Undeniably, Large Language Models (LLMs) have stirred an extraordinary wave of innovation in the machine learning research domain, resulting in substantial impact across diverse fields such as reinforcement learning, robotics, and computer vision. Their incorporation has been rapid and transformative, marking a significant paradigm shift in the field of machine learning research. However, the field of experimental design, grounded on black-box optimization, has been much less affected by such a paradigm shift, even though integrating LLMs with optimization presents a unique landscape ripe for exploration. In this position paper, we frame the field of black-box optimization around sequence-based foundation models and organize their relationship with previous literature. We discuss the most promising ways foundational language models can revolutionize optimization, which include harnessing the vast wealth of information encapsulated in free-form text to enrich task comprehension, utilizing highly flexible sequence models such as Transformers to engineer superior optimization strategies, and enhancing performance prediction over previously unseen search spaces.

Genetic algorithms constitute a family of black-box optimization algorithms, which take inspiration from the principles of biological evolution. While they provide a general-purpose tool for optimization, their particular instantiations can be heuristic and motivated by loose biological intuition. In this work we explore a fundamentally different approach: Given a sufficiently flexible parametrization of the genetic operators, we discover entirely new genetic algorithms in a data-driven fashion. More specifically, we parametrize selection and mutation rate adaptation as cross- and self-attention modules and use Meta-Black-Box-Optimization to evolve their parameters on a set of diverse optimization tasks. The resulting Learned Genetic Algorithm outperforms state-of-the-art adaptive baseline genetic algorithms and generalizes far beyond its meta-training settings. The learned algorithm can be applied to previously unseen optimization problems, search dimensions & evaluation budgets. We conduct extensive analysis of the discovered operators and provide ablation experiments, which highlight the benefits of flexible module parametrization and the ability to transfer ('plug-in') the learned operators to conventional genetic algorithms.

A shared goal of several machine learning communities like continual learning, meta-learning and transfer learning, is to design algorithms and models that efficiently and robustly adapt to unseen tasks. An even more ambitious goal is to build models that never stop adapting, and that become increasingly more efficient through time by suitably transferring the accrued knowledge. Beyond the study of the actual learning algorithm and model architecture, there are several hurdles towards our quest to build such models, such as the choice of learning protocol, metric of success and data needed to validate research hypotheses. In this work, we introduce the Never-Ending VIsual-classification Stream (NEVIS'22), a benchmark consisting of a stream of over 100 visual classification tasks, sorted chronologically and extracted from papers sampled uniformly from computer vision proceedings spanning the last three decades. The resulting stream reflects what the research community thought was meaningful at any point in time, and it serves as an ideal test bed to assess how well models can adapt to new tasks, and do so better and more efficiently as time goes by. Despite being limited to classification, the resulting stream has a rich diversity of tasks from OCR, to texture analysis, scene recognition, and so forth. The diversity is also reflected in the wide range of dataset sizes, spanning over four orders of magnitude. Overall, NEVIS'22 poses an unprecedented challenge for current sequential learning approaches due to the scale and diversity of tasks, yet with a low entry barrier as it is limited to a single modality and well understood supervised learning problems. Moreover, we provide a reference implementation including strong baselines and an evaluation protocol to compare methods in terms of their trade-off between accuracy and compute. We hope that NEVIS'22 can be useful to researchers working on continual learning, meta-learning, AutoML and more generally sequential learning, and help these communities join forces towards more robust models that efficiently adapt to a never ending stream of data.

Meta reinforcement learning (RL) attempts to discover new RL algorithms automatically from environment interaction. In so-called black-box approaches, the policy and the learning algorithm are jointly represented by a single neural network. These methods are very flexible, but they tend to underperform compared to human-engineered RL algorithms in terms of generalisation to new, unseen environments. In this paper, we explore the role of symmetries in meta-generalisation. We show that a recent successful meta RL approach that meta-learns an objective for backpropagation-based learning exhibits certain symmetries (specifically the reuse of the learning rule, and invariance to input and output permutations) that are not present in typical black-box meta RL systems. We hypothesise that these symmetries can play an important role in meta-generalisation. Building off recent work in black-box supervised meta learning, we develop a black-box meta RL system that exhibits these same symmetries. We show through careful experimentation that incorporating these symmetries can lead to algorithms with a greater ability to generalise to unseen action & observation spaces, tasks, and environments.

Meta-learning hyperparameter optimization (HPO) algorithms from prior experiments is a promising approach to improve optimization efficiency over objective functions from a similar distribution. However, existing methods are restricted to learning from experiments sharing the same set of hyperparameters. In this paper, we introduce the OptFormer, the first text-based Transformer HPO framework that provides a universal end-to-end interface for jointly learning policy and function prediction when trained on vast tuning data from the wild, such as Google’s Vizier database, one of the world’s largest HPO datasets. Our extensive experiments demonstrate that the OptFormer can simultaneously imitate at least 7 different HPO algorithms, which can be further improved via its function uncertainty estimates. Compared to a Gaussian Process, the OptFormer also learns a robust prior distribution for hyperparameter response functions, and can thereby provide more accurate and better calibrated predictions. This work paves the path to future extensions for training a Transformer-based model as a general HPO optimizer.

We show that the popular reinforcement learning (RL) strategy of estimating the state-action value (Q-function) by minimizing the mean squared Bellman error leads to a regression problem with confounding, the inputs and output noise being correlated. Hence, direct minimization of the Bellman error can result in significantly biased Q-function estimates. We explain why fixing the target Q-network in Deep Q-Networks and Fitted Q Evaluation provides a way of overcoming this confounding, thus shedding new light on this popular but not well understood trick in the deep RL literature. An alternative approach to address confounding is to leverage techniques developed in the causality literature, notably instrumental variables (IV). We bring together here the literature on IV and RL by investigating whether IV approaches can lead to improved Q-function estimates. This paper analyzes and compares a wide range of recent IV methods in the context of offline policy evaluation (OPE), where the goal is to estimate the value of a policy using logged data only. By applying different IV techniques to OPE, we are not only able to recover previously proposed OPE methods such as model-based techniques but also to obtain competitive new techniques. We find empirically that state-of-the-art OPE methods are closely matched in performance by some IV methods such as AGMM, which were not developed for OPE. We open-source all our code and datasets at https://github.com/liyuan9988/IVOPEwithACME.

This paper addresses the problem of policy selection in domains with abundant logged data, but with a restricted interaction budget. Solving this problem would enable safe evaluation and deployment of offline reinforcement learning policies in industry, robotics, and recommendation domains among others. Several off-policy evaluation (OPE) techniques have been proposed to assess the value of policies using only logged data. However, there is still a big gap between the evaluation by OPE and the full online evaluation in the real environment. Yet, large amounts of online interactions are often not possible in practice. To overcome this problem, we introduce active offline policy selection --- a novel sequential decision approach that combines logged data with online interaction to identify the best policy. This approach uses OPE estimates to warm start the online evaluation. Then, in order to utilize the limited environment interactions wisely we decide which policy to evaluate next based on a Bayesian optimization method with a kernel function that represents policy similarity. We use multiple benchmarks with a large number of candidate policies to show that the proposed approach improves upon state-of-the-art OPE estimates and pure online policy evaluation.

Many real-world problems, including multi-speaker text-to-speech synthesis, can greatly benefit from the ability to meta-learn large models with only a few task- specific components. Updating only these task-specific modules then allows the model to be adapted to low-data tasks for as many steps as necessary without risking overfitting. Unfortunately, existing meta-learning methods either do not scale to long adaptation or else rely on handcrafted task-specific architectures. Here, we propose a meta-learning approach that obviates the need for this often sub-optimal hand-selection. In particular, we develop general techniques based on Bayesian shrinkage to automatically discover and learn both task-specific and general reusable modules. Empirically, we demonstrate that our method discovers a small set of meaningful task-specific modules and outperforms existing meta- learning approaches in domains like few-shot text-to-speech that have little task data and long adaptation horizons. We also show that existing meta-learning methods including MAML, iMAML, and Reptile emerge as special cases of our method.

We learn recurrent neural network optimizers trained on simple synthetic functions by gradient descent. We show that these learned optimizers exhibit a remarkable degree of transfer in that they can be used to efficiently optimize a broad range of derivative-free black-box functions, including Gaussian process bandits, simple control objectives, global optimization benchmarks and hyper-parameter tuning tasks. Up to the training horizon, the learned optimizers learn to trade-off exploration and exploitation, and compare favourably with heavily engineered Bayesian optimization packages for hyper-parameter tuning.

PixelCNN achieves state-of-the-art results in density estimation for natural images. Although training is fast, inference is costly, requiring one network evaluation per pixel; O(N) for N pixels. This can be sped up by caching activations, but still involves generating each pixel sequentially. In this work, we propose a parallelized PixelCNN that allows more efficient inference by modeling certain pixel groups as conditionally independent. Our new PixelCNN model achieves competitive density estimation and orders of magnitude speedup – O(log N) sampling instead of O(N) – enabling the practical generation of 512x512 images. We evaluate the model on class-conditional image generation, text-to-image synthesis, and action-conditional video generation, showing that our model achieves the best results among non-pixel-autoregressive density models that allow efficient sampling.

Drawing a sample from a discrete distribution is one of the building components for Monte Carlo methods. Like other sampling algorithms, discrete sampling suffers from the high computational burden in large-scale inference problems. We study the problem of sampling a discrete random variable with a high degree of dependency that is typical in large-scale Bayesian inference and graphical models, and propose an efficient approximate solution with a subsampling approach. We make a novel connection between the discrete sampling and Multi-Armed Bandits problems with a finite reward population and provide three algorithms with theoretical guarantees. Empirical evaluations show the robustness and efficiency of the approximate algorithms in both synthetic and real-world large-scale problems.

The Gibbs sampler is one of the most popular algorithms for inference in statistical models. In this paper, we introduce a herding variant of this algorithm, called herded Gibbs, that is entirely deterministic. We prove that herded Gibbs has an $O(1/T)$ convergence rate for models with independent variables and for fully connected probabilistic graphical models. Herded Gibbs is shown to outperform Gibbs in the tasks of image denoising with MRFs and named entity recognition with CRFs. However, the convergence for herded Gibbs for sparsely connected probabilistic graphical models is still an open problem.

We argue that when faced with big data sets, learning and inference algorithms should compute updates using only subsets of data items. We introduce algorithms that use sequential hypothesis tests to adaptively select such a subset of data points. The statistical properties of this subsampling process can be used to control the efficiency and accuracy of learning or inference. In the context of learning by optimization, we test for the probability that the update direction is no more than 90 degrees in the wrong direction. In the context of posterior inference using Markov chain Monte Carlo, we test for the probability that our decision to accept or reject a sample is wrong. We experimentally evaluate our algorithms on a number of models and data sets.

Multivariate categorical data occur in many applications of machine learning. One of the main difficulties with these vectors of categorical variables is sparsity. The number of possible observations grows exponentially with vector length, but dataset diversity might be poor in comparison. Recent models have gained significant improvement in supervised tasks with this data. These models embed observations in a continuous space to capture similarities between them. Building on these ideas we propose a Bayesian model for the unsupervised task of distribution estimation of multivariate categorical data. We model vectors of categorical variables as generated from a non-linear transformation of a continuous latent space. Non-linearity captures multi-modality in the distribution. The continuous representation addresses sparsity. Our model ties together many existing models, linking the linear categorical latent Gaussian model, the Gaussian process latent variable model, and Gaussian process classification. We derive inference for our model based on recent developments in sampling based variational inference. We show empirically that the model outperforms its linear and discrete counterparts in imputation tasks of sparse data.

Bayesian nonparametric mixture models based on the Dirichlet process (DP) have been widely used for solving problems like clustering, density estimation and topic modelling. These models make weak assumptions about the underlying process that generated the observed data. Thus, when more data are collected, the complexity of these models can change accordingly. These theoretical properties often lead to superior predictive performance when compared to traditional finite mixture models. However, despite the increasing amount of data available, the application of Bayesian nonparametric mixture models is so far limited to relatively small data sets. In this paper, we propose an efficient distributed inference algorithm for the DP and the HDP mixture model. The proposed method is based on a variant of the slice sampler for DPs. Since this sampler does not involve a pre-determined truncation, the stationary distribution of the sampling algorithm is unbiased. We provide both local thread-level and distributed machine-level parallel implementations and study the performance of this sampler through an extensive set of experiments on image and text data. When compared to existing inference algorithms, the proposed method exhibits state-of-the-art accuracy and strong scalability with up to 512 cores.

Can we make Bayesian posterior MCMC sampling more efficient when faced with very large datasets? We argue that computing the likelihood for N datapoints in the Metropolis-Hastings (MH) test to reach a single binary decision is computationally inefficient. We introduce an approximate MH rule based on a sequential hypothesis test that allows us to accept or reject samples with high confidence using only a fraction of the data required for the exact MH rule. While this method introduces an asymptotic bias, we show that this bias can be controlled and is more than offset by a decrease in variance due to our ability to draw more samples per unit of time.

Darting Monte Carlo (DMC) is a MCMC procedure designed to effectively mix between multiple modes of a probability distribution. We propose an adaptive and distributed version of this method by using regenerations. This allows us to run multiple chains in parallel and adapt the shape of the jump regions as well as all other aspects of the Markov chain on the fly. We show that this significantly improves the performance of DMC because 1) a population of chains has a higher chance of finding the modes in the distribution, 2) jumping between modes becomes easier due to the adaptation of their shapes, 3) computation is much more efficient due to parallelization across multiple processors. While the curse of dimensionality is a challenge for both DMC and regeneration, we find that their combination ameliorates this issue slightly.

Bayesian estimation in Markov random fields is very hard due to the intractability of the partition function. The introduction of hidden units makes the situation even worse due to the presence of potentially very many modes in the posterior distribution. For the first time we propose a comprehensive procedure to address one of the Bayesian estimation problems, approximating the evidence of partially observed MRFs based on the Laplace approximation. We also introduce a number of approximate MCMC-based methods for comparison but find that the Laplace approximation significantly outperforms these.

We present a new method to combine possibly inconsistent locally (piecewise) trained conditional models p(yα|xα) into pseudo-samples from a global model. Our method does not require training of a CRF, but instead generates samples by iterating forward a weakly chaotic dynamical system. The new method is illustrated on image segmentation tasks where classifiers based on local appearance cues are combined with pairwise boundary cues.

The paper develops a connection between traditional perceptron algorithms and recently introduced herding algorithms. It is shown that both algorithms can be viewed as an application of the perceptron cycling theorem. This connection strengthens some herding results and suggests new (supervised) herding algorithms that, like CRFs or discriminative RBMs, make predictions by conditioning on the input attributes. We develop and investigate variants of conditional herding, and show that conditional herding leads to practical algorithms that perform better than or on par with related classifiers such as the voted perceptron and the discriminative RBM.