达到人类级别的AI：深度学习面临的挑战

Challenges for Deeping Learning towards Human-Level AI

演讲者：Yoshua Bengio，AI界三位大神之一

下面摘录精彩部分，一起分享。

Humans seem to be much more efficient than current AI at learning from unlabeled observations and interaction with their environment, and current machine learning system do not seem to understand their training data nearly as well as humans. 在学习无标签观测数据和与之相处的环境交互时，人类似乎比现阶段的AI更加高效，当前的机器学习系统几乎不能像人类一样去理解训练数据。

A core objective of deep learning is to come up with learning frameworks which can discover disentangled representations which explain the important variations in data. 深度学习的核心一个目标是提出一个学习框架，利用它我们能发现并理清错综复杂的表达，正是这些表示解释了数据背后的重要变量。

Progress in deep generative networks based on an adversarial criterion has been impressive and we show how these ideas can be used to estimate and optimize entropy and mutual information and how this could be used towards unsupervised learning of high-level abstractions.基于对抗标准和深度生成网络取得的成果令人印象深刻，已经证明了这些好的ideas如何去预测和优化熵和互信息, 以及如何用来解决抽象度更高的无监督学习。

This follows the ambitious objective of disentangling the underlying causal factors explaining the observed data. 更近一步，这些ideas可能会帮助实现我们的宏伟目标：理清背后的因果因素，以此解释观察到数据。

We argue that natural language understanding cannot come from current attempts purely based on text corpora.我们发现：自然语言理解是不能仅仅基于语料库的！

Instead, a learning agent must acquire information by acting in the world and jointly learn a model of the world and how language can be used to refer to it. 相反，一个学习中间件必须通过作用在现实中去获得信息，然后学得一个更贴近现实的模型，进而利用这个模型解决自然语言理解中的问题。

Natural language could be used as an additional hint about the abstract representations and disentangled factors which humans have discovered to explain their world. 自然语言可以看做是一个线索，借助它尝试发现人类自己已经学会的去解释世界的技能，比如抽象表达，理顺逻辑关系。

Some conscious thoughts also correspond to the kind of small nugget of AI. 一些潜在的意识表达也正是AI研究的重中之重。

This therefore raises the interesting possibility of addressing some of the objectives of classical symbolic AI focused on higher-level cognition using the deep learning machinery augmented by the architectural elements necessary to implement conscious thinking about disentangled causal factors. 总之，以上这些引人深思，它们可能帮助我们实现目标。比如更高级别认知的AI借助深度学习的机制，同时利用框架中的各个组建辅助解决，这些组建是必要的，能去实现潜在的思考能力，比如关于理顺因果逻辑关系。

文章翻译不准确之处，敬请指正。