Search

Language models trained on diverse datasets unlock generalization by in-context learning. Reinforcement Learning (RL) policies can achieve a similar effect by meta-learning within the memory of a sequence model. However, meta-RL research primarily focuses on adapting to minor variations of a single task. It is difficult to scale towards more general behavior without confronting challenges in multi-task optimization, and few solutions are compatible with meta-RL's goal of learning from large training sets of unlabeled tasks. To address this challenge, we revisit the idea that multi-task RL is bottlenecked by imbalanced training losses created by uneven return scales across different tasks. We build upon recent advancements in Transformer-based (in-context) meta-RL and evaluate a simple yet scalable solution where both an agent's actor and critic objectives are converted to classification terms that decouple optimization from the current scale of returns. Large-scale comparisons in Meta-World ML45, Multi-Game Procgen, Multi-Task POPGym, Multi-Game Atari, and BabyAI find that this design unlocks significant progress in online multi-task adaptation and memory problems without explicit task labels., Comment: NeurIPS 2024

For any fixed feedforward ReLU neural network architecture, it is well-known that many different parameter settings can determine the same function. It is less well-known that the degree of this redundancy is inhomogeneous across parameter space. In this work, we discuss two locally applicable complexity measures for ReLU network classes and what we know about the relationship between them: (1) the local functional dimension [14, 18], and (2) a local version of VC dimension that we call persistent pseudodimension. The former is easy to compute on finite batches of points; the latter should give local bounds on the generalization gap, which would inform an understanding of the mechanics of the double descent phenomenon [7]., Comment: 41 pages

In the late stages of accretion leading up to the formation of planetesimals, particles grew to pebbles the size of 1-mm to tens of cm. That is the same size range that dominates the present-day comet mass loss. Meteoroids that size cause visible meteors on Earth. Here, we hypothesize that the size distribution and the physical and chemical properties of young meteoroid streams still contain information about the conditions in the solar nebula during these late stages of accretion. From observations of 47 young meteor showers, we find that freshly ejected meteoroids from long-period comets tend to have low bulk density and are distributed with equal surface area per log-mass interval (magnitude distribution index chi ~ 1.85), suggesting gentle accretion conditions. Jupiter-family comets, on the other hand, mostly produce meteoroids twice as dense and distributed with a steeper chi ~ 2.15 or even chi ~ 2.5, which implies that those pebbles grew from particles fragmenting in a collisional cascade or by catastrophic collisions, respectively. Both comet populations contain an admixture of compact materials that are sometimes sodium-poor, but Jupiter-family comets show a higher percentage (~8% on average) than long-period comet showers (~4%), and a wider range. While there are exceptions in both groups, the implication is that most long-period comets formed under gentle particle growth conditions, possibly near the 30 AU edge of the Trans Neptunian Disk, while most Jupiter family comets formed closer to the Sun where pebbles reached or passed the fragmentation barrier. This is possible if the Scattered Disk represents all objects scattered by Neptune during its migration, while the present-day outer Oort cloud formed only during and after the Sun had moved away from sibling stars., Comment: 82 pages, 14 figures, 3 tables

The challenge of visual grounding and masking in multimodal machine translation (MMT) systems has encouraged varying approaches to the detection and selection of visually-grounded text tokens for masking. We introduce new methods for detection of visually and contextually relevant (concrete) tokens from source sentences, including detection with natural language processing (NLP), detection with object detection, and a joint detection-verification technique. We also introduce new methods for selection of detected tokens, including shortest $n$ tokens, longest $n$ tokens, and all detected concrete tokens. We utilize the GRAM MMT architecture to train models against synthetically collated multimodal datasets of source images with masked sentences, showing performance improvements and improved usage of visual context during translation tasks over the baseline model.

While most current work in multimodal machine translation (MMT) uses the Multi30k dataset for training and evaluation, we find that the resulting models overfit to the Multi30k dataset to an extreme degree. Consequently, these models perform very badly when evaluated against typical text-only testing sets such as the WMT newstest datasets. In order to perform well on both Multi30k and typical text-only datasets, we use a performant text-only machine translation (MT) model as the starting point of our MMT model. We add vision-text adapter layers connected via gating mechanisms to the MT model, and incrementally transform the MT model into an MMT model by 1) pre-training using vision-based masking of the source text and 2) fine-tuning on Multi30k.

A good evaluation framework should evaluate multimodal machine translation (MMT) models by measuring 1) their use of visual information to aid in the translation task and 2) their ability to translate complex sentences such as done for text-only machine translation. However, most current work in MMT is evaluated against the Multi30k testing sets, which do not measure these properties. Namely, the use of visual information by the MMT model cannot be shown directly from the Multi30k test set results and the sentences in Multi30k are are image captions, i.e., short, descriptive sentences, as opposed to complex sentences that typical text-only machine translation models are evaluated against. Therefore, we propose that MMT models be evaluated using 1) the CoMMuTE evaluation framework, which measures the use of visual information by MMT models, 2) the text-only WMT news translation task test sets, which evaluates translation performance against complex sentences, and 3) the Multi30k test sets, for measuring MMT model performance against a real MMT dataset. Finally, we evaluate recent MMT models trained solely against the Multi30k dataset against our proposed evaluation framework and demonstrate the dramatic drop performance against text-only testing sets compared to recent text-only MT models.

In this paper, we give a formula for the number of permutations that avoid the split patterns $3|12$ and $23|1$ with respect to a position $r$. Such permutations count the number of Schubert varieties for which the projection map from the flag variety to a Grassmannian induces a fiber bundle structure. We also study the corresponding bivariate generating function and show how it is related to modified Bessel functions., Comment: 10 pages

We introduce AMAGO, an in-context Reinforcement Learning (RL) agent that uses sequence models to tackle the challenges of generalization, long-term memory, and meta-learning. Recent works have shown that off-policy learning can make in-context RL with recurrent policies viable. Nonetheless, these approaches require extensive tuning and limit scalability by creating key bottlenecks in agents' memory capacity, planning horizon, and model size. AMAGO revisits and redesigns the off-policy in-context approach to successfully train long-sequence Transformers over entire rollouts in parallel with end-to-end RL. Our agent is scalable and applicable to a wide range of problems, and we demonstrate its strong performance empirically in meta-RL and long-term memory domains. AMAGO's focus on sparse rewards and off-policy data also allows in-context learning to extend to goal-conditioned problems with challenging exploration. When combined with a multi-goal hindsight relabeling scheme, AMAGO can solve a previously difficult category of open-world domains, where agents complete many possible instructions in procedurally generated environments., Comment: ICLR 2024

We present a new algorithm, Cross-Episodic Curriculum (CEC), to boost the learning efficiency and generalization of Transformer agents. Central to CEC is the placement of cross-episodic experiences into a Transformer's context, which forms the basis of a curriculum. By sequentially structuring online learning trials and mixed-quality demonstrations, CEC constructs curricula that encapsulate learning progression and proficiency increase across episodes. Such synergy combined with the potent pattern recognition capabilities of Transformer models delivers a powerful cross-episodic attention mechanism. The effectiveness of CEC is demonstrated under two representative scenarios: one involving multi-task reinforcement learning with discrete control, such as in DeepMind Lab, where the curriculum captures the learning progression in both individual and progressively complex settings; and the other involving imitation learning with mixed-quality data for continuous control, as seen in RoboMimic, where the curriculum captures the improvement in demonstrators' expertise. In all instances, policies resulting from CEC exhibit superior performance and strong generalization. Code is open-sourced at https://cec-agent.github.io/ to facilitate research on Transformer agent learning., Comment: To appear in NeurIPS 2023; The first two authors contributed equally

The parameter space for any fixed architecture of feedforward ReLU neural networks serves as a proxy during training for the associated class of functions - but how faithful is this representation? It is known that many different parameter settings can determine the same function. Moreover, the degree of this redundancy is inhomogeneous: for some networks, the only symmetries are permutation of neurons in a layer and positive scaling of parameters at a neuron, while other networks admit additional hidden symmetries. In this work, we prove that, for any network architecture where no layer is narrower than the input, there exist parameter settings with no hidden symmetries. We also describe a number of mechanisms through which hidden symmetries can arise, and empirically approximate the functional dimension of different network architectures at initialization. These experiments indicate that the probability that a network has no hidden symmetries decreases towards 0 as depth increases, while increasing towards 1 as width and input dimension increase., Comment: 27 pages, 11 figures, ICML 2023

Machine learning models fail to perform when facing out-of-distribution (OOD) domains, a challenging task known as domain generalization (DG). In this work, we develop a novel DG training strategy, we call PGrad, to learn a robust gradient direction, improving models' generalization ability on unseen domains. The proposed gradient aggregates the principal directions of a sampled roll-out optimization trajectory that measures the training dynamics across all training domains. PGrad's gradient design forces the DG training to ignore domain-dependent noise signals and updates all training domains with a robust direction covering main components of parameter dynamics. We further improve PGrad via bijection-based computational refinement and directional plus length-based calibrations. Our theoretical proof connects PGrad to the spectral analysis of Hessian in training neural networks. Experiments on DomainBed and WILDS benchmarks demonstrate that our approach effectively enables robust DG optimization and leads to smoothly decreased loss curves. Empirically, PGrad achieves competitive results across seven datasets, demonstrating its efficacy across both synthetic and real-world distributional shifts. Code is available at https://github.com/QData/PGrad.

Deep reinforcement learning algorithms have succeeded in several challenging domains. Classic Online RL job schedulers can learn efficient scheduling strategies but often takes thousands of timesteps to explore the environment and adapt from a randomly initialized DNN policy. Existing RL schedulers overlook the importance of learning from historical data and improving upon custom heuristic policies. Offline reinforcement learning presents the prospect of policy optimization from pre-recorded datasets without online environment interaction. Following the recent success of data-driven learning, we explore two RL methods: 1) Behaviour Cloning and 2) Offline RL, which aim to learn policies from logged data without interacting with the environment. These methods address the challenges concerning the cost of data collection and safety, particularly pertinent to real-world applications of RL. Although the data-driven RL methods generate good results, we show that the performance is highly dependent on the quality of the historical datasets. Finally, we demonstrate that by effectively incorporating prior expert demonstrations to pre-train the agent, we short-circuit the random exploration phase to learn a reasonable policy with online training. We utilize Offline RL as a launchpad to learn effective scheduling policies from prior experience collected using Oracle or heuristic policies. Such a framework is effective for pre-training from historical datasets and well suited to continuous improvement with online data collection.

The exponential growth in demand for digital services drives massive datacenter energy consumption and negative environmental impacts. Promoting sustainable solutions to pressing energy and digital infrastructure challenges is crucial. Several hyperscale cloud providers have announced plans to power their datacenters using renewable energy. However, integrating renewables to power the datacenters is challenging because the power generation is intermittent, necessitating approaches to tackle power supply variability. Hand engineering domain-specific heuristics-based schedulers to meet specific objective functions in such complex dynamic green datacenter environments is time-consuming, expensive, and requires extensive tuning by domain experts. The green datacenters need smart systems and system software to employ multiple renewable energy sources (wind and solar) by intelligently adapting computing to renewable energy generation. We present RARE (Renewable energy Aware REsource management), a Deep Reinforcement Learning (DRL) job scheduler that automatically learns effective job scheduling policies while continually adapting to datacenters' complex dynamic environment. The resulting DRL scheduler performs better than heuristic scheduling policies with different workloads and adapts to the intermittent power supply from renewables. We demonstrate DRL scheduler system design parameters that, when tuned correctly, produce better performance. Finally, we demonstrate that the DRL scheduler can learn from and improve upon existing heuristic policies using Offline Learning., Comment: Accepted at JSSPP-2022

Much of the existing work on text novelty detection has been studied at the topic level, i.e., identifying whether the topic of a document or a sentence is novel or not. Little work has been done at the fine-grained semantic level (or contextual level). For example, given that we know Elon Musk is the CEO of a technology company, the sentence "Elon Musk acted in the sitcom The Big Bang Theory" is novel and surprising because normally a CEO would not be an actor. Existing topic-based novelty detection methods work poorly on this problem because they do not perform semantic reasoning involving relations between named entities in the text and their background knowledge. This paper proposes an effective model (called PAT-SND) to solve the problem, which can also characterize the novelty. An annotated dataset is also created. Evaluation shows that PAT-SND outperforms 10 baselines by large margins., Comment: 28 pages, 2 figures

It is well-known that the parameterized family of functions representable by fully-connected feedforward neural networks with ReLU activation function is precisely the class of piecewise linear functions with finitely many pieces. It is less well-known that for every fixed architecture of ReLU neural network, the parameter space admits positive-dimensional spaces of symmetries, and hence the local functional dimension near any given parameter is lower than the parametric dimension. In this work we carefully define the notion of functional dimension, show that it is inhomogeneous across the parameter space of ReLU neural network functions, and continue an investigation - initiated in [14] and [5] - into when the functional dimension achieves its theoretical maximum. We also study the quotient space and fibers of the realization map from parameter space to function space, supplying examples of fibers that are disconnected, fibers upon which functional dimension is non-constant, and fibers upon which the symmetry group acts non-transitively., Comment: 51 pages, 1 figure

We develop a framework to establish benchmarks for machine learning and deep neural networks analyses of exclusive scattering cross sections (FemtoNet). Within this framework we present an extraction of Compton form factors for deeply virtual Compton scattering from an unpolarized proton target. Critical to this effort is a study of the effects of physics constraint built into machine learning (ML) algorithms. We use the Bethe-Heitler process, which is the QED radiative background to deeply virtual Compton scattering, to test our ML models and, in particular, their ability to generalize information extracted from data. We then use these techniques on the full cross section and compare the results to analytic model calculations. We propose a quantification technique, the random targets method, to begin understanding the separation of aleatoric and epistemic uncertainties as they are manifest in exclusive scattering analyses. We propose a set of both physics driven and machine learning based benchmarks providing a stepping stone towards applying explainable machine learning techniques with controllable uncertainties in a wide range of deeply virtual exclusive processes., Comment: 16 pages, 12 figures

We apply a generalized piecewise-linear (PL) version of Morse theory due to Grunert-Kuhnel-Rote to define and study new local and global notions of topological complexity for fully-connected feedforward ReLU neural network functions, F: R^n -> R. Along the way, we show how to construct, for each such F, a canonical polytopal complex K(F) and a deformation retract of the domain onto K(F), yielding a convenient compact model for performing calculations. We also give a construction showing that local complexity can be arbitrarily high., Comment: 40 pages, 8 figures; Sections 5 and 6 from v1 removed from v2. We plan to use constructions in those sections for a follow-up paper with a more computational focus

As more and more AI agents are used in practice, it is time to think about how to make these agents fully autonomous so that they can (1) learn by themselves continually in a self-motivated and self-initiated manner rather than being retrained offline periodically on the initiation of human engineers and (2) accommodate or adapt to unexpected or novel circumstances. As the real-world is an open environment that is full of unknowns or novelties, the capabilities of detecting novelties, characterizing them, accommodating/adapting to them, gathering ground-truth training data and incrementally learning the unknowns/novelties become critical in making the AI agent more and more knowledgeable, powerful and self-sustainable over time. The key challenge here is how to automate the process so that it is carried out continually on the agent's own initiative and through its own interactions with humans, other agents and the environment just like human on-the-job learning. This paper proposes a framework (called SOLA) for this learning paradigm to promote the research of building autonomous and continual learning enabled AI agents. To show feasibility, an implemented agent is also described., Comment: To appear in AI Magazine (AAAI), 2023. This draft is an extended and revised version of the previous work - "Self-initiated Open World Learning for Autonomous AI Agents" arXiv preprint arXiv:2110.11385 (2021), which was published in AAAI 2022 Spring Symposium Series

Accurate understanding of users in terms of predicative segments play an essential role in the day to day operation of modern internet enterprises. Nevertheless, there are significant challenges that limit the quality of data, especially on long tail predictive tasks. In this work, we present MetaCon, our unified predicative segments system with scalable, trillion concepts meta learning that addresses these challenges. It builds on top of a flat concept representation that summarizes entities' heterogeneous digital footprint, jointly considers the entire spectrum of predicative tasks as a single learning task, and leverages principled meta learning approach with efficient first order meta-optimization procedure under a provable performance guarantee in order to solve the learning task. Experiments on both proprietary production datasets and public structured learning tasks demonstrate that MetaCon can lead to substantial improvements over state of the art recommendation and ranking approaches.

As more and more AI agents are used in practice, it is time to think about how to make these agents fully autonomous so that they can learn by themselves in a self-motivated and self-supervised manner rather than being retrained periodically on the initiation of human engineers using expanded training data. As the real-world is an open environment with unknowns or novelties, detecting novelties or unknowns, characterizing them, accommodating or adapting to them, gathering ground-truth training data, and incrementally learning the unknowns/novelties are critical to making the agent more and more knowledgeable and powerful over time. The key challenge is how to automate the process so that it is carried out on the agent's own initiative and through its own interactions with humans and the environment. Since an AI agent usually has a performance task, characterizing each novelty becomes critical and necessary so that the agent can formulate an appropriate response to adapt its behavior to accommodate the novelty and to learn from it to improve the agent's adaptation capability and task performance. The process goes continually without termination. This paper proposes a theoretic framework for this learning paradigm to promote the research of building Self-initiated Open world Learning (SOL) agents. An example SOL agent is also described., Comment: Published in AAAI 2022 Spring Symposium Series

Recent Offline Reinforcement Learning methods have succeeded in learning high-performance policies from fixed datasets of experience. A particularly effective approach learns to first identify and then mimic optimal decision-making strategies. Our work evaluates this method's ability to scale to vast datasets consisting almost entirely of sub-optimal noise. A thorough investigation on a custom benchmark helps identify several key challenges involved in learning from high-noise datasets. We re-purpose prioritized experience sampling to locate expert-level demonstrations among millions of low-performance samples. This modification enables offline agents to learn state-of-the-art policies in benchmark tasks using datasets where expert actions are outnumbered nearly 65:1.

Optimization-based meta-learning typically assumes tasks are sampled from a single distribution - an assumption oversimplifies and limits the diversity of tasks that meta-learning can model. Handling tasks from multiple different distributions is challenging for meta-learning due to a so-called task ambiguity issue. This paper proposes a novel method, ST-MAML, that empowers model-agnostic meta-learning (MAML) to learn from multiple task distributions. ST-MAML encodes tasks using a stochastic neural network module, that summarizes every task with a stochastic representation. The proposed Stochastic Task (ST) strategy allows a meta-model to get tailored for the current task and enables us to learn a distribution of solutions for an ambiguous task. ST-MAML also propagates the task representation to revise the encoding of input variables. Empirically, we demonstrate that ST-MAML matches or outperforms the state-of-the-art on two few-shot image classification tasks, one curve regression benchmark, one image completion problem, and a real-world temperature prediction application. To the best of authors' knowledge, this is the first time optimization-based meta-learning method being applied on a large-scale real-world task.

Multivariate time series forecasting focuses on predicting future values based on historical context. State-of-the-art sequence-to-sequence models rely on neural attention between timesteps, which allows for temporal learning but fails to consider distinct spatial relationships between variables. In contrast, methods based on graph neural networks explicitly model variable relationships. However, these methods often rely on predefined graphs that cannot change over time and perform separate spatial and temporal updates without establishing direct connections between each variable at every timestep. Our work addresses these problems by translating multivariate forecasting into a "spatiotemporal sequence" formulation where each Transformer input token represents the value of a single variable at a given time. Long-Range Transformers can then learn interactions between space, time, and value information jointly along this extended sequence. Our method, which we call Spacetimeformer, achieves competitive results on benchmarks from traffic forecasting to electricity demand and weather prediction while learning spatiotemporal relationships purely from data.