Some fine resources about the issue of deep learning and its future are attached.

https://blog.keras.io/the-future-of-deep-learning.html

https://www.forbes.com/sites/quora/2017/07/27/how-will-deep-learning-algorithms-change-in-the-future/#70c2b48d3adf

https://www.quora.com/What-does-the-future-of-deep-learning-look-like-1

https://news.ycombinator.com/item?id=14797364

Some fine resources about the issue of deep learning and its future are attached.

https://blog.keras.io/the-future-of-deep-learning.html

https://www.forbes.com/sites/quora/2017/07/27/how-will-deep-learning-algorithms-change-in-the-future/#70c2b48d3adf

https://www.quora.com/What-does-the-future-of-deep-learning-look-like-1

https://news.ycombinator.com/item?id=14797364

How deep is this deep? In the initial days of BPN, I remember to to have read that a three layer BPN would serve for universal approximation. At one place we talk about green computing, energy conserving computing etc.. my anguish is about energy consumption by a deep neural network. There is no denigration, but deep learning networks should be expeditiously applied where ever, they are a must.

Hi M. J. Ebadi,

In my humble opinion, I see the tendency of deep learning in two ways. First, the huge advances in computational resources (CPU and GPU) allow the use of more complex and deeper models. The upper limit of this growth is difficult to predict. Second, as a researcher in machine learning I'm feeling a tendency towards the way the human learns. For example, currently there is a lot of research focused on Deep Reinforcement Learning, which integrates deep neural networks, such as CNNs, with reinforcement learning theory. These models have proven to be very effective (in some cases more than a human user) in several fields, for example playing video games (https://deepmind.com/research/dqn/).

Hope this insight helps.

Kind regards.

Carlos.

With facial recognition

(1) unsupervised learning features with deep models and then using discriminative methods

(2) directly using class labels as supervision of deep models

Source - Deep Learning Identity-Preserving Face Space

https://www.cv-foundation.org/openaccess/content_iccv_2013/papers/Zhu_Deep_Learning_Identity-Preserving_2013_ICCV_paper.pdf

Deep learning is one stream of exponential knowledge automation; the hidden layers of learning processes play a decisive role. The progress with images will be faster than that with language; there are still a lot of heuristic enigmas to be solved.

Causes of shame Education and genetics on isolation and treatment Education since childhood on honesty and confrontation Greetings Best Regards

Causes of shame Education and genetics on isolation and treatment Education since childhood on honesty and confrontation Greetings Best Regards wessam

Causes of shame Education and genetics on isolation and treatment Education since childhood on honesty and confrontation Greetings Best Regards wessam

Actually, Deep Learning is just complex architectures of neural networks. Sure, they have beaten many limits about precision during classification task but It is not possible to understand them and prove they are really learning. Then, DL could reach some limit and maybe it will be happen how neural networks (cold time) until people could understand them.

Dear Carlos Sampedro Pérez,

Thank you very much for your valuable answer.

Regards,

Javad.

Dear Ljubomir Jacić,

Thank you very much for your proposed resources.

Regards,

Javad.

Well, every other day I get to see the application of deep learning in different areas of science. As a machine learning student, I am myself interested and quite fascinated to learn and implement deep learning in my area of study i.e., Computational Molecular Biology. Since long the focus of machine learning is to learn more like human and deep learning is making it possible more than ever before.

Dear Romasa Qasim,

Thank you very much for sharing your opinion.

Regards,

Javad.

Dear Stephen I. Ternyik,

Thank you very much for your valuable answer.

Regards,

Javad.

Dear Krishnan Umachandran,

Thank you very much for sharing your opinion.

Regards,

Javad.

Dear Jayaram M.A,

Thank you very much for sharing your expertise and opinion on this issue.

Regards,

Javad.

In few words, I see the Deep Learning as a way to promote the future.

Dear Mohammed Chalouli,

Great!! Just with few words you draw up the future of Deep Learning. Thanks.

Regards,

Javad.

You are right dear Dr. Brenda Jacono , I do agree with your point of view.

Emad

I like this "I see the Deep Learning as a way to promote the future.". I agree with all RG Colleagues.

Best regards

Hi M. J. Ebadi,

If you look at the performance gap deep learning (DL) has created between itself and some other algorithms, it is massive. This is not to say other algorithms are now obsolete, no free lunch theorem remember, but DL has proven itself, empirically at least. If you check image classification/recognition and object detection algorithms and natural language processing (NLP), DL tops most algorithms and mostly by huge gaps. In some data sets it even rivals humans, that is why it is safe to say that most experts now are flocking to DL.

But it's not like when faced with any machine learning (ML) problem all you have to do is throw DL at it and drink coffee, no. Sometimes DL can be an overkill especially in smaller problems with little data availability and in some unsupervised learning problems, clustering still works. In most complex cases it is better to have a variety of algorithms or ensemble methods like in AlphaGo it was shown that tree search methods are pretty much alive and kicking. The Monte-Carlo tree search (MCTS) + DL were used in the AlphaGo system to quickly search for winning moves.

Now the most attractive thing about DL is that it is motivated by the way the world works, the world is compositional, atoms form molecules and molecules form all sorts of things and all sorts of things form all sorts of other things and so on.

In literature we have:

letters->words->sentences->paragraph

Just as in vision we have:

pixels->edges->parts->objects

You will find that most algorithms thought to have nothing to do with DL are actually a form of DL. Such as the deformable parts model (DPM), is actually a special form of convolutional neural networks (convNet) with a distance transform based pooling mechanism. So a lot of algorithms can be cast in this hierarchical compositional manner. We even do this when coding, we code low-level functions, then mid-level and then high-level functions.

This is the whole idea about DL, learning increasingly more abstract representations at every layer. This idea is somehow future proof, so even if it is difficult to say with high certainty, this compositional nature of DL algorithms is going to standard the test of time. It could be that maybe having 100+ layers won't be necessary in the future or some shallower wider networks may start to perform as well as deeper thinner networks as more better learning algorithms are discovered.

So what's the future of DL? Well it is certainly clear that DL is heavily supervised and requires soft differentiable objectives. So even if the concept of hierarchical layer-wise learning is less likely to change, the actual learning methods will shift from this desire to cast everything in a simplified differentiable manner.

Hope this helps

Source:

https://www.quora.com/What-is-the-future-of-deep-learning-Are-most-machine-learning-experts-turning-to-deep-learning

Concerning me I did not experience this type of learning but it seems that it is promising. Please refer to the following link for more information:-

https://en.wikipedia.org/wiki/Deep_learning#Deep_neural_networks

Besides the hype that surrounds Deep Learning and the over optimistic futurist views on the field, Deep learning will go through the same phases that each technological advances had witnessed. Theoretically, DL needs much work to fully grasp and prove (analytically) its convergence; on the other hand i.e. empirically it was demonstrated since the introduction of Convolutional Neural Networks (in parallel with the rise of connectionism in the 80s with backprop) that representation learning and wider hierarchical nets yield superior performances but on the cost of very large data and high computational power, now on the future of the field, DL has yet to prove its usefulness in areas where data are highly non-stationary and in more challenging tasks of unsupervised learning, Hinton has recently proposed to surpass backprop algorithm.

Why the future of deep learning depends on finding good data!

The future of deep learning may be to work toward unsupervised learning techniques. If we think about teaching babies and infants about the world, this makes sense; after all, while we do teach our children plenty, much of the most important learning we do as humans is experiential, ad hoc — unsupervised...

https://techcrunch.com/2017/07/21/why-the-future-of-deep-learning-depends-on-finding-good-data/

Dear Md. Samiul Haque Sunny,

Great!!! Thank you very much for your valuable and complete answer.

Regards,

Javad.

Dear Nabeel Hameed Al-Saat,

Thank you very much for sharing a useful link.

Regards,

Javad.

Dear Okba Bekhelifi,

Thank you very much for your useful answer.

Regards,

Javad.

Dear Ljubomir Jacić,

Thank you very much for your interesting answer.

Regards,

Javad.

Dear @Ebadi, this is fine reading on deep learning.

Google Brain chief: Deep learning takes at least 100,000 examples!

While the current class of deep learning techniques is helping fuel the AI wave, one of the frequently cited drawbacks is that they require a lot of data to work. But how much is enough data? ...

There are still plenty of hurdles that humans need to tackle before they can take the data they have and turn it into machine intelligence. In order to be useful for machine learning, data needs to be processed, which can take time and require (at least at first) significant human intervention.

“There’s a lot of work in machine learning systems that is not actually machine learning,” Dean said. “And so you still have to do a lot of that. You have to get the data together, maybe you have to have humans label examples, and then you have to write some data processing pipeline to produce the dataset that you will then do machine learning on.”...

https://venturebeat.com/2017/10/23/google-brain-chief-says-100000-examples-is-enough-data-for-deep-learning/

I think there will be no limit of development... at least if the quantum computer will be developed (although I am doubtful concerning quantum computers).

Best Regards

Andreas

A step toward artificial intelligence.

Deep Learning would penetrate into most of the soft computing tool based research work in near future!

No deep learning. Because by God Grace, life going smoothly with out planning.

I support the brilliant views of Prof. Carlos.

I will echo some of the previous posts in that this is just a phase of neural networks and covered with hype. If we go back to the promises of single layered networks, it came to a halt due to its inability to process the XOR problem. Then came the hybernation problem. Then came the Hopfield network, the back propagation NN and now the stacked restricted Boltzmann machines and convolutional NN's as Deep learning. This is just another phase for the following reasons:

1) While you can peer into what prototype representation is acquired by Stacked restricted Boltzmann machines you could approximate what is being learned but not directly obtain what is the generalization rule. This will be a limitation for adoption on which there is liability involved to the user of the system (nobody wants to be blamed for another "person's" mistake phenomenon .

2) While promises have been made into the miracles of NN's they are not magic, so junk-in junk-out. It will not solve that issue.

3) They are just pattern recognition algorithms. To evolve into more complex system will require hierarchies of systems. an area that is too complex for most thesis, therefore private sector will have to step in and revenue pressures will scrap most long term fundamental research on these systems.

Dear All

Thank you for the question and for nice answers.

I agree with Dr Calos opinion, however, I see it a recent case rather than a future.

Deep learning and CNNs are most common technique and structure in divers fields expanding from predicting weather, cloud ,..etc. to diagnoses of serious deases. But I didn't see who evaluate their computational complexity expenses and other disadvantages!

What is the future of deep learning? Charles Cadieu, Co-Founder and CEO of Bay Labs, answers your question in this edition of Ask a Scientist...

https://www.youtube.com/watch?v=fs_Z-y3myiw