Taught by Patrick Hebron at ITP, Fall 2016
Week 9:
Technical Questions:
We will start today's class by answering any questions that arose from your continued explorations of TensorFlow, LaunchBot and Docker.
Recurrent Neural Networks:
- Discussion of The Unreasonable Effectiveness of Recurrent Neural Networks article.
- Discussion of techniques for encoding non-textual information for use with Char-RNN (and other RNN implementations).
- Discussion of sequence-to-sequence learning, its relation to the nature of experience and its incredible range of potential applications.
From Exploration to Formalization:
Project Development Exercise:
- Each student will present his or her homework explorations to the class.
- Members of the class should respond to the work presented descriptively, not qualitatively
- Resist the urge to extrapolate on intent or draw inferences about where an idea might be headed.
- Stay with your direct observations and don't be afraid to say something that seems obvious:
- What are some features of the work?
- What themes do you notice in its components?
- Based on the class' observations, take a few moments to consider the following questions individually:
- Where did you succeed in bringing across an idea?
- Where was your idea seen differently from what you had been thinking about?
- What themes do you want to clarify or bring forward in the work?
- Which features of the work were extraneous to your conveying the idea?
- Spend some time talking with your classmates about these explorations:
- Run ideas past them.
- Ask questions about their earlier observations.
Homework:
Assignment:
- Go back to your explorations and begin to formalize your project idea.
- Emphasize the essential:
- Find a pure distillation of the ideas you were pursuing in the earlier exploration.
- Identity elements of your ideas that may require overly complex architectures or hard-to-obtain datasets.
- If your idea relates to a very large scope, think about what microcosm could serve as a testing ground for this larger context. e.g., If your idea requires temperature readings from around the globe, could you test the idea with temperature readings from around your neighborhood?
- Start to think about how to expand a feature set from your explorations and their formalizations:
- One way of formalizing a feature set is to write a list of "user stories" - simple statements like: "As a user, I should be able to submit an image and get back a textual prediction of the objects contained within the image."
- From this high-level view of the feature set, try to identify the key algorithms and datasets that can be used to implement the desired features. Go back to the dataset resources we've looked at in previous weeks. Consider how these datasets fit your idea or how they might be adapted to it.
- Consider the conceptual pathways that seem to lead inevitably from your idea. As Chekhov said, "If you say in the first chapter that there is a rifle hanging on the wall, in the second or third chapter it absolutely must go off."
Readings: