traditional Chinese Medicine
Forum Two: Artificial Intelligence and Medicine, Translational Medicine
Forum Three: Artificial Intelligence and Drug Development
Forum Four: Artificial Intelligence and Medical Devices
Forum Five: Artificial Intelligence and Investment
*For conference registration-related questions: please see the confirmation letter sent to your email;
*Except for invited guests, participants are responsible for their own transportation and accommodation costs. If you need help booking hotels for the conference, please specify in the confirmation letter.
Organizers
Biomedicine Branch of Zhejiang Overseas High-level Talent Association
Shao Yifu Hospital Affiliated to Zhejiang University
Contractors
Alpha Cat (Hangzhou) Artificial Intelligence Biotechnology Co., Ltd.
Haining Huaitai Kangming Biotechnology Co., Ltd.
Co-organizers
This conference aims to further strengthen the interaction and collaborative innovation between technical personnel in cutting-edge technology and doctors, promote interdisciplinary development, assist in the transformation of results, learn from and draw on the innovation and entrepreneurship model starting from “clinical needs” in Silicon Valley and Israel, seeking more possibilities for applying relevant technologies in the medical field to save more lives and promote the continuous progress of collaborative transformation in Chinese scientific research and clinical medicine.
Clinical doctors understand the needs on the front lines, while the advantages of “Thousand Talents Plan” experts lie in technology and engineering.A series of innovative entrepreneurship models in medicine indicate that when clinical needs are realized by technical personnel, only one more manager is needed to establish a company with a three-person model to achieve a startup product, rapidly advancing within a few years. This entrepreneurial model has been proven to be highly efficient in Israel.
Make friends with AI, let’s develop together
Surrounding environment of the venue
*This is the second round of meeting notifications; for specific agendas of the meeting, please pay attention to subsequent content from this platform.
Mo Chou’s dinner party, welcome to invite Mo Chou for dinner.