-
Transformer Circuit Faithfulness Metrics are not Robust
Authors:
Joseph Miller,
Bilal Chughtai,
William Saunders
Abstract:
Mechanistic interpretability work attempts to reverse engineer the learned algorithms present inside neural networks. One focus of this work has been to discover 'circuits' -- subgraphs of the full model that explain behaviour on specific tasks. But how do we measure the performance of such circuits? Prior work has attempted to measure circuit 'faithfulness' -- the degree to which the circuit repl…
▽ More
Mechanistic interpretability work attempts to reverse engineer the learned algorithms present inside neural networks. One focus of this work has been to discover 'circuits' -- subgraphs of the full model that explain behaviour on specific tasks. But how do we measure the performance of such circuits? Prior work has attempted to measure circuit 'faithfulness' -- the degree to which the circuit replicates the performance of the full model. In this work, we survey many considerations for designing experiments that measure circuit faithfulness by ablating portions of the model's computation. Concerningly, we find existing methods are highly sensitive to seemingly insignificant changes in the ablation methodology. We conclude that existing circuit faithfulness scores reflect both the methodological choices of researchers as well as the actual components of the circuit - the task a circuit is required to perform depends on the ablation used to test it. The ultimate goal of mechanistic interpretability work is to understand neural networks, so we emphasize the need for more clarity in the precise claims being made about circuits. We open source a library at https://github.com/UFO-101/auto-circuit that includes highly efficient implementations of a wide range of ablation methodologies and circuit discovery algorithms.
△ Less
Submitted 11 July, 2024;
originally announced July 2024.
-
Self-critiquing models for assisting human evaluators
Authors:
William Saunders,
Catherine Yeh,
Jeff Wu,
Steven Bills,
Long Ouyang,
Jonathan Ward,
Jan Leike
Abstract:
We fine-tune large language models to write natural language critiques (natural language critical comments) using behavioral cloning. On a topic-based summarization task, critiques written by our models help humans find flaws in summaries that they would have otherwise missed. Our models help find naturally occurring flaws in both model and human written summaries, and intentional flaws in summari…
▽ More
We fine-tune large language models to write natural language critiques (natural language critical comments) using behavioral cloning. On a topic-based summarization task, critiques written by our models help humans find flaws in summaries that they would have otherwise missed. Our models help find naturally occurring flaws in both model and human written summaries, and intentional flaws in summaries written by humans to be deliberately misleading. We study scaling properties of critiquing with both topic-based summarization and synthetic tasks. Larger models write more helpful critiques, and on most tasks, are better at self-critiquing, despite having harder-to-critique outputs. Larger models can also integrate their own self-critiques as feedback, refining their own summaries into better ones. Finally, we motivate and introduce a framework for comparing critiquing ability to generation and discrimination ability. Our measurements suggest that even large models may still have relevant knowledge they cannot or do not articulate as critiques. These results are a proof of concept for using AI-assisted human feedback to scale the supervision of machine learning systems to tasks that are difficult for humans to evaluate directly. We release our training datasets, as well as samples from our critique assistance experiments.
△ Less
Submitted 13 June, 2022; v1 submitted 12 June, 2022;
originally announced June 2022.
-
Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models
Authors:
Aarohi Srivastava,
Abhinav Rastogi,
Abhishek Rao,
Abu Awal Md Shoeb,
Abubakar Abid,
Adam Fisch,
Adam R. Brown,
Adam Santoro,
Aditya Gupta,
Adrià Garriga-Alonso,
Agnieszka Kluska,
Aitor Lewkowycz,
Akshat Agarwal,
Alethea Power,
Alex Ray,
Alex Warstadt,
Alexander W. Kocurek,
Ali Safaya,
Ali Tazarv,
Alice Xiang,
Alicia Parrish,
Allen Nie,
Aman Hussain,
Amanda Askell,
Amanda Dsouza
, et al. (426 additional authors not shown)
Abstract:
Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-futur…
▽ More
Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.
△ Less
Submitted 12 June, 2023; v1 submitted 9 June, 2022;
originally announced June 2022.
-
WebGPT: Browser-assisted question-answering with human feedback
Authors:
Reiichiro Nakano,
Jacob Hilton,
Suchir Balaji,
Jeff Wu,
Long Ouyang,
Christina Kim,
Christopher Hesse,
Shantanu Jain,
Vineet Kosaraju,
William Saunders,
Xu Jiang,
Karl Cobbe,
Tyna Eloundou,
Gretchen Krueger,
Kevin Button,
Matthew Knight,
Benjamin Chess,
John Schulman
Abstract:
We fine-tune GPT-3 to answer long-form questions using a text-based web-browsing environment, which allows the model to search and navigate the web. By setting up the task so that it can be performed by humans, we are able to train models on the task using imitation learning, and then optimize answer quality with human feedback. To make human evaluation of factual accuracy easier, models must coll…
▽ More
We fine-tune GPT-3 to answer long-form questions using a text-based web-browsing environment, which allows the model to search and navigate the web. By setting up the task so that it can be performed by humans, we are able to train models on the task using imitation learning, and then optimize answer quality with human feedback. To make human evaluation of factual accuracy easier, models must collect references while browsing in support of their answers. We train and evaluate our models on ELI5, a dataset of questions asked by Reddit users. Our best model is obtained by fine-tuning GPT-3 using behavior cloning, and then performing rejection sampling against a reward model trained to predict human preferences. This model's answers are preferred by humans 56% of the time to those of our human demonstrators, and 69% of the time to the highest-voted answer from Reddit.
△ Less
Submitted 1 June, 2022; v1 submitted 17 December, 2021;
originally announced December 2021.
-
Truthful AI: Developing and governing AI that does not lie
Authors:
Owain Evans,
Owen Cotton-Barratt,
Lukas Finnveden,
Adam Bales,
Avital Balwit,
Peter Wills,
Luca Righetti,
William Saunders
Abstract:
In many contexts, lying -- the use of verbal falsehoods to deceive -- is harmful. While lying has traditionally been a human affair, AI systems that make sophisticated verbal statements are becoming increasingly prevalent. This raises the question of how we should limit the harm caused by AI "lies" (i.e. falsehoods that are actively selected for). Human truthfulness is governed by social norms and…
▽ More
In many contexts, lying -- the use of verbal falsehoods to deceive -- is harmful. While lying has traditionally been a human affair, AI systems that make sophisticated verbal statements are becoming increasingly prevalent. This raises the question of how we should limit the harm caused by AI "lies" (i.e. falsehoods that are actively selected for). Human truthfulness is governed by social norms and by laws (against defamation, perjury, and fraud). Differences between AI and humans present an opportunity to have more precise standards of truthfulness for AI, and to have these standards rise over time. This could provide significant benefits to public epistemics and the economy, and mitigate risks of worst-case AI futures.
Establishing norms or laws of AI truthfulness will require significant work to: (1) identify clear truthfulness standards; (2) create institutions that can judge adherence to those standards; and (3) develop AI systems that are robustly truthful.
Our initial proposals for these areas include: (1) a standard of avoiding "negligent falsehoods" (a generalisation of lies that is easier to assess); (2) institutions to evaluate AI systems before and after real-world deployment; and (3) explicitly training AI systems to be truthful via curated datasets and human interaction.
A concerning possibility is that evaluation mechanisms for eventual truthfulness standards could be captured by political interests, leading to harmful censorship and propaganda. Avoiding this might take careful attention. And since the scale of AI speech acts might grow dramatically over the coming decades, early truthfulness standards might be particularly important because of the precedents they set.
△ Less
Submitted 13 October, 2021;
originally announced October 2021.
-
Evaluating Large Language Models Trained on Code
Authors:
Mark Chen,
Jerry Tworek,
Heewoo Jun,
Qiming Yuan,
Henrique Ponde de Oliveira Pinto,
Jared Kaplan,
Harri Edwards,
Yuri Burda,
Nicholas Joseph,
Greg Brockman,
Alex Ray,
Raul Puri,
Gretchen Krueger,
Michael Petrov,
Heidy Khlaaf,
Girish Sastry,
Pamela Mishkin,
Brooke Chan,
Scott Gray,
Nick Ryder,
Mikhail Pavlov,
Alethea Power,
Lukasz Kaiser,
Mohammad Bavarian,
Clemens Winter
, et al. (33 additional authors not shown)
Abstract:
We introduce Codex, a GPT language model fine-tuned on publicly available code from GitHub, and study its Python code-writing capabilities. A distinct production version of Codex powers GitHub Copilot. On HumanEval, a new evaluation set we release to measure functional correctness for synthesizing programs from docstrings, our model solves 28.8% of the problems, while GPT-3 solves 0% and GPT-J sol…
▽ More
We introduce Codex, a GPT language model fine-tuned on publicly available code from GitHub, and study its Python code-writing capabilities. A distinct production version of Codex powers GitHub Copilot. On HumanEval, a new evaluation set we release to measure functional correctness for synthesizing programs from docstrings, our model solves 28.8% of the problems, while GPT-3 solves 0% and GPT-J solves 11.4%. Furthermore, we find that repeated sampling from the model is a surprisingly effective strategy for producing working solutions to difficult prompts. Using this method, we solve 70.2% of our problems with 100 samples per problem. Careful investigation of our model reveals its limitations, including difficulty with docstrings describing long chains of operations and with binding operations to variables. Finally, we discuss the potential broader impacts of deploying powerful code generation technologies, covering safety, security, and economics.
△ Less
Submitted 14 July, 2021; v1 submitted 7 July, 2021;
originally announced July 2021.
-
Parallel Performance of ARM ThunderX2 for Atomistic Simulation Algorithms
Authors:
William Robert Saunders,
James Grant,
Eike Hermann Müller
Abstract:
Atomistic simulation drives scientific advances in modern material science and accounts for a significant proportion of wall time on High Performance Computing facilities. It is important that algorithms are efficient and implementations are performant in a continuously diversifying hardware landscape. Furthermore, they have to be portable to make best use of the available computing resource.
In…
▽ More
Atomistic simulation drives scientific advances in modern material science and accounts for a significant proportion of wall time on High Performance Computing facilities. It is important that algorithms are efficient and implementations are performant in a continuously diversifying hardware landscape. Furthermore, they have to be portable to make best use of the available computing resource.
In this paper we assess the parallel performance of some key algorithms implemented in a performance portable framework developed by us. We consider Molecular Dynamics with short range interactions, the Fast Multipole Method and Kinetic Monte Carlo. To assess the performance of emerging architectures, we compare the Marvell ThunderX2 (ARM) architecture to traditional x86_64 hardware made available through the Azure cloud computing service.
△ Less
Submitted 20 July, 2020;
originally announced July 2020.
-
Fast electrostatic solvers for kinetic Monte Carlo simulations
Authors:
William Robert Saunders,
James Grant,
Eike Hermann Müller,
Ian Thompson
Abstract:
Kinetic Monte Carlo (KMC) is an important computational tool in physics and chemistry. In contrast to standard Monte Carlo, KMC permits the description of time dependent dynamical processes and is not restricted to systems in equilibrium. Recently KMC has been applied successfully in modelling of novel energy materials such as Lithium-ion batteries and solar cells. We consider general solid state…
▽ More
Kinetic Monte Carlo (KMC) is an important computational tool in physics and chemistry. In contrast to standard Monte Carlo, KMC permits the description of time dependent dynamical processes and is not restricted to systems in equilibrium. Recently KMC has been applied successfully in modelling of novel energy materials such as Lithium-ion batteries and solar cells. We consider general solid state systems which contain free, interacting particles which can hop between localised sites in the material. The KMC transition rates for those hops depend on the change in total potential energy of the system. For charged particles this requires the frequent calculation of electrostatic interactions, which is usually the bottleneck of the simulation. To avoid this issue and obtain results in reasonable times, many studies replace the long-range potential by a short range approximation. This, however, leads to systematic errors and unphysical results. On the other hand standard electrostatic solvers such as Ewald summation or fast Poisson solvers are highly inefficient or introduce uncontrollable systematic errors at high resolution. In this paper we describe how the Fast Multipole Method by Greengard and Rokhlin can be adapted to overcome this issue by dramatically reducing computational costs. We exploit the fact that each update in the transition rate calculation corresponds to a single particle move and changes the configuration only by a small amount. This allows us to construct an algorithm which scales linearly in the number of charges for each KMC step, something which had not been deemed to be possible before. We demonstrate the performance and parallel scalability of the method by implementing it in a performance portable software library. We describe the high-level Python interface of the code which makes it easy to adapt to specific cases.
△ Less
Submitted 1 March, 2020; v1 submitted 10 May, 2019;
originally announced May 2019.
-
Long range forces in a performance portable Molecular Dynamics framework
Authors:
William R. Saunders,
James Grant,
Eike H. Müller
Abstract:
Molecular Dynamics (MD) codes predict the fundamental properties of matter by following the trajectories of a collection of interacting model particles. To exploit diverse modern manycore hardware, efficient codes must use all available parallelism. At the same time they need to be portable and easily extendible by the domain specialist (physicist/chemist) without detailed knowledge of this hardwa…
▽ More
Molecular Dynamics (MD) codes predict the fundamental properties of matter by following the trajectories of a collection of interacting model particles. To exploit diverse modern manycore hardware, efficient codes must use all available parallelism. At the same time they need to be portable and easily extendible by the domain specialist (physicist/chemist) without detailed knowledge of this hardware. To address this challenge, we recently described a new Domain Specific Language (DSL) for the development of performance portable MD codes based on a "Separation of Concerns": a Python framework automatically generates efficient parallel code for a range of target architectures.
Electrostatic interactions between charged particles are important in many physical systems and often dominate the runtime. Here we discuss the inclusion of long-range interaction algorithms in our code generation framework. These algorithms require global communications and careful consideration has to be given to any impact on parallel scalability. We implemented an Ewald summation algorithm for electrostatic forces, present scaling comparisons for different system sizes and compare to the performance of existing codes. We also report on further performance optimisations delivered with OpenMP shared memory parallelism.
△ Less
Submitted 3 August, 2017;
originally announced August 2017.
-
Trial without Error: Towards Safe Reinforcement Learning via Human Intervention
Authors:
William Saunders,
Girish Sastry,
Andreas Stuhlmueller,
Owain Evans
Abstract:
AI systems are increasingly applied to complex tasks that involve interaction with humans. During training, such systems are potentially dangerous, as they haven't yet learned to avoid actions that could cause serious harm. How can an AI system explore and learn without making a single mistake that harms humans or otherwise causes serious damage? For model-free reinforcement learning, having a hum…
▽ More
AI systems are increasingly applied to complex tasks that involve interaction with humans. During training, such systems are potentially dangerous, as they haven't yet learned to avoid actions that could cause serious harm. How can an AI system explore and learn without making a single mistake that harms humans or otherwise causes serious damage? For model-free reinforcement learning, having a human "in the loop" and ready to intervene is currently the only way to prevent all catastrophes. We formalize human intervention for RL and show how to reduce the human labor required by training a supervised learner to imitate the human's intervention decisions. We evaluate this scheme on Atari games, with a Deep RL agent being overseen by a human for four hours. When the class of catastrophes is simple, we are able to prevent all catastrophes without affecting the agent's learning (whereas an RL baseline fails due to catastrophic forgetting). However, this scheme is less successful when catastrophes are more complex: it reduces but does not eliminate catastrophes and the supervised learner fails on adversarial examples found by the agent. Extrapolating to more challenging environments, we show that our implementation would not scale (due to the infeasible amount of human labor required). We outline extensions of the scheme that are necessary if we are to train model-free agents without a single catastrophe.
△ Less
Submitted 17 July, 2017;
originally announced July 2017.
-
A Domain Specific Language for Performance Portable Molecular Dynamics Algorithms
Authors:
William R. Saunders,
James Grant,
Eike H. Müller
Abstract:
Developers of Molecular Dynamics (MD) codes face significant challenges when adapting existing simulation packages to new hardware. In a continuously diversifying hardware landscape it becomes increasingly difficult for scientists to be experts both in their own domain (physics/chemistry/biology) and specialists in the low level parallelisation and optimisation of their codes. To address this chal…
▽ More
Developers of Molecular Dynamics (MD) codes face significant challenges when adapting existing simulation packages to new hardware. In a continuously diversifying hardware landscape it becomes increasingly difficult for scientists to be experts both in their own domain (physics/chemistry/biology) and specialists in the low level parallelisation and optimisation of their codes. To address this challenge, we describe a "Separation of Concerns" approach for the development of parallel and optimised MD codes: the science specialist writes code at a high abstraction level in a domain specific language (DSL), which is then translated into efficient computer code by a scientific programmer. In a related context, an abstraction for the solution of partial differential equations with grid based methods has recently been implemented in the (Py)OP2 library. Inspired by this approach, we develop a Python code generation system for molecular dynamics simulations on different parallel architectures, including massively parallel distributed memory systems and GPUs. We demonstrate the efficiency of the auto-generated code by studying its performance and scalability on different hardware and compare it to other state-of-the-art simulation packages. With growing data volumes the extraction of physically meaningful information from the simulation becomes increasingly challenging and requires equally efficient implementations. A particular advantage of our approach is the easy expression of such analysis algorithms. We consider two popular methods for deducing the crystalline structure of a material from the local environment of each atom, show how they can be expressed in our abstraction and implement them in the code generation framework.
△ Less
Submitted 13 November, 2017; v1 submitted 11 April, 2017;
originally announced April 2017.