First speaker: Marisol
did outreach at UC Merced - workforce development; cybersecurity CECORE course based on a Stanford class
She did 20 years in CS; satellite image analysis.
Originally from a rural area of New Mexico. Was invited to join a 5-person team for a Los Alamos challenge when she was age 15. Said it took her a year to learn how to connect using a modem, but once she was in she realized she could make a computer do anything she wanted.
Mentoring:
- Ignite the spark
- Encourage each other
- Share knowledge
Women's empowerment lunch every 2 weeks
Daphne Koller - In Sitro, Stanford
Eroom's Law: there's an exponential decrease in pharmaceutical R&D productivity takes about $2.5B and 15 years per approved new drug
Binary classification: is there a bear here?
Multiclass classification: what is in this image?
Example of how models tend to latch onto subtle artifacts: study to detect broken bones in x-ray images ended up learning differences between the x-ray machines once they corrected for that, the ability to detect fractures was actually no better than a coin toss
Panel on ethics: FAITHE Fairness, Accountability, Integrity, Transparency, Honesty, Equity
Compounding risks: black box solutions + biased data + biased humans Don't just want virtue-signalling
Big data as a source of power inequity Women are especially aware since we've been on the receiving end of inequity for so long
What's public, what's private, and who decides
Alyson Fox - Skynet
Collaborative Autonomy in energy program (edge computing)
Cyber and infrastructure resilience to protect the power grid
Want reliable computing from unreliable hardware, like in multicellular biology (idea that losing one agent in a cluster doesn't take the whole thing down)
Decentralized averaging: each agent has a piece of data
Want techniques that are robust to:
- device failure
- network failure
- calculation error
- source data problems
Who does the calculation (which agent in the cluster)?
Can the method be network-agnostic, or does it need to know the topology of the graph?
Broadcast back: how does the device know when it has all the data and it's time to calculate?
Ex: Push-Sum Consensus, see Olshevsky et al. 2018
1.Each agent has a value and a weight 2.Wait and then communicate 3.Update the estimate
Each node is k physical agents - all the machine on the node are checking each other's calculations
Need historical context and domain context
example of a geometic mean with arcsin that is robust to outliers
Creating robust, iterative linear solvers
ex. Synchronous Jacobi: has to wait for all the data for each update
Actually want something asynchronous, e.g. Asynchronous Redundant Jacobi
testing: 30 agents where 1 is 10x slower than all the others
- too much redundancy slows things down, so there's a sweet spot
Fanny Chevalier - U Toronto
Decision making!
ex. 16 shark attacks per year vs. 1730 vending machine attacks per year
Mental shortcuts:
- affect (emotion)
- assumptions
- anecdotes
- Expectations
- Perspective
Interactive visualization tools
see Doccurate: Sultanum et al. VAST 2018 Phenolines.org: Glueck et al. VAST 2017
ex. Florida Dept. Law Enforcement charts with the y-axis upside-down re: stand your ground law, but misleading
see Ritchie et al. CHI 2019
- Make the data relatable ex. use soccer fields, not hectares see Climate Change Coloring Book
instead of 39g sugar in a can of coke, that's 10 sugar cubes
- Engage with data see Dear Data book Data Ink- personal visualization tool, collaboration with MSFT Xia et al. CHI 2018 https://www.microsoft.com/en-us/research/uploads/prod/2018/05/dataink.pdf
Data Quilt Zhang et al. CHI 2018
- Teach Data C'est la vis - tool for kids, see Alper et al.
Katie Schmidt - LLNL
Sensitivity Analysis
UQ - Uncertainty Quantification
- her thesis advisor wrote one of the key textbooks on the topic
connection to Model Calibration
-
assume error is normally distributed to start with
-
seems to be describing Frequentist as using "fixed parameters" (as if we don't update models??)
big vs. small change in output when we change a parameter
when doing dimensional reduction, if parameters are non-identifiable, the optimal parameter set may not be unique (there may be more than one set that is optimal)
local sensitivity: do partial derivatives
global sensitivity: Sobol Decomposition
decompose the variance to get Sobol indices: Si
and interaction sensitivities: S1-2, S2-3 etc.
total sensitivity: add up all the components
Kelli Humbird - Design Physicist at LLNL
Inertial Confinement Fusion (ICF)
shoot a laser at a 2mm capsule of deuterium and tritium
traditionally start with simulations because experiments are $$$
ML trained on a mapping of inputs and outputs, so use NNs w/transfer learning
30k low-fidelity sims --> low fidelity NN --> add 23 high fidelity sims and do a high-fidelity NN --> add 23 experiments and do your final NN
low fidelity here == simpler physics models (fewer parameters, less computationally expensive)
looking to maximize e.g. yield x (squared area density)
- naive: high power laser, 1D assumptions, just blast it
- slightly better: account for laser degradation at high power, use a thinner capsule and lower power
- after transfer learning: use a longer pulse at lower power, account for 3D effects
see github.com/llnl/djinn & I downloaded her IEEE paper.
Tsu-Jae King Liu, dean of engineering at Berkeley (she's an EE)
** her slides were an especially good, though depressing overview of current statistics on women in tech