NeurIPS 2023 highlights breadth of Microsoft’s machine learning innovation

Microsoft is proud to sponsor the 37th Conference on Neural Information Processing Systems (NeurIPS 2023). This interdisciplinary forum brings together experts in machine learning, neuroscience, statistics, optimization, computer vision, natural language processing, life sciences, natural sciences, social sciences, and other adjacent fields. We are pleased to share that Microsoft has over 100 accepted papers and is offering 18 workshops at NeurIPS 2023. 

This year’s conference includes three papers from Microsoft that were chosen for oral presentations, which feature groundbreaking concepts, methods, or applications, addressing pressing issues in the field. Additionally, our spotlights posters, also highlighted below, have been carefully curated by conference organizers, exhibiting novelty, technical rigor, and the potential to significantly impact the landscape of machine learning. This blog post celebrates those achievements.

Oral Presentations

Bridging Discrete and Backpropagation: Straight-Through and Beyond

Gradient computations are pivotal in deep learning’s success, yet they predominantly depend on backpropagation, a technique limited to continuous variables. The paper Bridging Discrete and Backpropagation: Straight-Through and Beyond, tackles this limitation. It introduces ReinMax, extending backpropagation’s capability to estimate gradients for models incorporating discrete variable sampling. Within extensive experiments of this study, ReinMax demonstrates consistent and significant performance gain over the state of the art. More than just a practical solution, the paper sheds light on existing deep learning practices. It elucidates that the ‘Straight-Through’ method, once considered merely a heuristic trick, is actually a viable first-order approximation for the general multinomial case. Correspondingly, ReinMax achieves second-order accuracy in this context without the complexities of second-order derivatives, thus having negligible computation overheads. 

The MineRL BASALT Competition on Learning from Human Feedback

The growth of deep learning research, including its incorporation into commercial products, has created a new challenge: How can we build AI systems that solve tasks when a crisp, well-defined specification is lacking? To encourage research on this important class of techniques, researchers from Microsoft led The MineRL BASALT Competition on Learning from Human Feedback (opens in new tab), an update to a contest first launched in 2021 (opens in new tab) by researchers at the University of California-Berkeley and elsewhere. The challenge of this competition was to complete fuzzy tasks from English language descriptions alone, with emphasis on encouraging different ways of learning from human feedback as an alternative to a traditional reward signal. 

The researchers designed a suite of four tasks in Minecraft for which writing hardcoded reward functions would be difficult. These tasks are defined by natural language: for example, “create a waterfall and take a scenic picture of it”, with additional clarifying details. Participants must train a separate agent for each task. Agents are then evaluated by humans who have read the task description.

The competition aimed to encourage development of AI systems that do what their designers intended, even when the intent cannot be easily formalized. Besides allowing AI to solve more tasks, this can also enable more effective regulation of AI systems, as well as making progress on value alignment problems, in which the specified objectives of an AI agent differ from those of its users.

Related

DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models

This comprehensive evaluation platform aims to answer the question: How trustworthy are generative pre-trained transformer (GPT) models? In DecodingTrust: A Comprehensive Assessment of Trustworthiness in GPT Models, researchers focus specifically on GPT-4, GPT-3.5, and a series of open LLMs. They consider diverse perspectives, including toxicity, stereotype bias, adversarial robustness, out-of-distribution robustness, robustness on adversarial demonstrations, privacy, machine ethics, and fairness.

The researchers’ evaluations identified previously unpublished vulnerabilities relating to trustworthiness. The team worked with Microsoft product groups to confirm that the potential vulnerabilities identified do not impact current customer-facing services. This is in part true because finished AI applications apply a range of mitigation approaches to address potential harms that may occur at the model level of the technology. They also shared their findings with GPT’s developer, OpenAI, which has noted the potential vulnerabilities in the system cards for relevant models.

This research aims to encourage others in the research community to utilize and build upon this work, potentially pre-empting adversaries who would exploit vulnerabilities to cause harm. To facilitate collaboration, the benchmark code is very extensible and easy to use: a single command is sufficient to run the complete evaluation on a new model.

Spotlight Posters

Differentially Private Approximate Near Neighbor Counting in High Dimensions

Differential privacy (DP) is a widely used tool for preserving the privacy of sensitive personal information. It allows a data structure to provide approximate answers to queries about the data it holds, while ensuring that the removal or addition of a single database entry does not significantly affect the outcome of any analysis.

Range counting (counting the number of data points falling into a given query ball) under differential privacy has been studied extensively. However, current algorithms for this problem come with challenges. One class of algorithms suffers from an additive error that is a fixed polynomial in the number of points. Another class of algorithms allows for polylogarithmic additive error, but the error grows exponentially in the dimension. To achieve the latter, the problem is relaxed to allow a “fuzzy” definition of the range boundary, e.g., a count of the points in a ball of radius r might also include points in a ball of radius cr for some c > 1.

In Differentially Private Approximate Near Neighbor Counting in High Dimensions, researchers present an efficient algorithm that offers a sweet spot between these two classes. The algorithm has an additive error that is an arbitrary small power of the data set size, depending on how fuzzy the range boundary is, as well as a small (1 + o(1)) multiplicative error. Crucially, the amount of noise added has no dependence on the dimension. This new algorithm introduces a variant of Locality-Sensitive Hashing, utilizing it in a novel manner.

Exposing Attention Glitches with Flip-Flop Language Modeling

Why do large language models sometimes output factual inaccuracies and exhibit erroneous reasoning? The brittleness of these models, particularly when executing long chains of reasoning, seems to be an inevitable price to pay for their advanced capabilities of coherently synthesizing knowledge, pragmatics, and abstract thought.

To help make sense of this fundamentally unsolved problem, Exposing Attention Glitches with Flip-Flop Language Modeling identifies and analyzes the phenomenon of attention glitches, in which the Transformer architecture’s inductive biases intermittently fail to capture robust reasoning. To isolate the issue, the researchers introduce flip-flop language modeling (FFLM), a parametric family of synthetic benchmarks designed to probe the extrapolative behavior of neural language models. This simple generative task requires a model to copy binary symbols over long-range dependencies, ignoring the tokens in between. This research shows how Transformer FFLMs suffer from a long tail of sporadic reasoning errors, some of which can be eliminated using various regularization techniques. The preliminary mechanistic analyses show why the remaining errors may be very difficult to diagnose and resolve. The researchers hypothesize that attention glitches account for some of the closed-domain errors occurring in natural LLMs.

In-Context Learning Unlocked for Diffusion Models

An emergent behavior of large language models (LLMs) is the ability to learn from context, or in-context learning. With a properly designed prompt structure and in-context learning, LLMs can combine the pre-training of multiple language tasks and generalize well to previously unseen tasks. While in-context learning has been extensively studied in natural language processing (NLP), its applications in the field of computer vision are still limited.

In-Context Learning Unlocked for Diffusion Models presents Prompt Diffusion, a framework for enabling in-context learning in diffusion-based generative models. Given a pair of task-specific example images and text guidance, this model understands the underlying task and performs the same task on a new query image following the text guidance. To achieve this, the researchers propose a vision-language prompt that can model a wide range of vision-language tasks, and a diffusion model that takes it as input. The diffusion model is trained jointly over six different tasks using these prompts. The resulting Prompt Diffusion model is the first diffusion-based vision-language foundation model capable of in-context learning. It demonstrates high-quality in-context generation on the trained tasks and generalizes to new, unseen vision tasks with their respective prompts. This model also shows compelling text-guided image editing results.

Optimizing Prompts for Text-to-Image Generation

Generative foundation models can be prompted to follow user instructions, including language models and text-to-image models. Well-designed prompts can guide text-to-image models to generate amazing images. However, the performant prompts are often model-specific and misaligned with user input. Instead of laborious human engineering, Optimizing Prompts for Text-to-Image Generation proposes prompt adaptation, a general framework that automatically adapts original user input to model-preferred prompts.

The researchers use reinforcement learning to explore better prompts with a language model. They define a reward function that encourages the policy network (i.e., language model) to generate more aesthetically pleasing images while preserving the original user intentions. Experimental results on Stable Diffusion show that this method outperforms manual prompt engineering in terms of both automatic metrics and human preference ratings. Reinforcement learning further boosts performance, especially on out-of-domain prompts.

Pareto Frontiers in Neural Feature Learning: Data, Compute, Width, and Luck

Algorithm design in deep learning can appear to be more like “hacking” than an engineering practice. There are numerous architectural choices and training heuristics, which can often modulate model performance and resource costs in unpredictable and entangled ways. As a result, when training large-scale neural networks (such as state-of-the-art language models), algorithmic decisions and resource allocations are foremost empirically-driven, involving the measurement and extrapolation of scaling laws. A precise mathematical understanding of this process is elusive, and cannot be explained by statistics or optimization in isolation.

In Pareto Frontiers in Neural Feature Learning: Data, Compute, Width, and Luck, researchers from Microsoft, Harvard, and the University of Pennsylvania explore these algorithmic intricacies and tradeoffs through the lens of a single synthetic task: the finite-sample sparse parity learning problem. In this setting, the above complications are not only evident, but also provable: intuitively, due to the task’s computational hardness, a neural network needs a sufficient combination of resources (“data × model size × training time × luck”) to succeed. This research shows that standard algorithmic choices in deep learning give rise to a Pareto frontier, in which successful learning is “bought” with interchangeable combinations of these resources. They show that algorithmic improvements on this toy problem can transfer to the real world, improving the data-efficiency of neural networks on small tabular datasets.

PDE-Refiner: Achieving Accurate Long Rollouts with Neural PDE Solvers

Time-dependent partial differential equations (PDEs) are ubiquitous in science and engineering. The high computational cost of traditional solution techniques has spurred increasing interest in deep neural network based PDE surrogates. The practical utility of such neural PDE solvers depends on their ability to provide accurate, stable predictions over long time horizons, which is a notoriously hard problem.

PDE-Refiner: Achieving Accurate Long Rollouts with Neural PDE Solvers presents a large-scale analysis of common temporal rollout strategies, identifying the neglect of non-dominant spatial frequency information, often associated with high frequencies in PDE solutions, as the primary pitfall limiting stable, accurate rollout performance. Motivated by recent advances in diffusion models, the researchers developed PDE-Refiner, a novel model class that enables more accurate modeling of all frequency components via a multistep refinement process. They validate PDE-Refiner on challenging benchmarks of complex fluid dynamics, demonstrating stable and accurate rollouts that consistently outperform state-of-the-art models, including neural, numerical, and hybrid neural-numerical architectures. They also demonstrate that PDE-Refiner greatly enhances data efficiency, since the denoising objective implicitly induces a novel form of spectral data augmentation. Finally, PDE-Refiner’s connection to diffusion models enables an accurate and efficient assessment of the model’s predictive uncertainty, allowing researchers to estimate when the surrogate becomes inaccurate.

Should I Stop or Should I Go: Early Stopping with Heterogeneous Populations

Randomized experiments are the gold-standard method of determining causal effects, whether in clinical trials to evaluate medical treatments or in A/B tests to evaluate online product offerings. But randomized experiments often need to be stopped prematurely when the treatment or test causes an unintended harmful effect. Existing methods that determine when to stop an experiment early are typically applied to the data in aggregate and do not account for treatment effect heterogeneity.

Should I Stop or Should I Go: Early Stopping with Heterogeneous Populations examines the early stopping of experiments for harm on heterogeneous populations. The paper shows that current methods often fail to stop experiments when the treatment harms a minority group of participants. The researchers use causal machine learning to develop Causal Latent Analysis for Stopping Heterogeneously (CLASH), the first broadly-applicable method for heterogeneous early stopping. They demonstrate CLASH’s performance on simulated and real data and show that it yields effective early stopping for both clinical trials and A/B tests.

Survival Instinct in Offline Reinforcement Learning

In offline reinforcement learning (RL), an agent optimizes its performance given an offline dataset. Survival Instinct in Offline Reinforcement Learning presents a novel observation: on many benchmark datasets, offline RL can produce well-performing and safe policies even when trained with “wrong” reward labels, such as those that are zero everywhere or are negatives of the true rewards. This phenomenon cannot be easily explained by offline RL’s return maximization objective. Moreover, it gives offline RL a degree of robustness that is uncharacteristic of its online RL counterparts, which are known to be sensitive to reward design.

This research demonstrates that this surprising robustness property is attributable to an interplay between the notion of pessimism in offline RL algorithms and a certain bias implicit in common data collection practices. This work shows that this pessimism endows the agent with a “survival instinct”, i.e., an incentive to stay within the data support in the long term, while the limited and biased data coverage further constrains the set of survival policies. The researchers argue that the survival instinct should be taken into account when interpreting results from existing offline RL benchmarks and when creating future ones.

Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics

Molecular dynamics (MD) is a well-established technique for simulating physical systems at the atomic level. When performed accurately, it provides unrivalled insight into the detailed mechanics of molecular motion, without the need for wet lab experiments. MD is often used to compute equilibrium properties, which requires sampling from an equilibrium distribution such as the Boltzmann distribution (opens in new tab). However, many important processes, such as binding and folding, occur over timescales of milliseconds or beyond, and cannot be efficiently sampled with conventional MD. Furthermore, new MD simulations need to be performed from scratch for each molecular system studied.

Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics presents an enhanced sampling method which uses a normalizing flow as a proposal distribution in a Markov chain Monte Carlo method targeting the Boltzmann distribution. The flow is trained offline on MD trajectories and learns to make large steps in time, simulating the molecular dynamics of 10^5−10^6fs. Crucially, Timewarp is transferable between molecular systems: the researchers show that, once trained, Timewarp generalizes to unseen small peptides (2-4 amino acids), exploring their metastable states and providing wall-clock acceleration when sampling compared to standard MD. This new method constitutes an important step towards developing general, transferable algorithms for accelerating MD.

（在新选项卡中打开）