Facebook Fellowship 2021 application¶
Informantion¶
Links¶
Dates¶
Applications close 1st Octoboer 2020
Reference letters due 7th October 2020
Winners announced January 2021
Fellowship duration: autumn to spring 2021
Benefits¶
Tuition and fees paid for two years
$42,000 annual stipend (for one year?), subject to local tax laws
Paid visit to FB
Existing university funding seems to not be a problem:
If you are selected to receive the award, you may be asked to decline funding (i.e. research projects, fellowships, internships, contractor work, employment) from another company or industry competitor. Usually funding from your university, Facebook internship, or related to a teaching requirement is allowed. Mandatory military employment or Facebook contractor positions are reviewed on a case-by-case basis.
Application¶
Short plain-text research summary:
500-word research summary that clearly identifies the area of focus, importance to the field, and applicability to Facebook of the anticipated research during the award (reference the available fellowships below)
Two letters of recommendation, including one from an academic advisor
CV
Note: bibliographic references from research summary can be included with the CV, and do not count towards the word count.
Research summary guidance¶
Pages on the FB site have some advice from former fellowship winners
From Daricia Wilkinson:
Paragraph 1: Introduction
Present the problem
Identify who this impacts and why this is relevant in general and more specifically relevant to the company
One sentence summarizing your idea/approach
Paragraph 2: Body
What you plan to do
How you plan to do it
What you’ve done to show you can do this (optional)
Paragraph 3: Conclusion
Contribution to the community (academic and public)
Relevance to the mission/values of the company
From Moses Namara
For a concrete research statement, there are four key questions that you should try to answer in one or two sentences before you write out a full-fledged statement:
What are you trying to do?
How is it done today, and what are the limits of the current practice?
What’s new in your approach, and why do you think it will be successful?
Who cares about it? If you are successful, what difference will it make?
Plan¶
Area of focus¶
“Truth discovery”
Aggregating information from multiple sources of unknown trustworthiness
Aim is to infer trustworthiness based on the claims from sources, and their interaction with other claims
We can then hope for more accurate information by assigning higher weight to trustworthy sources
Existing work:
Crowdsourcing, statistics (EM algorithm, probabilistic graph models)
Judgement aggregation (logical framework, axiomatic approach)
My plan:
Investigate truth discovery algorithms (or trust-based aggregation more generally?) from a theoretical point of view
Study formal models of trust/truth to determine under what conditions the truth can be found
Develop algorithms for truth discovery applicable to large scale data
Importance to the field¶
There is potential to incorporate and expand upon work from several strands of literature:
Aggregation of expert judgements studied in a logical setting in judgement aggregation. Truth discovery does away with the idea that all ‘judges’ are equally and maximally reliable
Belief revision addresses how a rational agent should change beliefs upon receiving new information which may conflict with prior beliefs. Truth discovery does this on a larger scale by considering claims from many sources – which may conflict – and aggregating them to a (hopefully) consistent viewpoint. This is also related to belief merging (which in turn is close to judgement aggregation).
Truth discovery in crowdsourcing takes a statistical or optimisation-based approach. The goal of my work is to extend truth discovery to more general domains and focus not just on algorithm development, but theoretical analysis of the truth discovery problem in full generality.
Argumentation studies the ways in which one can draw coherent conclusions when faced with a collection of attacking arguments for and against various propositions. TODO: what else can be said?
Contribution to the field:
Apply principled theoretical analysis to a problem that has received a lot of practical focus in the crowdsourcing literature
Applicability to Facebook¶
Identifying untrustworthy sources is key to tackling Fake News™
Facebook already tackle fake news in several ways; 1
It can be hard for an individual to assess the trustworthiness of online sources, given the scale of the web
Truth discovery (and related problems) can help in at least two ways:
Aggregate information from multiple sources to obtain estimates of the truth
Rank (or otherwise evaluate) sources in terms of trustworthiness, to show warnings alongside content from questionable content that may be false or misleading
Beyond fake news and user-facing content, truth discovery can help improve the quality of crowdsourced data:
Opening times, business information on Facebook pages
Machine learning training data
Anticipated research during the award¶
Analyse classes of exiting and new algorithms with respect to their truth-finding capabilities in theory
Develop algorithms for truth discovery that implement such truth-finding, and test these via software implementation
TODO¶
- Facebook application:
Mention Facebook explicitly; particularly their efforts to rank sources by trustworthiness
Emphasise the source ranking aspect of TD as well as the aggregation side
Explainable TD by argumentation
On my work so far:
Showing axioms cannot hold simultaneously
Evaluating existing algorithms
Developing formal foundations for future study
Draft statement¶
In today’s digital age, we are surrounded by ever-increasing quantities of information. It is common for multiple sources to provide information on the same topic, especially on the web and social media, and the aggregation of such information has applications in many domains, e.g. news and crowdsourcing. However, the openness of the web has implications for the reliability of information. With minimal barriers to entry, misinformed sources may provide inaccurate information by mistake, and malicious sources may provide false information to mislead others. This presents a problem for aggregation: who should we trust and what should we believe?
Accordingly, truth discovery (TD) [1] has arisen as a popular research area within the crowdsourcing literature, focussing on how to jointly estimate trustworthiness of sources and reliability of information. The guiding principle is that information provided by trustworthy sources is likely to be reliable, and a source providing reliable information is likely to be trustworthy. TD algorithms obtain both high quality aggregated information by filtering out low-quality sources, and a ranking of sources in terms of trustworthiness. This ranking can be further used to prioritise future information from trusted sources, such as Facebook have done since 2018. [2]
While many TD algorithms have been proposed, important issues remain. For example, in what sense do these algorithms actually find the truth? What do trustworthiness scores really represent? How untrustworthy do sources need to be before the truth can no longer be found? A lack of theoretical analysis means the notions of trust and truth at play are often left unspecified. For example, many algorithms reward sources in agreement with others sources, and thus take oft-repeated claims to be reliable. However, this brings the risk of simply finding consensus as opposed to truth. On a more fundamental level, the existing literature lacks a unifying framework in which truth these questions and others can be studied in full generality, independently of any particular algorithmic approach.
In my research I aim to develop formal models of truth discovery to address such theoretical issues. By setting out a general framework in which different TD algorithms can be modelled, it will be possible to compare approaches based on their truth-finding abilities in a precise sense. This may offer additional insight beyond that which can be obtained from purely empirical analysis. For example, it may be that different methods find the truth in different models or under different conditions; such information will be valuable when deciding which algorithms to use in practise.
My work so far has applied the tools of computational social choice [3] – in particular the axiomatic approach, [4] also popular in computational economics – and argumentation [5] to study truth discovery by outlining ‘reasonable’ properties which may be used to evaluate TD algorithms, and through exchange of arguments. In future research I plan to further develop these approaches and apply the theoretical insights gained to develop new TD algorithms with strong theoretical backing. In particular, argumentation-based approaches show promise for developing explainable algorithms. [6]
~ words