View All Pages

Project Proposal

Due by 10pm Monday 9/26

Team Forming

The first step in starting your project is team formation. Teams should ideally have 2-3 members. Some of you may already have a team, others may be more interested in specific topics. Its easiest to communicate over Piazza if you are looking to form a team. Once you have formed your team, please create a group in bCourses. Under "People" on the left menu, select "Groups" and then under the Settings wheel "View User Groups". At this point an "Add User Group" button should appear (make sure you are actually logged into bCourses and not viewing it anonymously). Then you can name the group and invite people into it. Note: bCourses (i.e. Canvas) does not treat student-formed groups as "official" groups for assignment submission. Once all the student groups are created, we will copy them as formal project groups.

Project Topic

The recommended projects are included in last Wednesday's lecture:

/courses/1453965/files/69621037/download?verifier=o7mrBH1K3cfL9wEOOHBHlGZ2UQd4fxhn72WmqbNW

You can also do your own project about something not on this list. Implementations and extensions of other recent research papers are particularly encouraged. The best sources are conferences like NIPS Links to an external site., ICLR Links to an external site. and ICML which is archived in the Journal of Machine Learning Research Links to an external site.. If you have interests in specific areas you can try conferences like CVPR, ICCV and ECCV (computer vision), ACL and NAACL (natural language), ICRA and IROS (robotics) and Interspeech and ICASSP (speech recognition). Here are some other resources on research papers and datasets:

Writeup

For the project proposal assignment, we would like you to produce a 2+ page Project Proposal. You can either submit pure text or a pdf file. This proposal should include these sections:

The Team

Give a list of team members and their backgrounds.

Problem Statement and Background

A high-level statement of the problem you intend to address, i.e. recognition of objects in images. Be specific - i.e. "recognition" in this case means outputting a label for the object and also a bounding box for its location.

Give background on the problem you are solving: why it is important, what is the start of the art - ideally for both neural and non-neural approaches to the problem. Include some references to papers that describe the method you plan to use, or to build on.

The Data Source You Intend to Use

Describe the data source(s) you will use. If you're not doing one of the recommended projects, make sure you have access to the data you want to use *in the quantity and quality you need*. Ideally you should get all the data in your hands before starting your project. Several teams in related courses have run into trouble because they couldnt obtain the volume or quality of data they needed. Many online data sources (e.g. Yelp) strongly throttle the data volume you can get through their API. Others (Twitter) have a higher limit but have use restrictions that you must respect. Certain types of meta-data (e.g. location, demographics) in social media sources are both sparse and unreliable.

Describe how you plan to obtain the data, or how you got it if you already have it. Remember the goal of the project is to explore neural models, not to do data wrangling so you should do any such prep work asap.

Description of the Tools You Plan to Use

Describe the tools you plan to use for modeling in your project. You are free to use any tool, but our ability to help you (or for you to get peer help) falls off for less-popular tools. The ones you should have most luck with are:

  • Tensorflow
  • Caffe
  • Theano
  • Torch

Evaluation

How will you evaluate your model? Even for simple prediction tasks, there are a variety of tools you can use, e.g. ROC and lit plots and AUC measures.

Submission

Use this link to submit your writeup:

https://bcourses.berkeley.edu/courses/1453965/assignments/7743665

Make sure it includes Project Title and Project Members prominently at the top of the page. Please make sure project member names match exactly with registered campus names - i.e. the names in the "People" page in bCourses.