How to Start Collaborating with CSE to Solve Global Health Problems
Oliver Kennedy
- CSE: Research vs Implementation?
- Research Highlights.
- Behind the Buzzwords.
- CSE Resources for Your Benefit.
Computer Science & Engineering
- How hard is a particular type of problem to solve?
- Can a particular solution be scaled to bigger problems?
- How do I solve problems that fit a particular pattern?
Abstraction
CSE research is about creating general solutions
(motivated by specific problems)
CSE Publication
How does the new solution differ? Is it...
- ... more general?
- ... more efficient/reliable?
- ... more scalable?
- ... easier to use?
Do you need research or implementation?
UB-CSE can help with both,
but each takes a different approach.
Have clear, well-defined parameters for the problem.
Typical Implementation Problems
- Organizing data into a database.
- A mobile app to display information.
- Making an R script run faster.
Implementation Resources
CSE-611
Undergraduate Research
Invenst
But...
Implementation often inspires research topics.
(Ailamaki's 7 month rule)
Implementation may synergize with existing research.
(e.g., Motivating use cases)
(so talking to a friendly CSE faculty can still be useful)
Research Highlights
- Reproducible Datasets (Oliver Kennedy)
- Wireless Sensor Networks (Chang Wen Chen)
Reproducible Datasets
Oliver Kennedy
Data Errors Suck
→
→
↓
↓
Assumption
≠
Assumption
The Vizier Notebook
If you're using Python, R, SQL, or Jupyter, we can help...
- ...improve your dataset documentation
- ...make your workflows more reproducible
- ...make your code faster
talk to me afterwards!
Wireless Sensor Networks
Chang Wen Chen
IoT Devices Need To Communicate
- Cellular or LORA Networks
- Each device talks to a tower
(reliable, but requires infrastructure) - Store and Collect
- Each device stores data and is recovered
(no infrastructure, but requires physical visits) - Mesh Networks
- Each device communicates via nearby devices
(bandwidth/power limited)
UB-CSE is at the forefront of wireless research
Application: Monitoring the excessive antibiotic discharge into Missouri river from overdose usage as agricultural runoff
Problem: 1d "mesh" is even more limited
Buzzwords
Neural Networks / Deep Learning
Linear Regression
↓
Spline Fitting
↓
Graphical Models
↓
Neural Networks
$y=f(x)$ where $f$ has 100s or 1000s (or more) DoF
- The Good
- • Feasible to fit very complex functions (e.g., face?)
- • Minimal knowledge of problem structure required.
- The Bad
- • Need huge training data
- • Very easy to overfit
- • Not explainable (yet)
Differential Privacy
A way to mathematically prove to yourself how much PII could leak if an aggregate dataset is released.
"Can I create a statistically significant effect on the dataset by removing N individuals?"
Blockchain
Oliver's Blockchain PSA
- It is statistically unlikely that you need a blockchain.
(I've yet to see a use case apart from cryptocurrency that needs a blockchain) - Proof of work is a huge power sink.
(Bitcoin alone estimated at 77 TWh this year ~= the power consumption of Chile)
Please consult a Doctor (of Philosophy in CS)
before starting a blockchain project
Resources
- CSE 611: https://invenst.cse.buffalo.edu/viewideabank.php
- Undergraduates: Any CSE Faculty
- Invenst: Alan Hunt
- Centers
- CUBS/CEDAR (image/video data)
- CARA (structured data)
- CMIF (multisource fusion)
NSF CS+X Programs
- CSSI: Cyberinfrastructure for Sustained Scientific Innovation
- SCH: Smart and Connected Health
- SCC: Smart and Connected Communities