Neural networks are known to exploit spurious artifacts (or shortcuts) that co-occur with a target label, exhibiting heuristic memorization. On the other hand, networks have been shown to memorize training examples, resulting in example-level memorization. These kinds of memorization impede generalization of networks beyond their training distributions. Detecting such memorization could be challenging, often requiring researchers to curate tailored test sets. In this work, we hypothesize -- and subsequently show -- that the diversity in the activation patterns of different neurons is reflective of model generalization and memorization. We quantify the diversity in the neural activations through information-theoretic measures and find support for our hypothesis on experiments spanning several natural language and vision tasks. Importantly, we discover that information organization points to the two forms of memorization, even for neural activations computed on unlabelled in-distribution examples. Lastly, we demonstrate the utility of our findings for the problem of model selection. The associated code and other resources for this work are available at https://rachitbansal.github.io/information-measures.

Reinforcement learning (RL) provides a powerful framework for decision-making, but its application in practice often requires a carefully designed reward function. Adversarial Imitation Learning (AIL) sheds light on automatic policy acquisition without access to the reward signal from the environment. In this work, we propose Auto-Encoding Adversarial Imitation Learning (AEAIL), a robust and scalable AIL framework. To induce expert policies from demonstrations, AEAIL utilizes the reconstruction error of an auto-encoder as a reward signal, which provides more information for optimizing policies than the prior discriminator-based ones. Subsequently, we use the derived objective functions to train the auto-encoder and the agent policy. Experiments show that our AEAIL performs superior compared to state-of-the-art methods on both state and image based environments. More importantly, AEAIL shows much better robustness when the expert demonstrations are noisy.

A year after being cleared for the test, billionaire Elon Musk's Neuralink has implanted its wireless brain chip in a human for the first time. Musk announced that the patient received the implant on Sunday and "is recovering well". The device is meant to have several applications, from restoring motor functionality within people to enabling a brain-computer interface. There was no independent verification of Musk's claims, and Neuralink did not offer too many details. Musk has touted Neuralink as the future of technology and medicine, but ethical concerns have been raised around the chip and its testing.

The social media platform, formerly Twitter, was so slow to react that one image racked up 47m views before it was taken down. It was largely Swift's fans who mobilised and mass-reported the images, and there was a sense of public anger, with even the White House calling it "alarming". X eventually removed the images and blocked searches to the pop star's name on Sunday evening. For women who have been victims of the creation and sharing of nonconsensual deepfake pornography, the events of the past week will have been a horrible reminder of their own abuse, even if they may also hope that the spotlight will force legislators into action. But because the pictures were removed, Swift's experience is far from the norm.

Language models have become a critical technology to tackling a wide range of natural language processing tasks, yet many details about how the best-performing language models were developed are not reported. In particular, information about their pretraining corpora is seldom discussed: commercial language models rarely provide any information about their data; even open models rarely release datasets they are trained on, or an exact recipe to reproduce them. As a result, it is challenging to conduct certain threads of language modeling research, such as understanding how training data impacts model capabilities and shapes their limitations. To facilitate open research on language model pretraining, we release Dolma, a three trillion tokens English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. In addition, we open source our data curation toolkit to enable further experimentation and reproduction of our work. In this report, we document Dolma, including its design principles, details about its construction, and a summary of its contents. We interleave this report with analyses and experimental results from training language models on intermediate states of Dolma to share what we have learned about important data curation practices, including the role of content or quality filters, deduplication, and multi-source mixing. Dolma has been used to train OLMo, a state-of-the-art, open language model and framework designed to build and study the science of language modeling.

In recent years, significant progress has been made in the field of robotic reinforcement learning (RL), enabling methods that handle complex image observations, train in the real world, and incorporate auxiliary data, such as demonstrations and prior experience. However, despite these advances, robotic RL remains hard to use. It is acknowledged among practitioners that the particular implementation details of these algorithms are often just as important (if not more so) for performance as the choice of algorithm. We posit that a significant challenge to widespread adoption of robotic RL, as well as further development of robotic RL methods, is the comparative inaccessibility of such methods. To address this challenge, we developed a carefully implemented library containing a sample efficient off-policy deep RL method, together with methods for computing rewards and resetting the environment, a high-quality controller for a widely-adopted robot, and a number of challenging example tasks. We provide this library as a resource for the community, describe its design choices, and present experimental results. Perhaps surprisingly, we find that our implementation can achieve very efficient learning, acquiring policies for PCB board assembly, cable routing, and object relocation between 25 to 50 minutes of training per policy on average, improving over state-of-the-art results reported for similar tasks in the literature. These policies achieve perfect or near-perfect success rates, extreme robustness even under perturbations, and exhibit emergent recovery and correction behaviors. We hope that these promising results and our high-quality open-source implementation will provide a tool for the robotics community to facilitate further developments in robotic RL. Our code, documentation, and videos can be found at https://serl-robot.github.io/

Metabolic energy consumption of a powered lower-limb exoskeleton user mainly comes from the upper body effort since the lower body is considered to be passive. However, the upper body effort of the users is largely ignored in the literature when designing motion controllers. In this work, we use deep reinforcement learning to develop a locomotion controller that minimizes ground reaction forces (GRF) on crutches. The rationale for minimizing GRF is to reduce the upper body effort of the user. Accordingly, we design a model and a learning framework for a human-exoskeleton system with crutches. We formulate a reward function to encourage the forward displacement of a human-exoskeleton system while satisfying the predetermined constraints of a physical robot. We evaluate our new framework using Proximal Policy Optimization, a state-of-the-art deep reinforcement learning (RL) method, on the MuJoCo physics simulator with different hyperparameters and network architectures over multiple trials. We empirically show that our learning model can generate joint torques based on the joint angle, velocities, and the GRF on the feet and crutch tips. The resulting exoskeleton model can directly generate joint torques from states in line with the RL framework. Finally, we empirically show that policy trained using our method can generate a gait with a 35% reduction in GRF with respect to the baseline.

We present experimental results highlighting two key differences resulting from the choice of training algorithm for two-layer neural networks. The spectral bias of neural networks is well known, while the spectral bias dependence on the choice of training algorithm is less studied. Our experiments demonstrate that an adaptive random Fourier features algorithm (ARFF) can yield a spectral bias closer to zero compared to the stochastic gradient descent optimizer (SGD). Additionally, we train two identically structured classifiers, employing SGD and ARFF, to the same accuracy levels and empirically assess their robustness against adversarial noise attacks.

Multi-label loss functions are usually non-differentiable, requiring surrogate loss functions for gradient-based optimisation. The consistency of surrogate loss functions is not proven and is exacerbated by the conflicting nature of multi-label loss functions. To directly learn from multiple related, yet potentially conflicting multi-label loss functions, we propose a Consistent Lebesgue Measure-based Multi-label Learner (CLML) and prove that CLML can achieve theoretical consistency under a Bayes risk framework. Empirical evidence supports our theory by demonstrating that: (1) CLML can consistently achieve state-of-the-art results; (2) the primary performance factor is the Lebesgue measure design, as CLML optimises a simpler feedforward model without additional label graph, perturbation-based conditioning, or semantic embeddings; and (3) an analysis of the results not only distinguishes CLML's effectiveness but also highlights inconsistencies between the surrogate and the desired loss functions.

Opinion summarisation aims to summarise the salient information and opinions presented in documents such as product reviews, discussion forums, and social media texts into short summaries that enable users to effectively understand the opinions therein. Generating biased summaries has the risk of potentially swaying public opinion. Previous studies focused on studying bias in opinion summarisation using extractive models, but limited research has paid attention to abstractive summarisation models. In this study, using political bias as a case study, we first establish a methodology to quantify bias in abstractive models, then trace it from the pre-trained models to the task of summarising social media opinions using different models and adaptation methods. We find that most models exhibit intrinsic bias. Using a social media text summarisation dataset and contrasting various adaptation methods, we find that tuning a smaller number of parameters is less biased compared to standard fine-tuning; however, the diversity of topics in training data used for fine-tuning is critical.