Yunqi number information: [ Click for more Industry News ]
where you can find different industries first-hand information on the cloud, what are you waiting for, come!
Recently, a doctoral candidate artificial intelligence research in Australia published an article on the SARS-CoV-2 virus on LinkedIn. Since the very topic of conversation and called up to 97.5% accuracy rate, this article will soon be tens of thousands of comments, thumbs up and forward. However, such a model was clawed back only took 50 pictures of training.
Exactly one week to put up 97% of the model, the truth is?
Earlier, an artificial intelligence doctoral candidate in Australia announced the construction of a set of deep learning model can accurately rate of 97.5% to detect whether a patient infected with the virus from the lungs COVID-19 X-ray film. Because the spread of the epidemic and the lack of foreign medical facilities, so people are very concerned about this achievement, just a time to reap tens of thousands of comments, praise and point forward, it also created a working group Slack, get a lot of praise.
Message from the current release, the project has the following characteristics:
- A trained model PyTorch
- The container application code that
- A GitHub repository, and translated into several languages
- Web application being developed
- Mobile applications being developed
- Blueprint, no intention to use this model in the server infrastructure hosted in AWS
- In marketing and sponsorship as well as a large number of follow-up plan
And above all, we have quickly completed within a week. Then, there are several serious problems Reddit users Bachu this solution, and this was sort of rebuttal.
It only took 50 images training?
First, the potential of these neural networks were very complex, and therefore necessarily require a lot of training samples to complete the model training. But as of the time of submission, this COVID-19 detection tools have only seen 50 lung imaging.
For such a network includes 150 layers, more than 20 million argument is so limited training sample set is obviously extremely absurd.
Data sample in question
In addition, there may be a huge data sample bias, which does not include the 50 pictures are related persons infected with the virus, but only make a mark in accordance with the operating COVID-19 lung caused by acute cases. Unless the lungs have been damaged by the virus, otherwise the model is unable to detect the signs of infection. In addition, even if you have symptoms of pneumonia, if not already belong to acute symptoms, still unable to prove the accuracy of this model.
Image repeat, confusing code, the model in question
Finally, this model is based on high popularity COVID reference network ResNet-50. Although the latter does belong to image identification and classification scheme commonly used in the art, but generally pre-trained ResNet cover only object in everyday environments. In other words, the ResNet hidden layers in the network geometry better at identifying the color image, the X-ray images, we obviously can not find such a model. Precisely because of this, most medical neural network was built from scratch in only choose the way of development.
Further observation of this code base, we also found a lot of other problems. Training, validation and test data set contains duplicate images, most of the training process directly copy PyTorch tutorial, mixed with a lot of unnecessary code; Github issues have been entirely unable to understand ......
GitHub Address: https://github.com/ elcronos / COVID-19
Project leader responded: I said item is not available
Initially, when individual developers to communicate with the project leader and questioned the other responded:
xxx, hello, we've got the results of Canadian research institutions xxx radiologists support and recognition
However, as more and more voices of doubt, the project leader updated the introduction of GitHub, and said:
Although the results of the project "looks promising", but I clearly pointed out that the model is far from available, it should not be used for diagnosis or any medical decisions. This is a work in progress, we need to have the relevant skills to help people. I'm still GitHub repository pointed out, I'm looking for help developers to improve and collect better data set.
...
Unfortunately, this project attracted the attention of experts, and they did not pay attention model is not ready, and the need for better data collection and to help create a better model, did not read all our disclaimer. The project has accused misleading, even the suggestion that I do have commercial intent. This has resulted in some negative impact on my personal life, so I decided to take a step back, to temporarily withdraw from social media. At least in the next few days, I will not be active in this group.
Full version declarative referential: https://github.com/elcronos/COVID-19
However, the official had also trumpeted the project and initiate fund-raising. The project leader creates a Slack contain multiple sub-channels of discussion groups, which have a #marketing channel dedicated to communication and financing. In addition, # sponsors channel is responsible for communication with potential investors, to report to the future prospects of return on investment.
Slack discussion group: https://app.slack.com/client/T010AJ5H31N/learning-slack
In addition, the number of useful content channel called #datascientists's not, which is full of full of enthusiasm but little experience of the novice. The same, # doctors channel situation is similar, the only valuable content is opposition from health care professionals, for example, does not recommend the use of chest X-ray diagnostic COVID-19 infection. The last sub-channel #researchers is almost nobody.
On the other hand, UI / UX output channel content is actually quite rich. The plan now has 5 different logos, plus a dedicated interface for mobile and Web applications.
Thus, for the statement that most developers do not buy it, many people think that in the current exceptional circumstances, serious problems such items should not be published and widely publicized (and even ridicule developer workload road publicity It is probably the development of 20 times).
Medical diagnosis of deep learning
Convolution depth network in terms of diagnosis and treatment of disease does have a range of potential advantages. In recent years, numerous scientific publications published are highly concerned about this new direction:
in 2016, a group of researchers from London published a new method to contain the 80000 fundus photographs dataset based can 86% of patients diagnosed with accuracy due to diabetes-induced retinopathy.
That same year, researchers from Uganda using the data set contains 10,000 objects, assessed the convolutional neural network (CNN) capacity to analyze microscopic blood smears.
Two Japanese researchers included 550,000 cases a CT scan image data sets, lung nodules vast scale of a sorting operation.
But the new crown virus previously mentioned detection is completely different, a little browsing code base of its publication, we can see the depth of a serious shortage of cognitive learning and AI technology. Even worse is that many developers are questioning its obviously trying to take advantage of this outbreak on their own to promote.
Say good code to change the world?
Deep learning is by no means one hundred test Braun solutions. In recent years, many companies are not ready to rush to establish data within the team, only to find the cost of the rapid increase at the same time not get any meaningful output.
Earlier, Li Feifei in interviewed had mentioned:
the bubble does exist, over-exaggeration, hype can say overwhelming. As a scientist, I hope these bubbles are dissipated as soon as possible. Only a solid core of attention in order to promote the progress of AI and bring real benefits, which is particularly important in the field of health care and medicine.
In addition, we should not use technology to produce injustice, prejudice or expand pre-existing inequalities. For AI technology, I want to reduce it as much as possible contact threshold, increase fairness and relieve all sorts of related conflicts. If handled properly, we have the opportunity to create a better future use of AI techniques. Of course, the premise is that we have to carefully sort out the existing AI achievements, figure out what is fabricated, which is true.
[Yunqi number of products online classes] have technical experts to share every day!
Courses Address: https://yqh.aliyun.com/zhiboJoin the community, with experts face to face, to keep abreast of the latest developments course!
[Yunqi number online classroom community] https://c.tb.cn/F3.Z8gvnK
Original link
This article Yunqi community original content may not be reproduced without permission.
