7:00 – ASRC NASA/PhD
- Ted Radio Hour interview with Margaret Heffernan, who spoke about her book, Willful Blindness:
- “Companies that have been studied for willful blindness can be asked questions like, are there issues at work that people are afraid to raise? And when academics have done studies like this of corporations in the United States, what they find is 85 percent of people say yes. Eighty-five percent of people know there’s a problem, but they won’t say anything. And when I duplicated the research in Europe, asking all the same questions, I found exactly the same number. And what’s really interesting is that when I go to companies in Switzerland, they tell me this is a uniquely Swiss problem. And when I go to Germany, they say, oh yes, this is the German disease. And when I go to companies in England they say, oh yeah, the British are really bad at this. And the truth is, this is a human problem. We’re all, under certain circumstances, willfully blind.”
- I’ve been thinking about this a lot because when I say, well, why don’t people speak up? What I get is, oh, it’s the culture. And I think, well, what is the culture? The culture is the accumulation of everybody’s actions. And in many of the organizations I work with, change starts in very unexpected places because people just decide, I want to do this or I want to try this. And then they discover they don’t get shot. And then they discover that, actually, now, they’ve got a really exciting project. You know, I think the most dangerous thing in organizations is silence. It’s all those brains whizzing around full of observations and insight and ideas that are not being articulated.
- I think that that the 15% who do speak out are Nomads. They are mis-aligned with the culture and as such it’s 1) Easier to see problems and solutions. 2) an inability to not behave independently.
- Bayesian Layers: A Module for Neural Network Uncertainty
- We describe Bayesian Layers, a module designed for fast experimentation with neural network uncertainty. It extends neural network libraries with layers capturing uncertainty over weights (Bayesian neural nets), pre-activation units (dropout), activations (“stochastic output layers”), and the function itself (Gaussian processes). With reversible layers, one can also propagate uncertainty from input to output such as for flow-based distributions and constant-memory backpropagation. Bayesian Layers are a drop-in replacement for other layers, maintaining core features that one typically desires for experimentation. As demonstration, we fit a 10-billion parameter “Bayesian Transformer” on 512 TPUv2 cores, which replaces attention layers with their Bayesian counterpart.