Post-training with reinforcement learning (RL) typically optimizes a single scalar objective and ignores structure in how solutions are produced. We ask whether a scalar hint toward a canonical solver ordering, used only during RL post-training, improves performance even when fine-tuned on randomized solution sequences. On Sudoku, we train a Transformer with standard fine-tuning on randomized solving orders, then post-train it with Group Relative Policy Optimization (GRPO) with two rewards: cell accuracy and an ordering reward that increases when the model's emission order aligns with the solver order. To compare signals cleanly, we combine them via fixed mixtures and use a simple bootstrapped scaling to equalize component magnitudes at initialization. Mixed rewards generally outperform cell-only optimization--the best mixture yields substantially higher test accuracy than the fine-tuned-only model trained on random-order and approaches the fine-tuned-only model trained on solver-order sequences in accuracy. These results suggest that coarse ordering signals can steer RL post-training toward solver-order trajectories without modifying supervised data or architecture.

Multilingual large language models (LLMs) are known to more frequently generate non-faithful output in resource-constrained languages (Guerreiro et al., 2023 - arXiv:2303.16104), potentially because these typologically diverse languages are underrepresented in their training data. To mitigate unfaithfulness in such settings, we propose using computationally light auxiliary models to rescore the outputs of larger architectures. As proof of the feasibility of such an approach, we show that monolingual 4-layer BERT models pretrained from scratch on less than 700 MB of data without fine-tuning are able to identify faithful summaries with a mean accuracy of 88.33% in three genetically unrelated languages that differ in their morphological complexity - Vietnamese, Polish and Georgian. The same hyperparameter combination moreover generalises well to three other tasks, suggesting applications for rescoring beyond improving faithfulness. In order to inform typologically aware model selection, we also investigate how morphological complexity interacts with regularisation, model depth and training objectives, ultimately demonstrating that morphologically complex languages are more likely to benefit from dropout, while across languages downstream performance is enhanced most by shallow architectures as well as training using the standard BERT objectives.

Learned optimization has emerged as a promising alternative to hand-crafted optimizers, with the potential to discover stronger learned update rules that enable faster, hyperparameter-free training of neural networks. A critical element for practically useful learned optimizers, that can be used off-the-shelf after meta-training, is strong meta-generalization: the ability to apply the optimizers to new tasks. Recent state-of-the-art work in learned optimizers, VeLO (Metz et al., 2022), requires a large number of highly diverse meta-training tasks along with massive computational resources, 4000 TPU months, to achieve meta-generalization. This makes further improvements to such learned optimizers impractical. In this work, we identify several key elements in learned optimizer architectures and meta-training procedures that can lead to strong meta-generalization. We also propose evaluation metrics to reliably assess quantitative performance of an optimizer at scale on a set of evaluation tasks. Our proposed approach, Celo, makes a significant leap in improving the meta-generalization performance of learned optimizers and also outperforms tuned state-of-the-art optimizers on a diverse set of out-of-distribution tasks, despite being meta-trained for just 24 GPU hours.

We study the bias of Stochastic Gradient Descent (SGD) to learn low-rank weight matrices when training deep neural networks. Our results show that training neural networks with mini-batch SGD and weight decay causes a bias towards rank minimization over the weight matrices. Specifically, we show, both theoretically and empirically, that this bias is more pronounced when using smaller batch sizes, higher learning rates, or increased weight decay. Additionally, we predict and observe empirically that weight decay is necessary to achieve this bias. Unlike previous literature, our analysis does not rely on assumptions about the data, convergence, or optimality of the weight matrices and applies to a wide range of neural network architectures of any width or depth. Finally, we empirically investigate the connection between this bias and generalization, finding that it has a marginal effect on generalization.

Deep neural network models are massively deployed on a wide variety of hardware platforms. This results in the appearance of new attack vectors that significantly extend the standard attack surface, extensively studied by the adversarial machine learning community. One of the first attack that aims at drastically dropping the performance of a model, by targeting its parameters (weights) stored in memory, is the Bit-Flip Attack (BFA). In this work, we point out several evaluation challenges related to the BFA. First of all, the lack of an adversary's budget in the standard threat model is problematic, especially when dealing with physical attacks. Moreover, since the BFA presents critical variability, we discuss the influence of some training parameters and the importance of the model architecture. This work is the first to present the impact of the BFA against fully-connected architectures that present different behaviors compared to convolutional neural networks. These results highlight the importance of defining robust and sound evaluation methodologies to properly evaluate the dangers of parameter-based attacks as well as measure the real level of robustness offered by a defense.

Momentum based stochastic gradient methods such as heavy ball (HB) and Nesterov's accelerated gradient descent (NAG) method are widely used in practice for training deep networks and other supervised learning models, as they often provide significant improvements over stochastic gradient descent (SGD). Rigorously speaking, "fast gradient" methods have provable improvements over gradient descent only for the deterministic case, where the gradients are exact. In the stochastic case, the popular explanations for their wide applicability is that when these fast gradient methods are applied in the stochastic case, they partially mimic their exact gradient counterparts, resulting in some practical gain. This work provides a counterpoint to this belief by proving that there exist simple problem instances where these methods cannot outperform SGD despite the best setting of its parameters. These negative problem instances are, in an informal sense, generic; they do not look like carefully constructed pathological instances. These results suggest (along with empirical evidence) that HB or NAG's practical performance gains are a by-product of mini-batching. Furthermore, this work provides a viable (and provable) alternative, which, on the same set of problem instances, significantly improves over HB, NAG, and SGD's performance. This algorithm, referred to as Accelerated Stochastic Gradient Descent (ASGD), is a simple to implement stochastic algorithm, based on a relatively less popular variant of Nesterov's Acceleration. Extensive empirical results in this paper show that ASGD has performance gains over HB, NAG, and SGD.

BEIJING, CHINA--(Marketwired - Sep 12, 2016) - GPU Technology Conference China - NVIDIA (NASDAQ: NVDA) today unveiled the latest additions to its Pascal architecture-based deep learning platform, with new NVIDIA Tesla P4 and P40 GPU accelerators and new software that deliver massive leaps in efficiency and speed to accelerate inferencing production workloads for artificial intelligence services. Modern AI services such as voice-activated assistance, email spam filters, and movie and product recommendation engines are rapidly growing in complexity, requiring up to 10x more compute compared to neural networks from a year ago. Current CPU-based technology isn't capable of delivering real-time responsiveness required for modern AI services, leading to a poor user experience. The Tesla P4 and P40 are specifically designed for inferencing, which uses trained deep neural networks to recognize speech, images or text in response to queries from users and devices. Based on the Pascal architecture, these GPUs feature specialized inference instructions based on 8-bit (INT8) operations, delivering 45x faster response than CPUs1 and a 4x improvement over GPU solutions launched less than a year ago.2

Results will be integrated into the above repository eventually. The Recurrent Networks take as input a 3D Tensor batch_size x seq_length x hidden_size and output the last hidden state, compute a MSE loss and compute the gradients of error with respect to each parameter. The hidden_size specifies the size of the output and input layer of the networks. The code of the scripts we ran are available. The code for the regular theano RNN implementations were borrowed from the rnn-benchmarks repository.