Last week, I attended the CRA Snowbird conference for the first time. This conference is held every 2 years and is attended by chairs and
associate chairs from departments in the U.S. and Canada, faculty
member on boards of organizations like CRA, representatives from
industry and funding organizations, and several big name researchers.
The conference was much more interesting that I expected. Below are my notes from the conference which are (as always) sketchy, biased and incomplete.
1) John Hennessy – President of Stanford University (and former CS professor)
- Until the invention of the printing press, doubling rate for universities was about every 100 years
- Major university cost was the library
- After printing press, university cost decreased dramatically -> huge increase in access to education -> dramatic societal impact
- Currently, major university cost is faculty salaries. Research faculty are very expensive (faculty salaries track dentist salaries)
- [You can see where this is going...]
- Claim: Advent of online education -> decrease in number of faculty in U.S. -> number of research universities in the U.S. has peaked.
This talk generated a *huge* amount of controversy. Hennessy (purposefully?) ignored any actual research benefits from research faculty.
2) Salman Khan (of Khan academy) and Peter Norvig (Google fellow and online ML course instructor)
- - Online education has the potential to dramatically improve learning
- - Only need to read 1 paper on education [Bloom '84] (Bloom’s 2-sigmas paper), which I [Jared] will now sum up in one sentence: “Private tutoring improves student performance by two standard deviations over lecture-based instruction”
- If can “automate” effects of private tutoring, through interactive educational technology, will dramatically improve student performance.
- Example: Students who answer a question incorrectly in Stanford online class can be clustered, and pointed to information that can help them improve answer
- Norvig: motivation is extremely crucial in online classes. His primary focus is keeping students motivated and in contact with small groups of other students to create social dependencies
- Khan: want to avoid too much polish – polished lectures can be boring and inauthentic.
- Physically attending a university will always be a better experience when done right. Universities need to ensure it’s “done right” so that physical attendance offers something more than what can be obtained online.
- Online education will likely replace or at least supplement traditional textbooks
Farnam Jahanian (Assistant Director of NSF for CISE)
- NSF funding for computer science has been increasing by ~6% per year for the last 5 years. (yes, really)
- Contrasts with increases of ~3% per year for most other disciplines
- Big effort to maintain these increases. Helps immensely if the CS community actively publicizes research successes – go do this.
- Feels that CS will be better shielded than most disciplines from political vagaries and educational tsunamis
Jeffrey Dean (Google Fellow and co-inventor of MapReduce)
- “Make a reliable whole out of unreliable parts” “Make a low latency whole out of variable latency parts”. Discussed clever trick ensuring low latency on a massive server farm
- Described several applications of neural nets distributed over 1 Million CPUs. Neural nets are robust and inherently distributed so work on massive networks. Massive neural nets give huge improvements in accuracy for 1) speech recognition (“improvement is equivalent to 20 years of research in the field”) and 2) image recognition (double accuracy compared to state of the art)
Shwetak Patel (UW): Created device you plug into an electrical
outlet that tracks energy usage of all devices in your house. How?
Every type of appliance generates a unique EM signature. His device
uses this noisy signal to track each device’s energy usage with
surprising accuracy. Amazingly, can do the same thing for water
(using special sensor faucet) and gas with appropriate devices. This
was really cool!
Daphne Kohller (Stanford): Machine-learning using clinical data to
detect: 1) infant health with much higher accuracy than Apgar and
other more invasive tests – her analysis considers only time signal
data for infant respiration; 2) aggressiveness of breast cancer. In
both cases, results from the ML approach are vastly more accurate than results from traditional medical tests (e.g. Apgar). Her recent focus is applying ML to online education data to automatically determine what is the appropriate information to present to students in order to help them fix common mistakes.
Leave a Comment so far
Leave a comment