-
Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought
Authors:
James Chua,
Edward Rees,
Hunar Batra,
Samuel R. Bowman,
Julian Michael,
Ethan Perez,
Miles Turpin
Abstract:
While chain-of-thought prompting (CoT) has the potential to improve the explainability of language model reasoning, it can systematically misrepresent the factors influencing models' behavior--for example, rationalizing answers in line with a user's opinion without mentioning this bias. To mitigate this biased reasoning problem, we introduce bias-augmented consistency training (BCT), an unsupervis…
▽ More
While chain-of-thought prompting (CoT) has the potential to improve the explainability of language model reasoning, it can systematically misrepresent the factors influencing models' behavior--for example, rationalizing answers in line with a user's opinion without mentioning this bias. To mitigate this biased reasoning problem, we introduce bias-augmented consistency training (BCT), an unsupervised fine-tuning scheme that trains models to give consistent reasoning across prompts with and without biasing features. We construct a suite testing nine forms of biased reasoning on seven question-answering tasks, and find that applying BCT to GPT-3.5-Turbo with one bias reduces the rate of biased reasoning by 86% on held-out tasks. Moreover, this model generalizes to other forms of bias, reducing biased reasoning on held-out biases by an average of 37%. As BCT generalizes to held-out biases and does not require gold labels, this method may hold promise for reducing biased reasoning from as-of-yet unknown biases and on tasks where supervision for ground truth reasoning is unavailable.
△ Less
Submitted 8 March, 2024;
originally announced March 2024.
-
The minimal computational substrate of fluid intelligence
Authors:
Amy PK Nelson,
Joe Mole,
Guilherme Pombo,
Robert J Gray,
James K Ruffle,
Edgar Chan,
Geraint E Rees,
Lisa Cipolotti,
Parashkev Nachev
Abstract:
The quantification of cognitive powers rests on identifying a behavioural task that depends on them. Such dependence cannot be assured, for the powers a task invokes cannot be experimentally controlled or constrained a priori, resulting in unknown vulnerability to failure of specificity and generalisability. Evaluating a compact version of Raven's Advanced Progressive Matrices (RAPM), a widely use…
▽ More
The quantification of cognitive powers rests on identifying a behavioural task that depends on them. Such dependence cannot be assured, for the powers a task invokes cannot be experimentally controlled or constrained a priori, resulting in unknown vulnerability to failure of specificity and generalisability. Evaluating a compact version of Raven's Advanced Progressive Matrices (RAPM), a widely used clinical test of fluid intelligence, we show that LaMa, a self-supervised artificial neural network trained solely on the completion of partially masked images of natural environmental scenes, achieves human-level test scores a prima vista, without any task-specific inductive bias or training. Compared with cohorts of healthy and focally lesioned participants, LaMa exhibits human-like variation with item difficulty, and produces errors characteristic of right frontal lobe damage under degradation of its ability to integrate global spatial patterns. LaMa's narrow training and limited capacity -- comparable to the nervous system of the fruit fly -- suggest RAPM may be open to computationally simple solutions that need not necessarily invoke abstract reasoning.
△ Less
Submitted 14 August, 2023;
originally announced August 2023.
-
Deep forecasting of translational impact in medical research
Authors:
Amy PK Nelson,
Robert J Gray,
James K Ruffle,
Henry C Watkins,
Daniel Herron,
Nick Sorros,
Danil Mikhailov,
M. Jorge Cardoso,
Sebastien Ourselin,
Nick McNally,
Bryan Williams,
Geraint E. Rees,
Parashkev Nachev
Abstract:
The value of biomedical research--a $1.7 trillion annual investment--is ultimately determined by its downstream, real-world impact. Current objective predictors of impact rest on proxy, reductive metrics of dissemination, such as paper citation rates, whose relation to real-world translation remains unquantified. Here we sought to determine the comparative predictability of future real-world trans…
▽ More
The value of biomedical research--a $1.7 trillion annual investment--is ultimately determined by its downstream, real-world impact. Current objective predictors of impact rest on proxy, reductive metrics of dissemination, such as paper citation rates, whose relation to real-world translation remains unquantified. Here we sought to determine the comparative predictability of future real-world translation--as indexed by inclusion in patents, guidelines or policy documents--from complex models of the abstract-level content of biomedical publications versus citations and publication meta-data alone. We develop a suite of representational and discriminative mathematical models of multi-scale publication data, quantifying predictive performance out-of-sample, ahead-of-time, across major biomedical domains, using the entire corpus of biomedical research captured by Microsoft Academic Graph from 1990 to 2019, encompassing 43.3 million papers across all domains. We show that citations are only moderately predictive of translational impact as judged by inclusion in patents, guidelines, or policy documents. By contrast, high-dimensional models of publication titles, abstracts and metadata exhibit high fidelity (AUROC > 0.9), generalise across time and thematic domain, and transfer to the task of recognising papers of Nobel Laureates. The translational impact of a paper indexed by inclusion in patents, guidelines, or policy documents can be predicted--out-of-sample and ahead-of-time--with substantially higher fidelity from complex models of its abstract-level content than from models of publication meta-data or citation metrics. We argue that content-based models of impact are superior in performance to conventional, citation-based measures, and sustain a stronger evidence-based claim to the objective measurement of translational potential.
△ Less
Submitted 17 October, 2021;
originally announced October 2021.
-
Structured illumination microscopy with extended axial resolution through mirrored illumination
Authors:
James D. Manton,
Florian Ströhl,
Reto Fiolka,
Clemens F. Kaminski,
Eric J. Rees
Abstract:
Wide-field fluorescence microscopy, while much faster than confocal microscopy, suffers from a lack of optical sectioning and poor axial resolution. 3D structured illumination microscopy (SIM) has been demonstrated to provide optical sectioning and to double the achievable resolution both laterally and axially, but even with this the axial resolution is still worse than the lateral resolution of u…
▽ More
Wide-field fluorescence microscopy, while much faster than confocal microscopy, suffers from a lack of optical sectioning and poor axial resolution. 3D structured illumination microscopy (SIM) has been demonstrated to provide optical sectioning and to double the achievable resolution both laterally and axially, but even with this the axial resolution is still worse than the lateral resolution of unmodified wide-field detection. Interferometric schemes using two high numerical aperture objectives, such as 4Pi confocal and I5S microscopy, have improved the axial resolution beyond that of the lateral, but at the cost of a significantly more complex optical setup. Here we investigate a simpler dual-objective scheme which we propose can be easily added to an existing 3D-SIM microscope, providing lateral and axial resolutions in excess of 125 nm with conventional fluorophores.
△ Less
Submitted 10 October, 2018;
originally announced October 2018.
-
Experiences with efficient methodologies for teaching computer programming to geoscientists
Authors:
Christian T. Jacobs,
Gerard J. Gorman,
Huw E. Rees,
Lorraine Craig
Abstract:
Computer programming was once thought of as a skill required only by professional software developers. But today, given the ubiquitous nature of computation and data science it is quickly becoming necessary for all scientists and engineers to have at least a basic knowledge of how to program. Teaching how to program, particularly to those students with little or no computing background, is well-kn…
▽ More
Computer programming was once thought of as a skill required only by professional software developers. But today, given the ubiquitous nature of computation and data science it is quickly becoming necessary for all scientists and engineers to have at least a basic knowledge of how to program. Teaching how to program, particularly to those students with little or no computing background, is well-known to be a difficult task. However, there is also a wealth of evidence-based teaching practices for teaching programming skills which can be applied to greatly improve learning outcomes and the student experience. Adopting these practices naturally gives rise to greater learning efficiency - this is critical if programming is to be integrated into an already busy geoscience curriculum. This paper considers an undergraduate computer programming course, run during the last 5 years in the Department of Earth Science and Engineering at Imperial College London. The teaching methodologies that were used each year are discussed alongside the challenges that were encountered, and how the methodologies affected student performance. Anonymised student marks and feedback are used to highlight this, and also how the adjustments made to the course eventually resulted in a highly effective learning environment.
△ Less
Submitted 9 June, 2016; v1 submitted 20 May, 2015;
originally announced May 2015.
-
On a paper by Y.G.Zarhin
Authors:
Elmer Rees
Abstract:
Zarhin showed that a matrix constructed from a polynomial with distinct roots has co-rank one. Some striking properties of this matrix are used to give a direct proof of his result. An account is given of calculations carried out to try to understand the analogous matrix for a polynomial with multiple roots and a conjecture about its rank is stated.
Zarhin showed that a matrix constructed from a polynomial with distinct roots has co-rank one. Some striking properties of this matrix are used to give a direct proof of his result. An account is given of calculations carried out to try to understand the analogous matrix for a polynomial with multiple roots and a conjecture about its rank is stated.
△ Less
Submitted 15 May, 2011;
originally announced May 2011.
-
Frobenius $n$-homomorphisms, transfers and branched coverings
Authors:
V. M. Buchstaber,
E. G. Rees
Abstract:
The main purpose is to characterise continuous maps that are $n$-branched coverings in terms of induced maps on the rings of functions. The special properties of Frobenius $n$-homomorphisms between two function spaces that correspond to $n$-branched coverings are determined completely. Several equivalent definitions of a Frobenius $n$-homomorphism are compared and some of their properties are pr…
▽ More
The main purpose is to characterise continuous maps that are $n$-branched coverings in terms of induced maps on the rings of functions. The special properties of Frobenius $n$-homomorphisms between two function spaces that correspond to $n$-branched coverings are determined completely. Several equivalent definitions of a Frobenius $n$-homomorphism are compared and some of their properties are proved. An axiomatic treatment of $n$-transfers is given in general and properties of $n$-branched coverings are studied and compared with those of regular coverings.
△ Less
Submitted 4 August, 2006;
originally announced August 2006.
-
Rings of continuous functions, symmetric products, and Frobenius algebras
Authors:
V. M. Buchstaber,
E. G. Rees
Abstract:
Properties of higher characters are developed and applied to symmetric products and Frobenius algebras. A `constructive' proof of the Gel'fand-Kolmogorov theorem is given. Generalisations of that theorem and the Nullstellensatz to symmetric products are discussed.Applications to the theory of multi-symmetric functions are also discussed. It is proved that the first three characters determine the…
▽ More
Properties of higher characters are developed and applied to symmetric products and Frobenius algebras. A `constructive' proof of the Gel'fand-Kolmogorov theorem is given. Generalisations of that theorem and the Nullstellensatz to symmetric products are discussed.Applications to the theory of multi-symmetric functions are also discussed. It is proved that the first three characters determine the Jordan algebra associated to a Frobenius algebra and as a corollary one obtains the theorem of Hoehnke and Johnson that a finite group is determined by the first three characters of its regular representation.
△ Less
Submitted 22 March, 2004;
originally announced March 2004.
-
The Gelfand map and symmetric products
Authors:
V. M. Buchstaber,
E. G. Rees
Abstract:
If A is an algebra of functions on X, there are many cases when X can be regarded as included in Hom(A,C) as the set of ring homomorphisms. In this paper the corresponding results for the symmetric products of X are introduced. It is shown that the symmetric product Sym^n(X) is included in Hom(A,C) as the set of those functions that satisfy equations generalising f(xy)=f(x)f(y). These equations…
▽ More
If A is an algebra of functions on X, there are many cases when X can be regarded as included in Hom(A,C) as the set of ring homomorphisms. In this paper the corresponding results for the symmetric products of X are introduced. It is shown that the symmetric product Sym^n(X) is included in Hom(A,C) as the set of those functions that satisfy equations generalising f(xy)=f(x)f(y). These equations are related to formulae introduced by Frobenius and, for the relevant A, they characterise linear maps on A that are the sum of ring homomorphisms. The main theorem is proved using an identity satisfied by partitions of finite sets.
△ Less
Submitted 18 September, 2001;
originally announced September 2001.
-
Collector Failures on 350 MHz, 1.2 MW CW Klystrons at the Low Energy Demonstration Accelerator (LEDA)
Authors:
D. Rees,
W. Roybal,
J. Bradley
Abstract:
We are currently operating the front end of the accelerator production of tritium (APT) accelerator, a 7 MeV radio frequency quadrapole (RFQ) using three, 1.2 MW CW klystrons. These klystrons are required and designed to dissipate the full beam power in the collector. The klystrons have less than 1500 operational hours. One collector has failed and all collectors are damaged. This paper will dis…
▽ More
We are currently operating the front end of the accelerator production of tritium (APT) accelerator, a 7 MeV radio frequency quadrapole (RFQ) using three, 1.2 MW CW klystrons. These klystrons are required and designed to dissipate the full beam power in the collector. The klystrons have less than 1500 operational hours. One collector has failed and all collectors are damaged. This paper will discuss the damage and the difficulties in diagnosing the cause. The collector did not critically fail. Tube operation was still possible and the klystron operated up to 70% of full beam power with excellent vacuum. The indication that finally led us to the collector failure was variable emission. This information will be discussed. A hydrophonic system was implemented to diagnose collector heating. The collectors are designed to allow for mixed-phase cooling and with the hydrophonic test equipment we are able to observe: normal, single-phase cooling, mixed-phase cooling, and a hard boil. These data will be presented. The worst case beam profile from a collector heating standpoint is presented. The paper will also discuss the steps taken to halt the collector damage on the remaining 350 MHz klystrons and design changes that are being implemented to correct the problem.
△ Less
Submitted 15 August, 2000;
originally announced August 2000.
