Deep learning is subject to large-scale hype, and people can't wait to use neural networks everywhere, but does it work in every place? We'll discuss it in the following sections. After reading it, you'll know the main drawbacks of neural networks, and you'll have a rough guideline when choosing the right type of algorithm for your current machine learning problem. You will also learn about the main issues in machine learning we are facing now. Deep learning has four main reasons for ongoing hype, including data, computing power, algorithms themselves, and marketing. We will discuss each of them in the following sections. Data One factor that increases the popularity of deep learning is the large amount of data available in 2018 that has been collected over the past few years and decades. This allows neural networks to really realize their potential because the more data they get, the better. In contrast, traditional machine learning algorithms will certainly reach a level where more data does not improve its performance. The chart below illustrates this: 2. Computing power Another very important reason is the computing power available now, which allows us to process more data. According to Ray Kurzweil, the leader in artificial intelligence, computing power is multiplied by a constant factor (for example, doubling each year) in each time unit, rather than gradually increasing. This means that computing power is growing exponentially. 3. Algorithm The third factor that raises the popularity of Deep Learning is the advancement of the algorithm itself. The recent breakthroughs in algorithm development are mainly due to making them run faster than before, which makes it possible to use more and more data. 4. Marketing Marketing can also be a very important factor. The neural network has experienced some hype for decades (first proposed in 1944), but in the past it was in an era when no one wanted to believe and invest. The phrase “deep learning†gave it a new fancy name, which made new hype possible, which is why many people mistakenly believe that deep learning is a new creation. In addition, other factors have contributed to the marketing of deep learning. For example, the “humanoid†robot Sophia of Hansen robotics has caused widespread controversy among the public and several breakthroughs in the main fields of machine learning, making it a mass media, etc. . When you should use neural networks or traditional machine learning algorithms, this is a difficult question to answer because it depends a lot on the problem you are trying to solve. This is also due to "there is no free lunch theorem", which roughly indicates that there is no "perfect" machine learning algorithm that excels on any problem. For each problem, a specific method is appropriate and can achieve good results, while another method may fail, but this may be one of the most interesting parts of machine learning. That's why you need to be proficient in several algorithms, and why you can get good machine learning engineers or data scientists by practicing. In this article you will be provided with some guidelines to help you better understand when you should use which type of algorithm. The main advantage of neural networks is that they almost exceed the capabilities of all other machine learning algorithms, but there are some shortcomings that we will discuss and focus on in this article. As I mentioned before, deciding whether or not to use deep learning depends primarily on the problem you are trying to solve. For example, in cancer detection, high performance is critical because the better the performance, the more people can be treated. But also with machine learning problems, traditional algorithms provide more than just satisfactory results. Black box Perhaps the most well-known shortcoming of neural networks is their "black box" nature, which means you don't know how and why neural networks produce certain outputs. For example, when you put a cat's image into a neural network and predict that it is a car, it's hard to understand what caused it to produce this prediction. When you have human-interpretable features, it's much easier to understand the cause of the error. In comparison, algorithms like decision trees are very easy to understand. This is important because in some areas, interpretability is very important. This is why many banks do not use neural networks to predict whether a person is credible because they need to explain to customers why they are not getting a loan. Otherwise, this person may feel threatened by the bank's mistakes because he does not understand why he did not get a loan, which may lead him to change his opinion on the bank, as is the case with sites like Quora. If they decide to delete a user account because of machine learning algorithms, they need to explain to the user why they have completed it. I doubt if they will be satisfied with the answer given by the computer. Driven by machine learning, other scenarios will be important business decisions. Can you imagine that a CEO of a big company will make a multi-million dollar decision without knowing why it should be done? Just because "computer" says he needs to do this. 2. Development duration Although libraries like Keras make the development of neural networks very simple, sometimes you need more control over the details of the algorithm, for example, when you are trying to solve problems in machine learning. Then you might use Tensorflow, which gives you more opportunities, but because it's more complicated, development takes longer (depending on what you want to build). So for the company's management, if it's really worthwhile for their expensive engineers to spend weeks developing something, then the problem will arise, and the problem can be solved faster with simpler algorithms. 3. The amount of data Compared to traditional machine learning algorithms, neural networks usually require more data, at least thousands or even millions of labeled samples. This is not an easy problem to solve. Many other machine learning problems can be solved with less data if other algorithms are used. Although in some cases neural networks rarely process data, in most cases they do not. In this case, a simple algorithm like Naive Bayes can handle a small amount of data well. 4. Calculating expensive In general, neural networks are computationally more expensive than traditional algorithms. State-of-the-art deep learning algorithms that enable successful training of truly deep neural networks can take weeks to fully start training from scratch. Most traditional machine learning algorithms take less than a few minutes to a few hours or days. The computing power required for a neural network depends largely on the size of the data, but it also depends on the depth and complexity of the network. For example, a neural network with one layer and 50 neurons will be much faster than a random forest with 1000 trees. In contrast, a neural network with 50 layers will be much slower than a random forest with only 10 trees. Now you may know that neural networks are more suitable for certain tasks, but not necessarily for others. You learned that a lot of data, more computing power, better algorithms and smart marketing have increased the popularity of deep learning and made it one of the hottest areas. Most importantly, you have learned that neural networks can beat almost all other machine learning algorithms and the accompanying shortcomings. The biggest drawback is their "black box" nature, increased development time (depending on your problem), the amount of data required, and most of their computational cost. Deep learning may still be a bit over-hyped and exceed what is expected to be done. But that doesn't mean it's useless. I think we live in the renaissance of machine learning because it is becoming more and more democratized, and more and more people can use it to build useful products. Machine learning can solve many problems, and I believe this will happen in the next few years. One of the main problems is that only a few people understand what they can do with it and know how to build a successful data science team that brings real value to the company. On the one hand, we have doctoral engineers who are the theoretical genius behind machine learning, but may lack a commercial understanding. On the other hand, we have CEOs and management positions, they don't know what deep learning can do, and think it will solve all problems in the next few years. We need more people to fill this gap, which will produce more products that are useful to our society. Lightweight, Reusable Goggles, Indirect Vented (Splash proof)
Screened top and bottom ports circulate air
Feature: Eye Goggle,Safety Goggles Price,Eye Surgery Goggles,Medical Safety Goggles Guangzhou HangDeng Tech Co. Ltd , http://www.hangdengtech.com
Clear Vision with a Wide Flange and Latex-free
Vents reduce fogging for extended wear.
Fits over most prescription glasses.
1.Material:The parts the wearer touches are not made of materials that can cause skin irritation
2.Structure:Smooth surface, no burr, no acute Angle and other defects that may cause eye and face discomfort.
It has good air permeability.
Adjustable parts and structural parts are easy to adjust and replace.
3.Package:The products are properly packaged and are accompanied by product certificates and instructions
4.fixing band : The minimum width of the headband in contact with the wearer is 11.5mm. The headband is adjustable, soft and durable
5.The appearance quality of the lens: the surface of the lens is smooth and free from scratches, ripples, bubbles, impurities, and other obvious defects that may impair vision
6.Diopter: 0.04D
7.The difference between the prism degrees of the left and right eye lenses: 0.12
8.Visible light transmission ratio: colorless transparent lens 89.35
9.Impact resistance: qualified
10.Protection performance of chemical fog drops: there is no color spot on the test paper within the center of the lens
11.Irritant gas protection performance: there is no color spot on the test paper within the center of the lens