Abstract:A fundamental challenge in machine learning is the choice of a loss as it characterizes our learning task, is minimized in the training phase, and serves as an evaluation criterion for estimators. Proper losses are commonly chosen, ensuring minimizers of the full risk match the true probability vector. Estimators induced from a proper loss are widely used to construct forecasters for downstream tasks such as classification and ranking. In this procedure, how does the forecaster based on the obtained estimator perform well under a given downstream task? This question is substantially relevant to the behavior of the $p$-norm between the estimated and true probability vectors when the estimator is updated. In the proper loss framework, the suboptimality of the estimated probability vector from the true probability vector is measured by a surrogate regret. First, we analyze a surrogate regret and show that the strict properness of a loss is necessary and sufficient to establish a non-vacuous surrogate regret bound. Second, we solve an important open question that the order of convergence in p-norm cannot be faster than the $1/2$-order of surrogate regrets for a broad class of strictly proper losses. This implies that strongly proper losses entail the optimal convergence rate.
Abstract:Gradient descent and its variants are de facto standard algorithms for training machine learning models. As gradient descent is sensitive to its hyperparameters, we need to tune the hyperparameters carefully using a grid search, but it is time-consuming, especially when multiple hyperparameters exist. Recently, parameter-free methods that adjust the hyperparameters on the fly have been studied. However, the existing work only studied parameter-free methods for the stepsize, and parameter-free methods for other hyperparameters have not been explored. For instance, the gradient clipping threshold is also a crucial hyperparameter in addition to the stepsize to prevent gradient explosion issues, but none of the existing studies investigated the parameter-free methods for clipped gradient descent. In this work, we study the parameter-free methods for clipped gradient descent. Specifically, we propose Inexact Polyak Stepsize, which converges to the optimal solution without any hyperparameters tuning, and its convergence rate is asymptotically independent of L under L-smooth and $(L_0, L_1)$-smooth assumptions of the loss function as that of clipped gradient descent with well-tuned hyperparameters. We numerically validated our convergence results using a synthetic function and demonstrated the effectiveness of our proposed methods using LSTM, Nano-GPT, and T5.
Abstract:SimSiam is a prominent self-supervised learning method that achieves impressive results in various vision tasks under static environments. However, it has two critical issues: high sensitivity to hyperparameters, especially weight decay, and unsatisfactory performance in online and continual learning, where neuroscientists believe that powerful memory functions are necessary, as in brains. In this paper, we propose PhiNet, inspired by a hippocampal model based on the temporal prediction hypothesis. Unlike SimSiam, which aligns two augmented views of the original image, PhiNet integrates an additional predictor block that estimates the original image representation to imitate the CA1 region in the hippocampus. Moreover, we model the neocortex inspired by the Complementary Learning Systems theory with a momentum encoder block as a slow learner, which works as long-term memory. We demonstrate through analysing the learning dynamics that PhiNet benefits from the additional predictor to prevent the complete collapse of learned representations, a notorious challenge in non-contrastive learning. This dynamics analysis may partially corroborate why this hippocampal model is biologically plausible. Experimental results demonstrate that PhiNet is more robust to weight decay and performs better than SimSiam in memory-intensive tasks like online and continual learning.
Abstract:In preference-based reinforcement learning (PbRL), a reward function is learned from a type of human feedback called preference. To expedite preference collection, recent works have leveraged \emph{offline preferences}, which are preferences collected for some offline data. In this scenario, the learned reward function is fitted on the offline data. If a learning agent exhibits behaviors that do not overlap with the offline data, the learned reward function may encounter generalizability issues. To address this problem, the present study introduces a framework that consolidates offline preferences and \emph{virtual preferences} for PbRL, which are comparisons between the agent's behaviors and the offline data. Critically, the reward function can track the agent's behaviors using the virtual preferences, thereby offering well-aligned guidance to the agent. Through experiments on continuous control tasks, this study demonstrates the effectiveness of incorporating the virtual preferences in PbRL.
Abstract:When dealing with data from distinct locations, machine learning algorithms tend to demonstrate an implicit preference of some locations over the others, which constitutes biases that sabotage the spatial fairness of the algorithm. This unfairness can easily introduce biases in subsequent decision-making given broad adoptions of learning-based solutions in practice. However, locational biases in AI are largely understudied. To mitigate biases over locations, we propose a locational meta-referee (Meta-Ref) to oversee the few-shot meta-training and meta-testing of a deep neural network. Meta-Ref dynamically adjusts the learning rates for training samples of given locations to advocate a fair performance across locations, through an explicit consideration of locational biases and the characteristics of input data. We present a three-phase training framework to learn both a meta-learning-based predictor and an integrated Meta-Ref that governs the fairness of the model. Once trained with a distribution of spatial tasks, Meta-Ref is applied to samples from new spatial tasks (i.e., regions outside the training area) to promote fairness during the fine-tune step. We carried out experiments with two case studies on crop monitoring and transportation safety, which show Meta-Ref can improve locational fairness while keeping the overall prediction quality at a similar level.
Abstract:This paper studies online structured prediction with full-information feedback. For online multiclass classification, van der Hoeven (2020) has obtained surrogate regret bounds independent of the time horizon, or \emph{finite}, by introducing an elegant \emph{exploit-the-surrogate-gap} framework. However, this framework has been limited to multiclass classification primarily because it relies on a classification-specific procedure for converting estimated scores to outputs. We extend the exploit-the-surrogate-gap framework to online structured prediction with \emph{Fenchel--Young losses}, a large family of surrogate losses including the logistic loss for multiclass classification, obtaining finite surrogate regret bounds in various structured prediction problems. To this end, we propose and analyze \emph{randomized decoding}, which converts estimated scores to general structured outputs. Moreover, by applying our decoding to online multiclass classification with the logistic loss, we obtain a surrogate regret bound of $O(B^2)$, where $B$ is the $\ell_2$-diameter of the domain. This bound is tight up to logarithmic factors and improves the previous bound of $O(dB^2)$ due to van der Hoeven (2020) by a factor of $d$, the number of classes.
Abstract:The self-attention mechanism prevails in modern machine learning. It has an interesting functionality of adaptively selecting tokens from an input sequence by modulating the degree of attention localization, which many researchers speculate is the basis of the powerful model performance but complicates the underlying mechanism of the learning dynamics. In recent years, mainly two arguments have connected attention localization to the model performances. One is the rank collapse, where the embedded tokens by a self-attention block become very similar across different tokens, leading to a less expressive network. The other is the entropy collapse, where the attention probability approaches non-uniform and entails low entropy, making the learning dynamics more likely to be trapped in plateaus. These two failure modes may apparently contradict each other because the rank and entropy collapses are relevant to uniform and non-uniform attention, respectively. To this end, we characterize the notion of attention localization by the eigenspectrum of query-key parameter matrices and reveal that a small eigenspectrum variance leads attention to be localized. Interestingly, the small eigenspectrum variance prevents both rank and entropy collapse, leading to better model expressivity and trainability.
Abstract:We propose Easymark, a family of embarrassingly simple yet effective watermarks. Text watermarking is becoming increasingly important with the advent of Large Language Models (LLM). LLMs can generate texts that cannot be distinguished from human-written texts. This is a serious problem for the credibility of the text. Easymark is a simple yet effective solution to this problem. Easymark can inject a watermark without changing the meaning of the text at all while a validator can detect if a text was generated from a system that adopted Easymark or not with high credibility. Easymark is extremely easy to implement so that it only requires a few lines of code. Easymark does not require access to LLMs, so it can be implemented on the user-side when the LLM providers do not offer watermarked LLMs. In spite of its simplicity, it achieves higher detection accuracy and BLEU scores than the state-of-the-art text watermarking methods. We also prove the impossibility theorem of perfect watermarking, which is valuable in its own right. This theorem shows that no matter how sophisticated a watermark is, a malicious user could remove it from the text, which motivate us to use a simple watermark such as Easymark. We carry out experiments with LLM-generated texts and confirm that Easymark can be detected reliably without any degradation of BLEU and perplexity, and outperform state-of-the-art watermarks in terms of both quality and reliability.
Abstract:In recent years, large language models (LLMs) have achieved remarkable performances in various NLP tasks. They can generate texts that are indistinguishable from those written by humans. Such remarkable performance of LLMs increases their risk of being used for malicious purposes, such as generating fake news articles. Therefore, it is necessary to develop methods for distinguishing texts written by LLMs from those written by humans. Watermarking is one of the most powerful methods for achieving this. Although existing watermarking methods have successfully detected texts generated by LLMs, they significantly degrade the quality of the generated texts. In this study, we propose the Necessary and Sufficient Watermark (NS-Watermark) for inserting watermarks into generated texts without degrading the text quality. More specifically, we derive minimum constraints required to be imposed on the generated texts to distinguish whether LLMs or humans write the texts. Then, we formulate the NS-Watermark as a constrained optimization problem and propose an efficient algorithm to solve it. Through the experiments, we demonstrate that the NS-Watermark can generate more natural texts than existing watermarking methods and distinguish more accurately between texts written by LLMs and those written by humans. Especially in machine translation tasks, the NS-Watermark can outperform the existing watermarking method by up to 30 BLEU scores.
Abstract:Contrastive learning is a self-supervised representation learning framework, where two positive views generated through data augmentation are made similar by an attraction force in a data representation space, while a repulsive force makes them far from negative examples. Non-contrastive learning, represented by BYOL and SimSiam, further gets rid of negative examples and improves computational efficiency. While learned representations may collapse into a single point due to the lack of the repulsive force at first sight, Tian et al. (2021) revealed through the learning dynamics analysis that the representations can avoid collapse if data augmentation is sufficiently stronger than regularization. However, their analysis does not take into account commonly-used feature normalization, a normalizer before measuring the similarity of representations, and hence excessively strong regularization may collapse the dynamics, which is an unnatural behavior under the presence of feature normalization. Therefore, we extend the previous theory based on the L2 loss by considering the cosine loss, which involves feature normalization. We show that the cosine loss induces sixth-order dynamics (while the L2 loss induces a third-order one), in which a stable equilibrium dynamically emerges even if there are only collapsed solutions with given initial parameters. Thus, we offer a new understanding that feature normalization plays an important role in robustly preventing the dynamics collapse.