Criticism on Convolutional Neural Networks

I’ve been doing a lot of work with Convolutional Neural Networks, although it’s amateur or their quality is disputable, it is quite easy to catch some drawbacks.

In this article, we’ll analyse why Convolutional Neural Networks are not in the desired positions on the market and we will discuss the technical reasons behind it.

If you think that it is in the right position, this article argues why it hasn’t done better.

Facility Analysis

These drawbacks are all of the concern of the academia, but it is related to deep learning sector and market. Lots of impact of Convolutional Neural Networks, however we are not able to see such a hit on market. It is hype that most of the computer scientists, undergraduate students want to specialize and they have been working for two years.

There are lots of “free” courses that are taught by the grandmasters of machine learning. We see enormous playlists on YouTube, I also create contents for two years, thousands of students in Udemy courses. I’ve never seen such a learning hype, every people can be educated by Andrew NG, Geoffrey Hinton and others.

Udemy, Udacity and the competitors are standing up with neural network trend. Check these websites and see how neural networks courses are proudly presented.

It creates a facility to the communities, there are lots of Slack and Discord groups, they have subscribers of 15000 people. It is like a 15000 capacity class and they are doing interactive project.

Tools are free, promoted by the giant companies, Google, Facebook, Amazon, Sony, each have independent libraries and they share it free. They organize conferences and tutorials. Noone pays for tools. Learners are just paying the Udemy courses that they are impressed, or paying for Coursera certificates.

Statisticians is one of the highest paid job in U.S. market, and big companies purchases small AI startups for millions of dollars. From this two statements, we can not clearly argue that it is well-paid, however these kind of signs show that it is well paid. Actually, no need to give an example for that, companies fund universities, university labs, their own labs. Sponsor lots of events.

What does it show?

Education is free, being educated from the bests is free, lots of communities, lots of funds and sponsorships, crazy investments. What do we have, just Tesla?

As I observe from the Slack and Discord communities it is said that Deep Learning still couldn’t hit the market, couldn’t impact our life in a desired level. There are two main reasons for it.

New generation is enthusiastic

Although Convolutinal Neural Networks show its impact on 2012, Neural Networks trend has been in our life since 2016. It’s just three years old. Time must pass to get a certain level of impact.

Even though these algorithms exist since 1950s, people who expect more is young. 43 years old Andrew NG has lots of 20 years students, has an audience that the minimum age is 13. They are looking for a fast change due to the their ages, it makes the expectation of AI, ML, DL higher.

CNN’s are hungry

I need to indicate that the impact of AI on daily life depends on applications of Convolutional Neural Networks and Reinforcement Learning. You can predict the market share and optimize your advertising strategy with machine learning, but people see the difference of AI with your AI Tesla Car that you plan advertising strategies with machine learning and also use machine learning on the real product, again.

Convolutional Neural Networks are game changer of the machine learning. Tesla is a good example. But one example. I’m not repetitive but the market is.

There are several technical reasons of it.

CNN requires high amount of data and computational resources. Check the GAN work,PGGAN, by Nvidia. 1024×1024 fake human images are created, trained on lots of GPU and it took 4 months. 4 months, it is incredible time. Nvidia do not care about the GPU. Because of that, we need to wait for new methods that decrease the computational costs for GAN. Otherwise, only big companies can do some work with it.

Computational resources are real handicap. To do a CNN project,you need to collect the data, and apply CNN algorithm that has lots of parameters waiting for adjustment. Even one of the problems is solved, CNN products will be more than twice.

For CNN, the problem is same but it is more and more applicable. However, if you have the DATA, it’ll cost to train your models.

Data, hard and expensive to collect. Pricey to label, hard to find. Google and Facebook are able to collect high amount of data, but their superpower is they can collect data that fit for the purpose. They identify some problem, and can collect the data for that.

Interpretability is a big concern. If you’ve some knowledge about softmax, you should know that it is hoped to represent probabilities. We can not see systematic and honest comments on hypotheses such as Bayesian Approach at CNN applications. Thus, Tesla prohibits you to use Tesla Car hands-free, or they can not explain the reasons of the Tesla Car crash.

Considering all the drawbacks, CNN’s will have impact on our life more and more, and we will see it after 4 years. Most of the problems of the CNN will not be a problem. After 7 years, I think that it’ll be part of our life and hype won’t live that much. After 7 years, we’ll discuss more about the smart robots. That’ll be the hype of 2025.

Criticism on Convolutional Neural Networks” için 2 yorum

  • Mart 24, 2019 tarihinde, saat 5:49 pm


    Sanırım makaleleri kendiniz yazıyorsunuz. İngilizcenin yanında Türkçe olarak da yayınlasanız daha faydalı olur diye düşünüyorum.

    • Mart 25, 2019 tarihinde, saat 6:12 am

      Kim okuyor Türkçe’yi hocam malesef Allah aşkına. Vidyosunu çekiyorum zaten ingilizce olanların. Bu daha genel bir blog olacak.


Bir cevap yazın

E-posta hesabınız yayımlanmayacak. Gerekli alanlar * ile işaretlenmişlerdir