We are excited to share the latest report and prototype from our machine
intelligence R team: Multi-Task Learning.
Wax on.. face off!
When humans learn new tasks, we take advantage of knowledge we’ve gained from
learning, or having learned, related tasks. Take the 1984 movie Karate Kid,
where Mr. Miyagi takes on Daniel as his martial arts student. He begins
Daniel’s training by having him complete various lengthy menial chores, from
waxing cars to sanding floors. When Daniel expresses frustration, Mr. Miyagi
reveals that these chores, by training his muscle memory, will help Daniel
master martial arts blocks. This fictional example is a great illustration of
how humans draw on mastery of related tasks when learning new tasks.
Machines struggle to take advantage of such task relationships. To date, most
machine learning algorithms are trained to master one—and only one—task.
Because such single-task-trained algorithms only know one task, they do not,
and cannot, benefit from the relationships between tasks.
Multi-task learning is an alternative approach to training machine learning
algorithms that allows machines to master more than one task; machines gain the
ability to benefit from task relationships. Machine learning becomes, a little
bit more, like human learning - capable of taking on more complex challenges
involving richer representations of reality.
The Benefits of Multi-Task Learning
Newsie, the prototype that accompanies this report, classifies news articles
into categories (e.g., news, sports, entertainment). It was trained using
multi-task learning to correctly classify buttoned-up broadsheet articles
(Task 1) and more sensationalist tabloid ones (Task 2).
Newsie’s article view shows how the model arrives at the classification on a
word-by-word level. From surrounding sentences to place of publication,
context changes word meaning. Enabled by multi-task learning, Newsie provides
users with a window into the differences in coverage and language use across
the buttoned-up and sensationalist press; it puts current news into
perspective.
The ability to perform a fine-grained analysis of the news is just one way of
using the emerging capabilities of multi-task learning. Our report and
prototype is valuable for anyone looking for better solutions to complex
classification problems, efficient model performance, reduced model
maintenance, or the ability to train one model on several different, related
sets of data (as we did for Newsie): multi-task learning has many benefits,
some of them quite subtle. To inspire future applications, in a recent
post,
we reviewed the successes of multi-task learning across a range of applications
and industries. From healthcare and finance to robotics and agriculture,
multi-task learning has the potential to take classification to the next level.
Multi-Task Learning, Why Now?
Like many concepts in machine learning, multi-task learning is not new; Rich
Caruana’s review
article from 1997
(!) remains one of the best introductions to the topic. Instead, multi-task
learning is newly exciting.
Multi-task learning is an approach to training machines, not an algorithm. From
decision trees and random forests to neural networks, we can train these
algorithms using multi-task learning. But, simpler algorithms, such as decision
trees and random forests, require strongly related task to reap the benefits of
multi-task learning; they lack the capacity to benefit from complex
relationships between loosely related tasks. By contrast, neural networks can
learn complex relationships in data, including complex relationships between
(loosely related) tasks. We explore different architectures for multi-task
neural networks highlighting the value of recent technical innovations, such as
adversarial learning, popularized by Generative Adversarial
Networks, and skip connections, used to train
very deep
networks, for
multi-task networks. Neural networks make multi-task learning newly exciting.
Combined with multi-task learning, adversarial learning allows us to learn task-agnostic representation, representations that support all tasks and yet contain no task-specific information (here, the shared LSTM).
We demonstrate how current, modular deep learning framework simplify the task
of architecting a multi-task neural network. Multi-task learning not only
offers better solutions for complex problems, it has become easier to
implement, too.
Hear us speak about multi-task learning:
