In this paper we present the first safe system for full control of self-driving vehicles trained from human demonstrations and deployed in challenging, real-world, urban environments.Current industry-standard solutions use rulebased systems planning.Although they perform reasonably well common scenarios, engineering complexity renders approach incompatible with human-level performance.On other hand, performance machine-learned (ML) planning can be improved by simply adding more exemplar...

Continual reinforcement learning (CRL) is the study of optimal strategies for maximizing rewards in sequential environments that change over time. This particularly crucial domains such as robotics, where operational environment inherently dynamic and subject to continual change. Nevertheless, research this area has thus far concentrated on off-policy algorithms with replay buffers are capable amortizing impact distribution shifts. Such an approach not feasible on-policy learn solely from...

The problem of reducing processing time large deep learning models is a fundamental challenge in many real-world applications. Early exit methods strive towards this goal by attaching additional Internal Classifiers (ICs) to intermediate layers neural network. ICs can quickly return predictions for easy examples and, as result, reduce the average inference whole model. However, if particular IC does not decide an answer early, its are discarded, with computations effectively being wasted. To...

In this work we are the first to present an offline policy gradient method for learning imitative policies complex urban driving from a large corpus of real-world demonstrations. This is achieved by building differentiable data-driven simulator on top perception outputs and high-fidelity HD maps area. It allows us synthesize new experiences existing demonstrations using mid-level representations. Using then train network in closed-loop employing gradients. We our proposed 100 hours expert...

Continual learning (CL) -- the ability to continuously learn, building on previously acquired knowledge is a natural requirement for long-lived autonomous reinforcement (RL) agents. While such agents, one needs balance opposing desiderata, as constraints capacity and compute, not catastrophically forget, exhibit positive transfer new tasks. Understanding right trade-off conceptually computationally challenging, which we argue has led community overly focus catastrophic forgetting. In...

We propose a semi-supervised generative model, SeGMA, which learns joint probability distribution of data and their classes is implemented in typical Wasserstein autoencoder framework. choose mixture Gaussians as target latent space, provides natural splitting into clusters. To connect Gaussian components with correct classes, we use small amount labeled classifier induced by the distribution. SeGMA optimized efficiently due to Cramer-Wold distance maximum mean discrepancy penalty, yields...

Fine-tuning is a widespread technique that allows practitioners to transfer pre-trained capabilities, as recently showcased by the successful applications of foundation models. However, fine-tuning reinforcement learning (RL) models remains challenge. This work conceptualizes one specific cause poor transfer, accentuated in RL setting interplay between actions and observations: forgetting capabilities. Namely, model deteriorates on state subspace downstream task not visited initial phase...

Although modern generative models achieve excellent quality in a variety of tasks, they often lack the essential ability to generate examples with requested properties, such as age person photo or weight generated molecule. To overcome these limitations we propose PluGeN (Plugin Generative Network), simple yet effective technique that can be used plugin for pre-trained models. The idea behind our approach is transform entangled latent representation using flow-based module into...

Large Language Models (LLMs) and Vision (VLMs) possess extensive knowledge exhibit promising reasoning abilities; however, they still struggle to perform well in complex, dynamic environments. Real-world tasks require handling intricate interactions, advanced spatial reasoning, long-term planning, continuous exploration of new strategies-areas which we lack effective methodologies for comprehensively evaluating these capabilities. To address this gap, introduce BALROG, a novel benchmark...

The ability of continual learning systems to transfer knowledge from previously seen tasks in order maximize performance on new is a significant challenge for the field, limiting applicability solutions realistic scenarios. Consequently, this study aims broaden our understanding and its driving forces specific case reinforcement learning. We adopt SAC as underlying RL algorithm Continual World suite continuous control tasks. systematically how different components (the actor critic,...

The problem of reducing processing time large deep learning models is a fundamental challenge in many real-world applications. Early exit methods strive towards this goal by attaching additional Internal Classifiers (ICs) to intermediate layers neural network. ICs can quickly return predictions for easy examples and, as result, reduce the average inference whole model. However, if particular IC does not decide an answer early, its are discarded, with computations effectively being wasted. To...

Modern generative models achieve excellent quality in a variety of tasks including image or text generation and chemical molecule modeling. However, existing methods often lack the essential ability to generate examples with requested properties, such as age person photo weight generated molecule. Incorporating additional conditioning factors would require rebuilding entire architecture optimizing parameters from scratch. Moreover, it is difficult disentangle selected attributes so that...

Active Visual Exploration (AVE) is a task that involves dynamically selecting observations (glimpses), which critical to facilitate comprehension and navigation within an environment. While modern AVE methods have demonstrated impressive performance, they are constrained fixed-scale glimpses from rigid grids. In contrast, existing mobile platforms equipped with optical zoom capabilities can capture of arbitrary positions scales. To address this gap between software hardware capabilities, we...

A new breed of gated-linear recurrent neural networks has reached state-of-the-art performance on a range sequence modeling problems. Such models naturally handle long sequences efficiently, as the cost processing input is independent length. Here, we explore another advantage these stateful models, inspired by success model merging through parameter interpolation. Building parallels between fine-tuning and in-context learning, investigate whether can treat internal states task vectors that...

Visual perspective-taking (VPT), the ability to understand viewpoint of another person, enables individuals anticipate actions other people. For instance, a driver can avoid accidents by assessing what pedestrians see. Humans typically develop this skill in early childhood, but it remains unclear whether recently emerging Vision Language Models (VLMs) possess such capability. Furthermore, as these models are increasingly deployed real world, understanding how they perform nuanced tasks like...

We introduce bio-inspired artificial neural networks consisting of neurons that are additionally characterized by spatial positions. To simulate properties biological systems we add the costs penalizing long connections and proximity in a two-dimensional space. Our experiments show case where network performs two different tasks, naturally split into clusters, each cluster is responsible for processing task. This behavior not only corresponds to systems, but also allows further insight...

Experience replay is a simple and well-performing strategy for continual learning problems, often used as basis more advanced methods. However, the dynamics of experience are not yet well understood. To showcase this, we focus on single component this problem, namely choosing batch size buffer samples. We find that small batches perform much better at stopping forgetting than larger batches, contrary to intuitive assumption it recall samples from past avoid forgetting. show phenomenon does...

Schedae Informaticae » 2018 Volume 27 Deep learning-based initialization for object packing A

We introduce a new training paradigm that enforces interval constraints on neural network parameter space to control forgetting. Contemporary Continual Learning (CL) methods focus networks efficiently from stream of data, while reducing the negative impact catastrophic forgetting, yet they do not provide any firm guarantees performance will deteriorate uncontrollably over time. In this work, we show how put bounds forgetting by reformulating continual learning model as contraction its space....