The capability of eliciting high quality information is crucial for many applications. For instance, the quality of a trained machine learning depends heavily on the quality of its input data. We’d like to understand the capability and power of elicitation and aggregation approaches, in several critical application settings, including building forecasting system, reproducibility scoring system, and peer review/grading systems.
Hybrid Forecasting Competition
This is an IARPA funded project aiming to build a hybrid forecasting system that is able to acquire high quality predictions from human participants and create an active interaction loop between human participants and machine algorithms. A set of relevant questions or challenges include: eliciting high quality prediction from human participants without waiting for the outcome of event to realize; engaging human participation in forecasting long-term events; information aggregation/verification without ground truth; eliciting other useful information from human participants.
Information Market for Reproducibility Scoring
The reproducibility crisis has largely bothered different research community. Reproducing the results from scratch is very expensive. Even if it is affordable to carry out the reproducibility studies, it’s very easy to ask the question that how reproducible these reproducibility studies are. Machine learning algorithms wouldn’t be able to receive enough high quality training labels to build automatic predictors. We seek a cheap way to gather human judgement information to help us label the reproducibility of a particular piece of research article. The involved questions include: Build a combinatorial prediction market to elicit opinion for reproducibility Aggregate information to generate confidence scores for each research article Robust aggregation in face of uninformative information Truth discovery when the majority opinion is wrong
This is a jointly funded project with Yiling Chen at Harvard.
No training data is perfectly clean, even if it’s close to be so. Take ImageNet as an example. While people often taken ImageNet as the ground truth data, it is not always the case. Aggregated labels from people contain different (endogenous or exogenous) sources of noises. This is bothering - without understanding the noises in the training data, a good number of results are only validated upon a set of biased ones.
We argue that by default a learning algorithm should assume the existence of noise (then the clean setting corresponds to the case when there is zero noise). This project aims to first understand the robustness of machine learning models under noisy and adversarial inputs. We will then develop approaches to learn a robust model via learning and leveraging the knowledges of the noises.
The fast progress of machine learning techniques have also raised concerns on their fairness, transparency, and accountability when the outcomes can significantly affect people and our society. For instance, many times those above mentioned issues are due to the existing bias in the collected data. Machine learning systems are known to have the property of “garbage-in-garbage-out”. When the collected training data is sampled in a biased way that misrepresents a population, the only thing a machine learning algorithm can do is to reinforce this bias.
There is a series of questions awaiting answers such as how to make a machine-made decision fair, how to explain a machine-made decision to people, and who should hold accountability when an error is made by the machine.