We present an algorithm for skill discovery from expert demonstrations. The algorithm first utilizes Large Language Models (LLMs) to propose an initial segmentation of the trajectories. Following that, a hierarchical variational inference framework incorporates the LLM-generated segmentation information to discover reusable skills by merging trajectory segments. To further control the trade-off between compression and reusability, we introduce a novel auxiliary objective based on the Minimum Description Length principle that helps guide this skill discovery process. Our results demonstrate that agents equipped with our method are able to discover skills that help accelerate learning and outperform baseline skill learning approaches on new long-horizon tasks in BabyAI, a grid world navigation environment, as well as ALFRED, a household simulation environment.
We view large language models (LLMs) as stochastic \emph{language layers} in a network, where the learnable parameters are the natural language \emph{prompts} at each layer. We stack two such layers, feeding the output of one layer to the next. We call the stacked architecture a \emph{Deep Language Network} (DLN). We first show how to effectively perform prompt optimization for a 1-Layer language network (DLN-1). We then show how to train 2-layer DLNs (DLN-2), where two prompts must be learnt. We consider the output of the first layer as a latent variable to marginalize, and devise a variational inference algorithm for joint prompt training. A DLN-2 reaches higher performance than a single layer, sometimes comparable to few-shot GPT-4 even when each LLM in the network is smaller and less powerful. The DLN code is open source: https://github.com/microsoft/deep-language-networks .
The growing utilization of machine learning (ML) in decision-making processes raises questions about its benefits to society. In this study, we identify and analyze three axes of heterogeneity that significantly influence the trajectory of ML products. These axes are i) values, culture and regulations, ii) data composition, and iii) resource and infrastructure capacity. We demonstrate how these axes are interdependent and mutually influence one another, emphasizing the need to consider and address them jointly. Unfortunately, the current research landscape falls short in this regard, often failing to adopt a holistic approach. We examine the prevalent practices and methodologies that skew these axes in favor of a selected few, resulting in power concentration, homogenized control, and increased dependency. We discuss how this fragmented study of the three axes poses a significant challenge, leading to an impractical solution space that lacks reflection of real-world scenarios. Addressing these issues is crucial to ensure a more comprehensive understanding of the interconnected nature of society and to foster the democratic and inclusive development of ML systems that are more aligned with real-world complexities and its diverse requirements.
Actor-critic (AC) methods are widely used in reinforcement learning (RL) and benefit from the flexibility of using any policy gradient method as the actor and value-based method as the critic. The critic is usually trained by minimizing the TD error, an objective that is potentially decorrelated with the true goal of achieving a high reward with the actor. We address this mismatch by designing a joint objective for training the actor and critic in a decision-aware fashion. We use the proposed objective to design a generic, AC algorithm that can easily handle any function approximation. We explicitly characterize the conditions under which the resulting algorithm guarantees monotonic policy improvement, regardless of the choice of the policy and critic parameterization. Instantiating the generic algorithm results in an actor that involves maximizing a sequence of surrogate functions (similar to TRPO, PPO) and a critic that involves minimizing a closely connected objective. Using simple bandit examples, we provably establish the benefit of the proposed critic objective over the standard squared error. Finally, we empirically demonstrate the benefit of our decision-aware actor-critic framework on simple RL problems.
We consider minimizing functions for which it is expensive to compute the (possibly stochastic) gradient. Such functions are prevalent in reinforcement learning, imitation learning and adversarial training. Our target optimization framework uses the (expensive) gradient computation to construct surrogate functions in a target space (e.g. the logits output by a linear model for classification) that can be minimized efficiently. This allows for multiple parameter updates to the model, amortizing the cost of gradient computation. In the full-batch setting, we prove that our surrogate is a global upper-bound on the loss, and can be (locally) minimized using a black-box optimization algorithm. We prove that the resulting majorization-minimization algorithm ensures convergence to a stationary point of the loss. Next, we instantiate our framework in the stochastic setting and propose the $SSO$ algorithm, which can be viewed as projected stochastic gradient descent in the target space. This connection enables us to prove theoretical guarantees for $SSO$ when minimizing convex functions. Our framework allows the use of standard stochastic optimization algorithms to construct surrogates which can be minimized by any deterministic optimization method. To evaluate our framework, we consider a suite of supervised learning and imitation learning problems. Our experiments indicate the benefits of target optimization and the effectiveness of $SSO$.
Parameter-efficient fine-tuning (PEFT) methods can adapt large language models to downstream tasks by training a small amount of newly added parameters. In multi-task settings, PEFT adapters typically train on each task independently, inhibiting transfer across tasks, or on the concatenation of all tasks, which can lead to negative interference. To address this, Polytropon (Ponti et al.) jointly learns an inventory of PEFT adapters and a routing function to share variable-size sets of adapters across tasks. Subsequently, adapters can be re-combined and fine-tuned on novel tasks even with limited data. In this paper, we investigate to what extent the ability to control which adapters are active for each task leads to sample-efficient generalization. Thus, we propose less expressive variants where we perform weighted averaging of the adapters before few-shot adaptation (Poly-mu) instead of learning a routing function. Moreover, we introduce more expressive variants where finer-grained task-adapter allocation is learned through a multi-head routing function (Poly-S). We test these variants on three separate benchmarks for multi-task learning. We find that Poly-S achieves gains on all three (up to 5.3 points on average) over strong baselines, while incurring a negligible additional cost in parameter count. In particular, we find that instruction tuning, where models are fully fine-tuned on natural language instructions for each task, is inferior to modular methods such as Polytropon and our proposed variants.
We use functional mirror ascent to propose a general framework (referred to as FMA-PG) for designing policy gradient methods. The functional perspective distinguishes between a policy's functional representation (what are its sufficient statistics) and its parameterization (how are these statistics represented) and naturally results in computationally efficient off-policy updates. For simple policy parameterizations, the FMA-PG framework ensures that the optimal policy is a fixed point of the updates. It also allows us to handle complex policy parameterizations (e.g., neural networks) while guaranteeing policy improvement. Our framework unifies several PG methods and opens the way for designing sample-efficient variants of existing methods. Moreover, it recovers important implementation heuristics (e.g., using forward vs reverse KL divergence) in a principled way. With a softmax functional representation, FMA-PG results in a variant of TRPO with additional desirable properties. It also suggests an improved variant of PPO, whose robustness and efficiency we empirically demonstrate on MuJoCo. Via experiments on simple reinforcement learning problems, we evaluate algorithms instantiated by FMA-PG.
We investigate the impact of aliasing on generalization in Deep Convolutional Networks and show that data augmentation schemes alone are unable to prevent it due to structural limitations in widely used architectures. Drawing insights from frequency analysis theory, we take a closer look at ResNet and EfficientNet architectures and review the trade-off between aliasing and information loss in each of their major components. We show how to mitigate aliasing by inserting non-trainable low-pass filters at key locations, particularly where networks lack the capacity to learn them. These simple architectural changes lead to substantial improvements in generalization on i.i.d. and even more on out-of-distribution conditions, such as image classification under natural corruptions on ImageNet-C [11] and few-shot learning on Meta-Dataset [26]. State-of-the art results are achieved on both datasets without introducing additional trainable parameters and using the default hyper-parameters of open source codebases.
We study the stochastic bilinear minimax optimization problem, presenting an analysis of the Stochastic ExtraGradient (SEG) method with constant step size, and presenting variations of the method that yield favorable convergence. We first note that the last iterate of the basic SEG method only contracts to a fixed neighborhood of the Nash equilibrium, independent of the step size. This contrasts sharply with the standard setting of minimization where standard stochastic algorithms converge to a neighborhood that vanishes in proportion to the square-root (constant) step size. Under the same setting, however, we prove that when augmented with iteration averaging, SEG provably converges to the Nash equilibrium, and such a rate is provably accelerated by incorporating a scheduled restarting procedure. In the interpolation setting, we achieve an optimal convergence rate up to tight constants. We present numerical experiments that validate our theoretical findings and demonstrate the effectiveness of the SEG method when equipped with iteration averaging and restarting.
Adversarial robustness is an open challenge in deep learning, most often tackled using adversarial training. Adversarial training is computationally costly, involving alternated optimization with a trade-off between standard generalization and adversarial robustness. We explore training robust models without adversarial training by revisiting a known result linking maximally robust classifiers and minimum norm solutions, and combining it with recent results on the implicit bias of optimizers. First, we show that, under certain conditions, it is possible to achieve both perfect standard accuracy and a certain degree of robustness without a trade-off, simply by training an overparameterized model using the implicit bias of the optimization. In that regime, there is a direct relationship between the type of the optimizer and the attack to which the model is robust. Second, we investigate the role of the architecture in designing robust models. In particular, we characterize the robustness of linear convolutional models, showing that they resist attacks subject to a constraint on the Fourier-$\ell_\infty$ norm. This result explains the property of $\ell_p$-bounded adversarial perturbations that tend to be concentrated in the Fourier domain. This leads us to a novel attack in the Fourier domain that is inspired by the well-known frequency-dependent sensitivity of human perception. We evaluate Fourier-$\ell_\infty$ robustness of recent CIFAR-10 models with robust training and visualize adversarial perturbations.