02.04.2013 Views

Statistical Language Models based on Neural Networks - Faculty of ...

Statistical Language Models based on Neural Networks - Faculty of ...

Statistical Language Models based on Neural Networks - Faculty of ...

SHOW MORE
SHOW LESS

You also want an ePaper? Increase the reach of your titles

YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.

have to discover while learning a language. The sec<strong>on</strong>d step would be to design learning<br />

algorithms that can learn given patterns using limited amount <strong>of</strong> resources (number <strong>of</strong><br />

training examples, computati<strong>on</strong>al time). Simplicity <strong>of</strong> such proposal is actually an advan-<br />

tage: importantly, after the definiti<strong>on</strong> <strong>of</strong> the incrementally more difficult tasks, it should<br />

be possible to objectively compare many different approaches and techniques developed<br />

by independent researchers.<br />

There are <strong>of</strong> course many possible modificati<strong>on</strong>s and extensi<strong>on</strong>s <strong>of</strong> this proposal that<br />

may be crucial for successful learning <strong>of</strong> the language using reas<strong>on</strong>able amount <strong>of</strong> resources.<br />

For example, it might be important to learn the language together with observing the real<br />

world (or simulated world) situati<strong>on</strong>s [10]. Also, it might be important to use an active<br />

system that does not just passively predict future events. Measuring success <strong>of</strong> different<br />

systems <str<strong>on</strong>g>based</str<strong>on</strong>g> <strong>on</strong> their ability to predict future events (words) might be sufficient in the<br />

beginning, but also can be insufficient in the l<strong>on</strong>g run - an AI system should probably aim<br />

to maximize some reward functi<strong>on</strong>.<br />

The representati<strong>on</strong> <strong>of</strong> the AI system itself is another important topic: due to compu-<br />

tati<strong>on</strong>al limitati<strong>on</strong>s, traditi<strong>on</strong>al neural network architectures might be insufficient. Repre-<br />

senting AI by a Turing machine or a computer program that is automatically c<strong>on</strong>structed<br />

during training can be a better opti<strong>on</strong>, although with other difficulties as discussed in [43].<br />

Clearly, the deeper we go, the more questi<strong>on</strong>s arise - thus, my opini<strong>on</strong> is that the prob-<br />

lem <strong>of</strong> language understanding by computers should be solved step by step, starting with<br />

simple problems where various techniques that can be easily compared.<br />

108

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!