> In DeepMind's AlphaGo, for example, most of the "intelligence" on display is designed and hard-coded by expert programmers (e.g. Monte-Carlo tree search);<p>Not true. This paraphrases the original paper:
<a href="https://www.tastehit.com/blog/google-deepmind-alphago-how-it-works/" rel="nofollow">https://www.tastehit.com/blog/google-deepmind-alphago-how-it...</a><p>> They tested their best-performing policy network against Pachi, the strongest open-source Go program, and which relies on 100,000 simulations of MCTS at each turn. AlphaGo's policy network won 85% of the games against Pachi! I find this result truly remarkable. A fast feed-forward architecture (a convolutional network) was able to outperform a system that relies extensively on search.<p>Also, this article reeked of AGI ideas. Deep learning isn't trying to solve AGI. Reasoning and abstraction and high level AGI concepts that I don't think apply to deep learning. I don't know the path to AGI but I don't think it'll be deep learning. I think it would have to be fundamentally different.
What about the future of <i>jobs</i> in the field of deep learning?<p>EDIT: I'm thinking deep learning will become much like web development is today. Everybody can do it, and only a few experts will work at the technological frontier and develop tools and libraries for everybody else to use.<p>Therefore, if one invests time in DL, then I suppose it better be a serious effort (at research level), rather than at the level of invoking a library, because soon everybody can do that.
I enjoyed part 1 of Chollet's two articles today but am less fond of this one. It suggests that deep learning will expand from its present capabilities of recognizing patterns to one day master logical relations and employ a rich knowledge base of general facts, growing into a situated general problem solver that one day may equal or surpass human cognition. Maybe. But he then proposes that deep nets will rise to these heights of self-organization and purposefulness using one of the weakest and slowest forms of AI, namely evolutionary strategies?<p>I don't think so.<p>The many problems bedeviling the expansion of an AI's competence at one specific task into mastery of more general and more complex tasks are legend. Alas neither deep nets nor genetic algorithms have shown any way to address classic AGI roadblocks like: 1) the enormity of the possible solution space when synthesizing candidate solutions, and 2) the enormous number of training examples needed to learn the multitude of common sense facts common to all problem spaces, and 3) how to translate existing specific problem solutions into novel general ones. Wait, wait, there's more...<p>These roadblocks are common to all forms of AI. The prospect of replacing heuristic strategies with zero knowledge techniques (like GA trial and error) or curated knowledge bases with only example-based learning is unrealistic and infeasible. Likewise, the notion that a sufficient number of deep nets can span all the info and problem spaces that will be needed for AGI is <i>quite</i> implausible. While quite impressive at the lowest levels of AI (pattern matching), deep learning has yet to address intermediate and high level AI implementation challenges like these. Until it does, there's little reason to believe DL will be equally good at implementing executive cognitive functions.<p>Yes DeepMind solved Go using AlphaGo's deep nets (and monte carlo tree search). But 10 and 20 years before that IBM Watson solved Jeopardy and IBM Deep Blue solved chess. At the time, everyone was duly impressed. Yet today nobody is suggesting that the AI methods at the heart of those game solutions will one day pave the yellow brick road to AI Oz.<p>In another 10 years, I predict it's just as likely that AlphaGo's deep nets will be a bust as a boom, at least when it comes to building deep AI like HAL 9000.
TLDR is that models will become more abstract (current pattern recognition will blend with formal reasoning and abstraction), modular (think transfer learning, but taken to its extreme - every trained model's learned representations should be applicable to other tasks), and automated (ML experts will spend less time in the repetitive training/optimization cycle, instead focusing more on how models apply to their specific domain).
This is part 2 from the post yesterday:
<a href="https://news.ycombinator.com/item?id=14790251" rel="nofollow">https://news.ycombinator.com/item?id=14790251</a><p>And the author posted a comment on hn:<p>"fchollet: Hardly a "made-up" conclusion -- just a teaser for the next post, which deals with how we can achieve "extreme generalization" via abstraction and reasoning, and how we can concretely implement those in machine learning models."<p>I like the ideas presented in the post, but its not concrete or new at all.Basically he writes "everything will get better".<p>I do agree with the point that we need to move away from strictly differential learning though. All deep learning problems only work on systems that have derivates so we can do backpropagation. I dont think the brain learns with backpropagation at all.<p>* AutoML, there are dozens of these type of systems already, he mentions one already in the post called HyperOpt. So we will continue to use this systems and they will get smarter? Many of these systems are basically grid search/brute force. Do you think the brain is doing brute force at all? We have to use these now because there are no universal correct hyperparameters for tuning these models. As long as we build AI models the way we do now, we will have to do this hyperparameter tuning. Yes, these will get better, again, nothing new here.<p>* He talks about reusable modules. Everyone in the deep learning community has been talking about this a lot, its called transfer learning and people are using it now, and working on making it better all the time. We currently have "model zoos" which are databases of pretrained models that you can use. If you want to see a great scifi short piece on what neural network mini programs could look like written by the head of computer vision at tesla, check out this post: <a href="http://karpathy.github.io/2015/11/14/ai/" rel="nofollow">http://karpathy.github.io/2015/11/14/ai/</a>
Interesting article, in a difficult topic. Speculating about the future of deep learning. The author deserves recognition for writing about this. In my personal opinion, within the next 10 years, there will be systems exhibiting basic general intelligence behavior. I am currently doing early hobbist research on it, and I see it as feasible. These system will not be very powerful initially. They will exist and work in simpler simulated environments. Eventually we will be able to make these systems powerful enough to handle the real world. Although that will probably not be easy.<p>I somewhat disagree with the author. I don't think that deep learning systems of the future are going to generate "programs", composed of programming primitives. In my speculative view, the key for general intelligence is not very far from our current knowledge. Deep learning, as currently we have, is a good enough basic tool. There are no magic improvements to the current deep learning algorithms, hidden around the corner. Rather what I think will enable general intelligence, is assembling systems of deep learning networks in the right setup. Some of the structure of these systems will be similar to traditional programs. But the models they generate will not resemble computer programs. They will be more like data graphs.<p>I expect within 10 years there will be computer agents capable of communicating in simplified, but functional languages. Full human language capability will come after that. And within 20 years I expect artificial general intelligence to exist. At least in a basic form. That is my personal view. I am currently working on this.
Glad to see Deep Learning "coming down to earth". This is the first high profile post I've seen that spells out exactly how DL models will become reconfigurable, purpose-built tools, and what a workflow might look like. We're still a long way aways from treating them like software components.
Here are the results of my research into program synthesis using genetic algorithms.<p>Using Artificial Intelligence to Write Self-Modifying/Improving Programs<p><a href="http://www.primaryobjects.com/2013/01/27/using-artificial-intelligence-to-write-self-modifying-improving-programs/" rel="nofollow">http://www.primaryobjects.com/2013/01/27/using-artificial-in...</a><p>There is always a research paper, if you prefer the sciency format.<p>BF-Programmer: A Counterintuitive Approach to Autonomously Building Simplistic Programs Using Genetic Algorithms<p><a href="http://www.primaryobjects.com/bf-programmer-2017.pdf" rel="nofollow">http://www.primaryobjects.com/bf-programmer-2017.pdf</a>
Seeing how gradient descent is such a pinnacle of deep learning, I can't help wondering: is this how our brain learns? If not, then what prevents us from implementing deep learning the same way?
Regarding logic and DL, there is NeSy workshop in London
<a href="http://neural-symbolic.org/" rel="nofollow">http://neural-symbolic.org/</a>
Are there popular modern libraries that do program synthesis? Although I've thought about this and read about the concept on hn, I've not heard it mentioned seriously or frequently or strenuously as a thing to do either just for fun or to get a job doing it. This could be a popular way to solve programming problems without needing programmers. I think this truly would kick off AI as a very personal experience for the masses because they would use AI basically like they do already do now with a search engine. People would use a virtual editor to design their software using off the shelf parts freely available. The level of program complexity could really skyrocket as people now have more control over what and how they run programs because they can easily design it themselves. Everyone could design their own personal Facebook or Twitter and probably a whole new series of websites too complex or for other reasons not invented yet.<p>For instance, you want to program the personality of a toy, so you search around using the AI search engine for parts that might work. Or you want a relationship advice coach so you put it together using personalities you like, taking only the parts you want from each personality. Or another example would be just to make remixes of media you like. Because everything works without programming anyone can participate.
I'm also interested to see how the worlds of program synthesis (specifically type directed, proof-checking, dependently typed stuff) can combine with deep learning. If recent neural nets have such great results on large amounts of unstructured data, imagine what they can do with a type lattice.
About libraries of models, it would be useful to have open source pre-trained models which can be augmented through github-like push requests of training data together with label sets.<p>It would allow to maintain versioned versions of always improving models everyone can update with a `npm update`, `git pull` or equivalent.
Self-driving cars are expected to take over the roads, however no programmer is able to write code that does this directly, without machine learning. However, programmers have built all kinds of software of great value, from operating systems to databases, desktop software and so on. Much of this software is open source and artificial systems can learn from it. Therefore, it could well be that, in the end, it would be easier to build artificial systems that learn to automatically develop such software than systems that autonomously drive cars, if the right methodologies are used. The author is right to say that neural program synthesis is the next big thing, and this also motivated me to switch my research to this field. If you have a PhD and are interested in working in neural program synthesis, please check out these available positions: <a href="http://rist.ro/job-a3" rel="nofollow">http://rist.ro/job-a3</a>
I'm wondering if we will ever figure out how nature performs the equivalent of backpropagation, and if that will change how we work with artificial neural networks.
I'm excited for the easy to use tools that have to be coming out relatively soon. There are a lot right now, but the few I've used weren't super intuitive like I feel like they could be.
Compression.<p>That one word disrupts his whole point of view. This idea that we need orders and orders of magnitude more data seems insane. What we need is to figure out how to be more effective with each layer of data, and be able to have compression between the tensor layers.<p>The brain does a great job of throwing away information, and yet we can reconstruct pretty detailed memories. Somehow I find it hard to believe that all of that data is orders of magnitude above where we are today. Much more efficient, yes. And that's through compression.