Fall 2016 Computer Science 458: 11/14/2016

Administrivia

Per piazza posting, Jieung Kim office hours: Monday and Wednesdays, 5:30 pm to 7:30pm. AKW 311.

Upcoming Speakers

Wednesday November 16th: David Swensen, Yale Investment Office. Reading: Yale Endowment 2015 Reading: Pioneering Portfolio Management, Chapter 3: Investment and Spending Goals

Dinner guests:
  Fed Panfilov
  Pauli Xu
  Stephanie Hickman
  Davey Proctor
  Kenneth Seals-Nutt
Alternate:
  Chris Gunther (plus one)

Today's Guest Speaker:Alborz Geramifard, Research Scientist, Amazon Alexa

Bio:
Alborz Geramifard is currently a senior research scientist and manager at Amazon leading the Boston branch of Alexa's conversational AI. Before joining Amazon, he was a postdoctoral associate at MIT's Laboratory for Information and Decision Systems. Alborz received his PhD from MIT working with Jonathan How and Nicholas Roy on representation learning and safe exploration in large scale sensitive sequential decision-making problems in 2012. He finished his MSc at University of Alberta under the supervision of Richard Sutton and Michael Bowling in 2008, where he worked on data efficient online reinforcement learning techniques. His research interests lie at machine learning with the focus on reinforcement learning, natural language understanding, planning, and brain and cognitive sciences. Alborz was the recipient of the NSERC postgraduate scholarships 2010-2012 program.

Abstract:
Echo is high performance fully voice controlled wireless speaker created by Amazon that is designed for the home. It is very convenient, always plugged in and ready to use. The technology behind Echo lives on device as well as in the cloud and represents some of the best in Natural Language Processing (NLP) technologies today. Echo is extensible and supports a broad range of functionalities out-of-the-box, such as music, wikipedia, to-do and shopping lists, sports and weather information, generic question answering and more. It features high performance keyword spotting, automatic speech recognition, natural language understanding, question answering and text-to-speech. This talk provides an overview of speech and natural language processing activities at Amazon around Echo and describe some of the core technologies and research challenges our teams are facing.

Dinner guests:
  Fed Panfilov
  Pauli Xu
  Kenneth Seals-Nutt
  Bhavani Ananthabhotla

Alborz's cheat sheet for acing the Amazon machine learning interview.

Final Project: Additional Specifications

Final projects are due the last day of reading period: Thursday December 15th.

Be prepared to give a brief presentation of your project to the class after Thanksgiving break. It should be 5-10 minutes, and describe what you intend to do. If you have already implemented something, all the better.

Post project proposal to class*v2 under assignments Once you get approval, you may commence work. Submit your proposal this week. Leave time for an iterative review process.

Your project should run in the zoo. If you need additional libraries or software installed, you must clear it with the TF.
You need to include a write-up explaining your project, including how to run it, what it does, what is interesting or important about it, how it explains the decision. Screen shots might be helpful as well.
You can create a web site. In this case, it may not need to run in the zoo. However, you need to submit all the code so that it could theoretically be run in the zoo.
Can you leverage a project for another course? It depends. You need to check with me first.
Can you change your project after your proposal? Well, yes. The real question is what happens if you do that? The answer is again, it depends. If your initial idea hits a wall and does not work, you probably should figure out another way to go.

Final Project: Criteria

In no particular order:

Originality. I appreciate the degree to which you come up with new ideas. There are lots of ways to tackling the projects.
Functionality. The program should work. It is best if we can run it ourselves, such as on the zoo. At the very least, you should provide ample evidence that the program does what it is supposed to do.
Explanation. The program should explain its behavior. A good way to think about this is how you would explain your program's output to someone observing your program in action.
Relevance to course. The program should have a connection to a topic or topics discussed in class, including the guest speakers. The more direct the connection, the better. Note that the best connection is through a method or technique, not a domain. Thus, applying Monte Carlo simulation to medicine is better than applying astrology to finance.
Clarity. This is valued in thought, word, and deed. How clear is your concept? Do you describe it well in your write-up? Does your program clearly reflect your idea?
Success. Actually, not so much. In a freshman chemistry course, I selected a project of making polywater - a polymer form of water which had been reported by Soviet scientists. The professor encouraged me to pursue this and it failed. I was unable to replicate the Soviet experiment. (As it turns out, neither was anyone else.) My point is that you may learn from failure as well as from success.

Lecture

Self-driving Cars: an introduction

Cognitive Models of Decision Making: VOTE

A Realistic Model of Rationality (More readable PDF version of above.)

Running VOTE on the zoo

An Intentional Arithmetic for Qualitative Decision Making

Affect

human.java (Java human class with emotions)

[Home]