Monday, 14 December 2015

Week 5 [14-20.12.2015] Machines are learning how to learn - a novel approach to Machine Learning

I would like you to "read" or at least "go through" this fascinating article:
http://www.sciencemag.org/content/350/6266/1332.full. I am sorry, but this is more like scientific paper-ish than scientific article, however I still think it's worth reading.

To cut  a long story short, authors describe novel approach to Machine Learning called Bayesian Program Learning. This approach builds richer models of things it learns and tries to devise the connections between artificial concepts (or abstractions) it creates rather than use brute-force approach like Deep Learning (that is learning from massive sets of examples). It tends to outperform Deep Learning (so far the best known Machine Learning algorithms) in a number of tasks, especially so-called one-shot classification (learning to classify things based on just one example). The method is comparable to human in this regards (similar error rates).

Based on this article, can you comment on these statements:

1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

Do you think these statements are true or false, or at least plausible?

31 comments:

  1. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    Well this is a reason why I decided to study Neural Networks.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    This might happen.

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    Yes, this is true. Like many revolution such as industrial and computer revolution we are on the edge of new breakthrough. But lot of this problems was discussed with Marvin Minsky. We have efficient computers and lot of emerging technologies like optic processors and quantum CPUs. However from 1960 we are asking same questions.

    ReplyDelete
    Replies
    1. Doesn't Neural Networks require enormous number of examples?
      With you're risking overfitting or underfitting (that is inability to generalize well or over-generalization).

      Ideally we would like to find something (well, that could be Neural Networks) that given relatively small number of examples will be able to find abstractions and learn about their relations. Neural Networks aren't there yet.

      I must admit that Machine Learning didn't make the expected progress. If we compare Machine Learning abilities 40 years ago and now, keeping in mind the exponential growth of the computational power, we're definitely not doing very good job.

      We are probably missing something, when the best algorithm so far to train Neural Networks is back propagation combined with Statistical Gradient Descent. There have to be a better way; that's not how we learn :)

      Delete
  2. This comment has been removed by the author.

    ReplyDelete
  3. Hi,
    the article you posted is very interesting. I do not fully agree with it though. I think that cognitive models are the future, but in my opinion building such models requires some unsupervised pre-training. Just like a child observes a world before it is tought to classify objects, reconize objects and so on, a ML model will have to be pre-trained with some amount of training data. Without that it's impossible to build a context to generalize and generate new patterns.
    Current state of art deep learning algorithms have less than half percent error rate (0,21% - http://cs.nyu.edu/~wanli/dropc/ ).
    Problem with getting it better is that sometimes it's very hard even for a human to classify a character. Handwriting recognition and voice recognition problems are difficult, because handwriting can be scribbled and voice recordings can be noisy.
    I think it will take a while before we'll deploy intelligent robots and cut down employment of unqualified staff. Do you imagine robots cleaning streets in Poland? It'd be stolen, taken apart and sold on scrapyard in minutes :)

    ReplyDelete
    Replies
    1. Now it's getting interesting. We're about to face the old philosophical battle :)

      Do you think child's brain is pre-programmed or is it a clean sheet of paper (or "card blanche" as philosophers called that)?

      Delete
  4. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    In my opinion this statesments is true because we can't find all of cases for examples, because there are many factors which create a example, so better is induction which can make a model for Artificial Intelligence.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    It is very difficulty statesments, it might happen of course but in my opinion will be easier to master Voice Recognition then Handwritting because now computer better cope with Voice Recognition then Handwritting

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    In my opinion this statesments is true but it is future , first of all we must have one robot who replace people in one job. From this point politicians must be think about it

    ReplyDelete
    Replies
    1. >> 2. In a relatively short time (say 10 years) computers >> will be able to beat humans in Voice Recognition and
      >> Handwriting Recognition.

      > It is very difficulty statesments, it might happen of
      > course but in my opinion will be easier to master Voice
      > Recognition then Handwritting because now computer
      > better cope with Voice Recognition then Handwritting

      That's what I thought. However, my PhD advisor said something else :)
      When it comes to recognizing surnames, it is not as easy as we think - the Voice Recognition systems are build to recognize full sentence and actively use its structure. The problem with surnames is the fact, that it might be hard to recognize if the computer is hearing it for the first or second time - it is exactly the situation mentioned in the paper: one-shoot learning.
      And it seems to be not as big problem with Connected Handwriting Recognition (although recognizing different writing styles and writing systems is a very challenging problem).

      Delete
  5. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    The statement may be correct. However I don’t agree with this (at least not now). It’s one thing to use a framework to generalise and extrapolate on simple symbols (the subject of the research presented in the article), but it gets much more ‘experience’ (i.e. bunch of samples, basic ideas) to come up with the wider concept or new idea based on some snapshot. I’d rather say that cognitive models using learning concepts AND enormous number of examples are the future of AI.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    I could agree with this statement.

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    No, unskilled repetitive task based jobs are already being replaced by automations. It is already happening using the current state of the art. No further Machine Learning is required to achieve this. Consider it done. But when these rich models you mention will materialise in an available technology, then the vast amount of more skilled jobs will be at risk.

    ReplyDelete
    Replies
    1. > I’d rather say that cognitive models using learning
      > concepts AND enormous number of examples are
      > the future of AI.

      Do we need very large number of examples to learn certain concept? Or we need just a few? A few here may still mean large number, but I tend to believe that it could be tens of thousands rather than millions.

      Delete
  6. I agree with previous comments about future of AI is composed by cognitive models using learning concepts AND enormous number of examples. Example with child describes my way of thinking about that point.
    Such a huge advance in voice and handwriting recognition sounds very likely to happen. I wonder what will come instead of today's captcha (and how more annoying will it be).
    I'm not sure what to think about your last question. That technology will be surely very expensive, so potential "replace" would last for a long time. I think the main demand would be generated by military area.

    ReplyDelete
    Replies
    1. Actually, Google is using something else. "Are you a human". It seems that they know whether you're human or a bot, just by the way you "click" the button.

      Delete
  7. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    I suppose that more common way to learn AI is use more rather induction than enormous number of examples. Artificial neural net should in my opinion be similar to biological neural net. Then the learning process will be more natural and similar to biological learning process.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    It's not possible, if human can't recognize something, that means it's hard or impossible to recognize. In learning process the human assessment is indicator of learning grade during training.

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    That's totally not true. There will be job cuts for sure, but there will be also new worksite in robots management. Workers will have to be more qualified and specialized in robots and automation engineering.

    ReplyDelete
  8. I think the ability of human being to learn from even one example, sparse, limited data cannot be beaten by machines which have to Take In enormous number of examples to reach the same level of recognition. Humans can derive new concepts from simple examples whereas machines cannot.
    It is fascinating that BPL framework as a kind of artificial intelligence enables machines to construct new programs by reusing pieces of existing ones and is able to approach a real world processes.
    Machines can recognize a handwriting or a voice using some complicated schema of parts, subparts and relations. They take into consideration a lot of combinations: movements, flowing of ink from a pen, stroke trajectories to mention a few.
    Yet humans are far better at learning new concepts. They need fewer examples
    I don’t think politicians care about unskilled workers. But there is a threat that worse - educated people will be replaced by machines in the future.

    ReplyDelete
  9. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    I think that's true but it gonna be in very far future. Near future will be based on learning enormous number of examples even if it's gathered by robot. After this stage AI can figure out answer to questions even without knowing the topic. You can see some similar situation in Watson computer.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.
    I think it will be possible.

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    This is quite true. In some large IT companies there are people responsible for turning on and off switch. So even if there is such capability for replacing humans with program, companies sometimes prefer to hire a person for this kind of task.

    ReplyDelete
  10. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    I think it would be interesting to compare results of both approaches.
    As it was already mentioned they might turn out to be complementary.


    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    I’m sure about it.


    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    Thank you for taking this point up. This statement makes me ask more questions. I agree with one on my predecessors saying that we do need to wait long to see extinction of low-income unskilled labour jobs as this is happening now and in the future more and more complex tasks will be replaced by computers. I would not count on politicians as the jobs cut perspective exceeds their mandates in most of cases. But the question is still valid. I believe there will always be something good to do for those who want it. Should the jobs be ensured indeed? If yes, is closing people in corporations a really good idea ? How to find morally strong managers ? What should be done in order to enhance people’s creativity in a country perspective? How should future education look like ?

    ReplyDelete
  11. 2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.
    As Speech Recognition technology improves in terms of accuracy, vocabulary, and its ability to understand natural language, in relatively short time, we will see the concept of interactive machines in every arena. From assembly line mechanical tools to intelligent microwave ovens to "writing" a check, we will have the power to use our voice to instruct the electronic devices we encounter every day. Nowadays humans are trying to create a computing environment in which the computer learns from the user instead of one where the user must learn how to use the computer. Speech Recognition technology is the next obvious step in an attempt to integrate computing into a "natural" way of life. This effective means of communication, even when perfected, will still present limitations as to how humans can express themselves. The bottleneck of the future will be the physical constraint of not being able to speak all of one's thoughts in a coherent and sensible fashion.

    ReplyDelete
  12. Hey,

    Very interesting paper, I am excited for the future possibilities that this will open the path to.
    Frankly speaking I am not really convinced by the idea of decreasing the required number of examples in order to perform classification. What if we are given a special case of an example for which certain criteria are totally different from typical ones? Of course lowering the threshold of required number of examples at which we achieve acceptable error rates is a good practice and it should continue to be researched.
    It would be great to have AI surpass our capabilities, although we need to take it with a pinch of salt, many things can happen during those 10 years that will hinder or support advancement in this area.
    Which jobs are particularly threatened by this breakthrough? I guess people should adapt to using tools which will become available and which will help them in their work. Other than that I don't think it will cause major job cuts.

    ReplyDelete
  13. I believe that these statements / sentences are likely. coprawda I do not follow this subject very intensely. I consider, and I presented in my speech to the Scientific Supervisors. That's what I think about progress. I was showing, among others, Moore's law, which says Moore's law is the observation That the number of transistors in a dense integrated circuit doubles every two years Approximately. The observation is named after Gordon Moore, the co-founder of Intel. His prediction provede accurate for several decades, and the law was used in the semiconductor industry to guide long-term planning and to set targets for research and development. Advancements in digital electronics are strongly linked that Moore's law: quality-adjusted prices microprocessor, memory capacity, sensors and even the number and size of pixels in digital cameras.
    I think, that from month to month, from year to year, we gain amazing possibilities previously didn't have.

    ReplyDelete
  14. The issues of ML and AI have been already discussed on this blog. Nobody needs to convince me that development of both ML and AI technologies is progressing rapidly and faster each year. Their popularity and demand is growing too, as can be observe on many job hunting pages; the ML skills are among top 5 in demand.

    ReplyDelete
  15. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.
    It's hard to predict, maybe both of them? Using machine learning is very hard, you need a lot of resources, examples and so on. Maybe learning by induction will cut some costs. Who knows?

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.
    It's good, better voice recognition can be very usefull, same as handwriting recognition. For example, using voice recognition in our cars or just in our daily work. Like writing code by using our voice :)
    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.
    It's good to read this article:
    http://observer.com/2015/08/stephen-hawking-elon-musk-and-bill-gates-warn-about-artificial-intelligence/
    In my opinion we can create next generation of ourselves that can replace us.

    ReplyDelete
  16. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    The future of learning lays in mummification of human brain learning model, becous we don't know yet an entity with better learingng abilities in this universe.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    It is possible if I we could imitate human brain learning process, then computer can beat the human if we add its computational capabilities like speed and precision,

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    I supouse then we won't need to work, wor will be a human's choice becouse robots will work for us - am I right ?

    ReplyDelete
  17. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.
    I think that AI needs both of them. We are the best examples. We often learn on examples :) so if we can do thay why machines should not ? if we want to create a true AI we need to allow them to have as many ways of learning as possible, we do it in the same way

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.
    why not. i think that this can happen even faster. This two things are based on specifi alghoritms, and mathematical operations and we all know that machines are better then us in these things

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.
    Yes, AI is very important element of our future and we should react before it will be available. If we do it to late than it will be harder to handle with the new technology

    ReplyDelete
  18. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    I think I agree with this sentence. Of course like everything in IT is is difficult to say it for sure, but I like the idea because this approach is much more alike to human brain.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    I am sure it will happened, but probably in more than 10 years (but who knows, we will see.. ;)).

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    In my opinion it is not a new issue. Technology progress always influenced on people's jobs. Some of them, especially low qualified, are disappearing. Automation is profitable for the employer. That is one side of this, but in the same time new jobs and areas are created, so I do not predict a big catastrophe on the labor market.

    ReplyDelete
  19. 1. The future of Artificial Intelligence lies in rich cognitive models built using
    learning concepts by induction, rather than using enormous number of examples.
    It may be true but we have to remember that a human brain has started building a representation of objects since people are children. Therefore, I wouldn't say that after only one example human brain is able to success from one example.
    2. In a relatively short time (say 10 years) computers will be able to beat
    humans in Voice Recognition and Handwriting Recognition.
    I believe that it is possible to happen. I am not sure if it will happen simultaneously. Someday it will happen and will be discussing another advanced techniques which are being developed.
    3. This kind of advancement in Machine Learning is a call for action for the politicians.
    They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.
    There are many benefits and drawbacks of this occurance. The most popular cause of any inventions is humans' laziness. One of the side effects is higher unemployment in some branches. The threat of unemployment may lead to the stimulation of the economy and innovations.
    Is that good to keep low-income unskilled labour jobs? The would have to happen a kind of a shift in the employemnt structure regarding the skills the employees have. There will be a need for employees who will know how to serve new technology, developers, consultants, etc.

    ReplyDelete
  20. 1. I think it will depend on a task, I don't think any of those approaches is final yet and each of them provides some new knowledge and maybe together will involve in a new and even more efficient methods. 2. I would say definately but I was dissapointed with such predictions so many times that this time I will say it is possible. 3. We have been experiencing technological progress for so many years and people still have jobs. So I think people are very good in adjusting to new situations. There will be a lot of new jobs for people fixing, maintaining and cleaning ; ) those machines.

    ReplyDelete
  21. This comment has been removed by the author.

    ReplyDelete
  22. 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.
    I agree with previous speakers, also I suppose that these approaches can complement each other.

    2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.
    I think that even before (not counting the handwriting of doctors)

    3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.
    Automation is cost effective for the employer. Typically, in factories weakest, the most unreliable, the least exact item is now man. However, if society would be richer (I hope) that there will always be new services which will be in demand. The problem lies in the education system set further knowledge for solving tests and not to apply this knowledge. Today, access to information is almost unlimited but you need to be able to take advantage of it.

    ReplyDelete
  23. > 1. The future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, rather than using enormous number of examples.

    I thought both are used in combination. Maybe the better answer was; we developed the former thanks to latter ?

    > 2. In a relatively short time (say 10 years) computers will be able to beat humans in Voice Recognition and Handwriting Recognition.

    I agree.

    > 3. This kind of advancement in Machine Learning is a call for action for the politicians. They should act now, or we will face tremendous job cuts, basically an extinction of low-income unskilled labour jobs.

    I do have privacy concerns above all the others.

    ReplyDelete
  24. This comment has been removed by the author.

    ReplyDelete
  25. 1. I'd say it probably depends on the problem being solved. Learning concepts by induction does sound "human" but isn't learning from enormous numbers of examples also the way humans learn? How do you explain seemingly obvious and natural things like love, trust, jealousy, sarcasm etc. to a child? Most likely by giving examples!

    2. Of course! After all, humans are not 100% accurate either. I'm pretty sure most of us have had trouble understanding what their doctors had scribbled or have heard even Polish songs they couldn't understand.

    3. I wouldn't worry about this too much. People have had this concern for quite some time now but a catastrophe has yet to happen, even though many tasks have already been taken over by machines. People adapt and they will always find stuff to do. Many professions that used to exist don't exist anymore and I'm sure there are many that don't exist yet.

    ReplyDelete
  26. 1. I agree that the future of Artificial Intelligence lies in rich cognitive models built using learning concepts by induction, especially if computers would be able to observe the world with multiple sense, exactly in a way humans do. For example they could experience a vehicle (the example from the article) not only on a picture but also in real life, as a moving object, having a certain influence on it's surroundings.

    2. Certainly computers are already very good in handwriting recognition, so 10 years is a possible timeline. I have no idea about voice recognition.

    3. Obviously this is a call for action for politicians and for people in general to open up to new kinds of jobs possible with this kind of technology - a lot of jobs connected to better city management just to name one example. I would see the process as an opportunity for new kind of performance in every domain, rather that a threat.

    ReplyDelete