1）算法（Algorithm），或很可能在未来被统称为“ AI”的技术，将逐渐成为我们与日益充塞资讯的世界接口的唯一界面。
2）深度学习模型非常脆弱（brittle)、非常耗费数据（extremely data hungry)、并且不能推广来概括（generalize）数据分布（data distribution）以外范围的演算。
3）认为现有的深度学习技术代表了AI所有的潜力和发展结果是错误的。不论是从构造（construction），或者是建模的训练（training），深度学习要做的是查找过去的数据（looking up past data）并执行数据的插值（interpolation）。这充其量只能实现局部的概括。这些深度学习系统可以稳健地完成他们训练有素的工作，可以理解他们之前所看到的内容，可以解决创建者计划之内的不确定性。但是，真正的智能需要的不仅仅是插值，它更需要外推来实现广泛甚至极端范围的概括（extrapolation)。举个例子，希望深度学习的能耐无所不包，就好比期望如果汽车的车轮旋转得够快就会开始飞行一样不切实际。诚然，汽车可能非常有用，但若我们认为汽车可以在任何地方行驶，并且是唯一需要的交通工具，那就错了。
4）在当前框限人工智能研究的主要障碍不是缺少硬体设备，而是缺乏不同的构想。如果你只有有限的资源，不要花时间担心图像处理器（GPU）不够用，反之，请多花心思思索你是否正在解决正确的问题，问对的问题。
5）机器学习在某种层度上同时被过度渲染和低估。在一方面，人们倾向高估机器学习系统的智能以及归纳数据的能力 -- 错误的希冀机器学习如同你向任何问题一挥，就能使问题解决的魔术棒。这是很大的错误 -- 机器学习的算法的实质智能含量其实非常少，而且机器学习可以应用的范围也极其狭小。但与此同时，大多数人也低估了今时今日相对“生涩”的机器学习所能够发挥的作用 -- 尤其是如果我们得以系统性的运用机器学习于每一个它有潜力解决的问题上。机器学习在某种方式上，是我们时代的蒸汽动力，是一个相当基本但在大规模应用的情况下却足以大幅度改变世界的机制。
6）在10年之内，你将会买到妥善总结人工智能于2010至2020年间每一个发展跃进历程的教科书。今时今日大量泛滥的人工智能内容看似重要，但多数是噪音。因此，请专注在大问题。
7）深度学习建模就像巨大的参数化和微分模型，让数据经由输入空间到输出空间，以梯度下降法来培训。它学习数据从输入向量空间到输出空间的连续几何变形。由于这过程是点带点的完成，一个深度神经网络只能理解在几何空间中非常靠近它培训数据串之位置的输入数据。深度学习充其量可以进行跨点插值。但是，这意味着要有效的训练人工神经网络建模，你需要为输入数据进行密集的采样，而且几乎是逐点采样 -- 如果你要处理的是诸如自动驾驶或机器人等复杂的现实世界问题，这会非常的昂贵。反之，你可以查看非常简单的规则导向算法。如果你有适用的符号规则，由于它是抽象表述而不是靠点带点逐点映射来运算，它实际上可应用于非常大的输入数据。简而言之，深度学习很类似逐点几何变形。另一方面，抽象规则可以更好地概括。我们未来可以将两者结合起来（应用）。
[QUOTE] Google Scientist cum Deep Learning Framework Keras inventor, François Chollet:
1) Algorithms, and in general "AI", will increasingly serve as our sole interface to a world that is increasingly made of information.
2) Deep learning models are brittle, extremely data-hungry, and do not generalize beyond their training data distribution.
3) It would be a mistake to believe that existing deep learning techniques represent the end-all-be-all of AI. By construction, by training, what deep learning does is looking up past data and performing interpolation. This can implement local generalization -- at best, systems that can robustly do what they're trained to do, that can make sense of what they've seen before, that can handle the kind of uncertainty that their creators have planned for. But intelligence as I formally define it in the paper needs to feature extrapolation rather than mere interpolation -- it needs to implement broad or even extreme generalization, to adapt to unknown unknowns across previously unknown tasks.
They can at best encode the abstractions we explicitly train them to encode, they cannot autonomously produce new abstraction. They simply don't have the machinery for it -- it's like expecting a car to start flying if only its wheel would turn fast enough. Cars can be very useful, but if you think they can go anywhere and are the only vehicle we're ever going to need, you're mistaken.
The state of our knowledge is the same at project completion as it was when the project started.
4) The main thing that’s holding back AI research right now is not a lack of hardware, it’s a lack of diversity of thought. If you have limited resources, don’t spend your time worrying about GPUs, rather, spend it worrying whether you’re working on the right problem and asking the right questions.
5) Machine learning is, in a way, simultaneously overhyped and underrated. On one hand, people tend to vastly overestimate the intelligence and the generalization power of current machine learning systems, perceiving machine learning as a kind of magic wand that you can wave at arbitrary problems to make them disappear. This is, of course, largely false, there is very little actual intelligence in our algorithms, and their scope of application is extremely narrow. But at the same time, most people still underestimate how much can be achieved with the relatively crude systems we have today, if we apply them systematically to every problem they can potentially solve. Machine learning is, in a way, the steam power of our era: a pretty basic mechanism that nonetheless has the potential to profoundly change the world when used at scale.
6) In 10 years you’ll be able to buy a textbook that will neatly sum up every AI advance that has happened from 2010 to 2020. The flood of content being published today may look important, but most of it is noise. Focus on the big questions.
7) Deep Learning models are like huge parametric and differentiable models that go from an input space to an output space, trained with gradient descent. They are learning a continuous geometric morphing from an input vector space to an output space. As this is done point by point; a deep neural network can only make sense of points in space that are very close to things that it has already seen in string data. At best it can do the interpolation across points.
However, that means in order to train your network you need a dense sampling of the input, almost a point-by-point sampling which can be very expensive if you’re dealing with complex real-world problems like autonomous driving or robotics. In contrast to this, you can look at very simple rules algorithms. If you have a symbolic rule it can actually apply to a very large set of inputs because it is abstract, it is not obtained by doing a point by point mapping.
Deep learning is really like point by point geometric morphings. Meanwhile, abstract rules can generalize much better. I think the future is which can combine the two.
[/QUOTE]
参考资料/Reference:
[1]ZDNet： Keras inventor Chollet charts a new direction for AI - a Q&A https://www.zdnet.com/article/keras-creator-chollets-new-direction-for-ai-a-q-a/
[2]François Chollet@Tweet
https://mobile.twitter.com/fchollet/status/1144674106706558976
[3]François Chollet: Keras, Deep Learning, and the Progress of AI | Artificial Intelligence Podcast http://youtube.com/watch?v=Bo8MY4JpiXE
[4]François Chollet, creator of Keras on TensorFlow 2.0 and Keras integration, tricky design decisions in Deep Learning, and more
https://hub.packtpub.com/francois-chollet-tensorflow-2-0-keras-integration-tricky-design-decisions-deep-learning/
[5]Interview with The Creator of Keras, AI Researcher: François Chollet
https://hackernoon.com/interview-with-the-creator-of-keras-ai-researcher-fran%C3%A7ois-chollet-823cf1099b7c
Posts
0 Comments