One day machines can communicate with us the
same way that we communicate with each other, and that way they will change our lives. My name is Regina Barzilay and I am professor
of computer science and electrical engineering at MIT. I develop machine learning methods that enable
machines [to] understand human language and perform useful tasks like finding relevant information
or summarizing the documents. So machine learning is a mechanism that enables
machine to do the mapping from the beginning input conditions to the output conditions. So instead of kind of specifying a very detailed
mechanism–how to go from A to B–you give them the input and the desired output, you
train them, showing them multiple examples and it enables them to figure out, kind of
the mapping from input to output. I read my first book on human decipherment
of language, and I read how painful it was for humans to do it. It’s really sometimes decades-long process. And I wanted to see if we can change it. And that’s what inspired us to look at Ugaritic. And obviously there are some cases where you
don’t have enough data. In Ugaritic you do have enough data, and it
belongs to the same semitic language as Hebrew. So machines, if you structure the model appropriately,
can learn the patterns and identify how the alphabets map to each other and how the related
words from these alphabets can be mapped and that’s what enabled us to solve this model. My research on language grounding is motivated
by the fact that whenever we as children are learning language, we are actually always learning
it in context, and we’re getting feedback from interacting with others. So we started looking in the virtual environment,
where the semantic of language actually is pretty much defined in the actions that you
take. It could be the actions that you take in a
strategy game like Civilization. It means that you are actually–if you understood
the instruction, you solved them correctly. And in this case machine gets feedback when
these instructions are actually executed. Eventually the models, in a smart way, can
be guided by this feedback. I got breast cancer three years ago. And when I went through the treatment, I noticed
we don’t utilize data at all. So what we did was take the three decades
of pathology reports of women who were screened and translated them into databases. I think it will contribute to both preventative
methods and also for over-treatment. We are looking at creating models that can
read mammograms and predict whether the woman– who doesn’t have currently cancer, whose mammogram
is still very normal–what are her chances to get cancer in five years. So we’re using deep learning to do these kind[s]
of predictions related to disease progression. Despite the fact that, you know, it’s clear
to me and it’s clear to people at MIT that machine learning can really revolutionize
this work, this award would help me to continue to do research in cancer and enable me to
move this direction forward.

Computer Scientist Regina Barzilay | 2017 MacArthur Fellow
Tagged on:                 

Leave a Reply

Your email address will not be published. Required fields are marked *