Neural Information Processing Systems http://nips.cc/

We would like to sincerely thank you for your important ideas and constructive comments. It is not related to the deep learning domain. We will clearly state these contributions in the paper. As you suggest, we will define B2N, RAI, and GGT in the paper. Optimizing for a specific loss hinders other objectives, e.g., accuracy and calibration.

Neural Information Processing Systems http://nips.cc/

On an otherwise routine July day, inside a laboratory at Duke University, two rhesus monkeys sat in separate rooms, each watching a computer screen that featured an image of a virtual arm in two-dimensional space. The monkeys' task was to guide the arm from the center of the screen to a target, and when they did so successfully, the researchers rewarded them with sips of juice. But there was a twist. The monkeys were not provided with joysticks or any other devices that could manipulate the arm. Rather, they were relying on electrodes implanted in portions of their brains that influence movement.

Quantifying and measuring uncertainty in deep neural networks, despite recent important advances, is still an open problem. Bayesian neural networks are a powerful solution, where the prior over network weights is a design choice, often a normal distribution or other distribution encouraging sparsity. However, this prior is agnostic to the generative process of the input data, which might lead to unwarranted generalization for out-of-distribution tested data. We suggest treating the generative process of the input data as a confounder for the relation between the input and the discriminative function, thereby conditioning the prior of the network weights on the distribution of the input. We propose an algorithm for modeling this confounder through neural connectivity patterns. This approach is ultimately translated into a new deep architecture---a compact hierarchy of networks. We demonstrate that sampling networks from this hierarchy, proportionally to their posterior, is efficient and enables estimating various types of uncertainties. Empirical evaluations of our method demonstrate significant improvement compared to state-of-the-art calibration and out-of-distribution detection methods.

That moment was the culmination of two decades of work in brain-machine interface technology, a research field I pioneered with my colleagues at Duke University. Early experiments involved rats and monkeys moving levers, robots and avatar bodies using their thoughts. My colleagues and I believe that we can apply what we've learned about neuroplasticity--the ability of the brain to change over time--to a range of neurological diseases, including Parkinson's disease, epilepsy, stroke, cerebral palsy and even autism. Scientists from university labs to Silicon Valley are working on two additional ideas conceived in my lab: connecting brains to form a network, or brainet, and developing a communication method that lets people message one another directly brain-to-brain. Once brains are connected they could become a hackable system in which the thoughts and actions of connected individuals can be accessed and manipulated.

The first true brain-to-brain communication in people could start next year, thanks to huge recent advances. Early attempts won't quite resemble telepathy as we often imagine it. Our brains work in unique ways, and the way each of us thinks about a concept is influenced by our experiences and memories. This results in different patterns of brain activity, but if neuroscientists can learn one individual's patterns, they may be able to trigger certain thoughts in that person's brain. In theory, they could then use someone else's brain activity to trigger these thoughts.