What happens when a new piece of knowledge is introduced into the training data and how long does it last while a large language model (LM) continues to train? We investigate this question by injecting facts into LMs from a new probing dataset, "Outlandish", which is designed to permit the testing of a spectrum of different fact types. When studying how robust these memories are, there appears to be a sweet spot in the spectrum of fact novelty between consistency with world knowledge and total randomness, where the injected memory is the most enduring. Specifically we show that facts that conflict with common knowledge are remembered for tens of thousands of training steps, while prompts not conflicting with common knowledge (mundane), as well as scrambled prompts (randomly jumbled) are both forgotten much more rapidly. Further, knowledge-conflicting facts can "prime'' how the language model hallucinates on logically unrelated prompts, showing their propensity for non-target generalization, while both mundane and randomly jumbled facts prime significantly less. Finally, we show that impacts of knowledge-conflicting facts in LMs, though they can be long lasting, can be largely erased by novel application of multi-step sparse updates, even while the training ability of the model is preserved. As such, this very simple procedure has direct implications for mitigating the effects of data poisoning in training.

We present MELODI, a novel memory architecture designed to efficiently process long documents using short context windows. The key principle behind MELODI is to represent short-term and long-term memory as a hierarchical compression scheme across both network layers and context windows. Specifically, the short-term memory is achieved through recurrent compression of context windows across multiple layers, ensuring smooth transitions between windows. In contrast, the long-term memory performs further compression within a single middle layer and aggregates information across context windows, effectively consolidating crucial information from the entire history. Compared to a strong baseline - the Memorizing Transformer employing dense attention over a large long-term memory (64K key-value pairs) - our method demonstrates superior performance on various long-context datasets while remarkably reducing the memory footprint by a factor of 8.

At present, the mechanisms of in-context learning in Transformers are not well understood and remain mostly an intuition. In this paper, we suggest that training Transformers on auto-regressive objectives is closely related to gradient-based meta-learning formulations. We start by providing a simple weight construction that shows the equivalence of data transformations induced by 1) a single linear self-attention layer and by 2) gradient-descent (GD) on a regression loss. Motivated by that construction, we show empirically that when training self-attention-only Transformers on simple regression tasks either the models learned by GD and Transformers show great similarity or, remarkably, the weights found by optimization match the construction. Thus we show how trained Transformers become mesa-optimizers i.e. learn models by gradient descent in their forward pass. This allows us, at least in the domain of regression problems, to mechanistically understand the inner workings of in-context learning in optimized Transformers. Building on this insight, we furthermore identify how Transformers surpass the performance of plain gradient descent by learning an iterative curvature correction and learn linear models on deep data representations to solve non-linear regression tasks. Finally, we discuss intriguing parallels to a mechanism identified to be crucial for in-context learning termed induction-head (Olsson et al., 2022) and show how it could be understood as a specific case of in-context learning by gradient descent learning within Transformers. Code to reproduce the experiments can be found at https://github.com/google-research/self-organising-systems/tree/master/transformers_learn_icl_by_gd .

Stochastic gradient descent plays a fundamental role in nearly all applications of deep learning. However its ability to converge to a global minimum remains shrouded in mystery. In this paper we propose to study the behavior of the loss function on fixed mini-batches along SGD trajectories. We show that the loss function on a fixed batch appears to be remarkably convex-like. In particular for ResNet the loss for any fixed mini-batch can be accurately modeled by a quadratic function and a very low loss value can be reached in just one step of gradient descent with sufficiently large learning rate. We propose a simple model that allows to analyze the relationship between the gradients of stochastic mini-batches and the full batch. Our analysis allows us to discover the equivalency between iterate aggregates and specific learning rate schedules. In particular, for Exponential Moving Average (EMA) and Stochastic Weight Averaging we show that our proposed model matches the observed training trajectories on ImageNet. Our theoretical model predicts that an even simpler averaging technique, averaging just two points a many steps apart, significantly improves accuracy compared to the baseline. We validated our findings on ImageNet and other datasets using ResNet architecture.

We focus on the problem of learning without forgetting from multiple tasks arriving sequentially, where each task is defined using a few-shot episode of novel or already seen classes. We approach this problem using the recently published HyperTransformer (HT), a Transformer-based hypernetwork that generates specialized task-specific CNN weights directly from the support set. In order to learn from a continual sequence of tasks, we propose to recursively re-use the generated weights as input to the HT for the next task. This way, the generated CNN weights themselves act as a representation of previously learned tasks, and the HT is trained to update these weights so that the new task can be learned without forgetting past tasks. This approach is different from most continual learning algorithms that typically rely on using replay buffers, weight regularization or task-dependent architectural changes. We demonstrate that our proposed Continual HyperTransformer method equipped with a prototypical loss is capable of learning and retaining knowledge about past tasks for a variety of scenarios, including learning from mini-batches, and task-incremental and class-incremental learning scenarios.

Decentralized learning with private data is a central problem in machine learning. We propose a novel distillation-based decentralized learning technique that allows multiple agents with private non-iid data to learn from each other, without having to share their data, weights or weight updates. Our approach is communication efficient, utilizes an unlabeled public dataset and uses multiple auxiliary heads for each client, greatly improving training efficiency in the case of heterogeneous data. This approach allows individual models to preserve and enhance performance on their private tasks while also dramatically improving their performance on the global aggregated data distribution. We study the effects of data and model architecture heterogeneity and the impact of the underlying communication graph topology on learning efficiency and show that our agents can significantly improve their performance compared to learning in isolation.

Deep neural networks have revolutionized many areas of machine intelligence and are now used for many vision tasks that even few years ago were considered nearly impenetrable (Krizhevsky et al., 2012; Simonyan & Zisserman, 2014; Liu et al., 2016). Advances in neural networks and hardware is resulting in much of the computation being shifted to consumer devices, delivering faster response, and better security and privacy guarantees (Konečný et al., 2016; Howard et al., 2017). As the space of deep learning applications expands and starts to personalize, there is a growing need for the ability to quickly build and customize models. While model sizes have dropped dramatically from 50M parameters of the pioneering work of AlexNet (Krizhevsky et al., 2012) and VGG (Simonyan & Zisserman, 2014) to 5M of the recent Mobilenet (Sandler et al., 2018; Howard et al., 2017) and ShuffleNet (Zhang et al., 2017; Ma et al., 2018), the accuracy of models has been improving. However, delivering, maintaining and updating hundreds of models on the embedded device is still a significant expense in terms of bandwidth, energy and storage costs. While there still might be space for improvement in designing smaller models, in this paper we explore a different angle: we would like to be able to build models that require only a few parameters to be trained in order to be re-purposed to a different task, with minimal loss in accuracy compared to a model trained from scratch. While there is ample existing work on compressing models and learning as few weights as possible (Rosenfeld & Tsotsos, 2018; Sandler et al., 2018; Howard et al., 2017) to solve a single task, to the best of our awareness, there is no prior work that tries to minimize the number of model parameters when solving many tasks together. Our contribution is a novel learning paradigm in which each task carries its own model patch - a small set of parameters - that, along with a shared set of parameters constitutes the model for that task (Figure 1, left side).

CycleGAN (Zhu et al. 2017) is one recent successful approach to learn a transformation between two image distributions. In a series of experiments, we demonstrate an intriguing property of the model: CycleGAN learns to "hide" information about a source image into the images it generates in a nearly imperceptible, high-frequency signal. This trick ensures that the generator can recover the original sample and thus satisfy the cyclic consistency requirement, while the generated image remains realistic. We connect this phenomenon with adversarial attacks by viewing CycleGAN's training procedure as training a generator of adversarial examples and demonstrate that the cyclic consistency loss causes CycleGAN to be especially vulnerable to adversarial attacks.