On the other hand, recent finding on neural tangent kernel enables us to approximate a wide neural network with a linear model of the network's random features. In this paper, we theoretically analyze the knowledge distillation of a wide neural network. First we provide a transfer risk bound for the linearized model of the network. Then we propose a metric of the task's training difficulty, called data inefficiency.

In a model inversion (MI) attack, an adversary abuses access to a machine learning (ML) model to infer and reconstruct private training data. Remarkable progress has been made in the white-box and black-box setups, where the adversary has access to the complete model or the model's soft output respectively. However, there is very limited study in the most challenging but practically important setup: Label-only MI attacks, where the adversary only has access to the model's predicted label (hard label) without confidence scores nor any other model information. In this work, we propose LOKT, a novel approach for label-only MI attacks. Our idea is based on transfer of knowledge from the opaque target model to surrogate models.

On the other hand, recent finding on neural tangent kernel enables us to approximate a wide neural network with a linear model of the network's random features. In this paper, we theoretically analyze the knowledge distillation of a wide neural network. First we provide a transfer risk bound for the linearized model of the network. Then we propose a metric of the task's training difficulty, called data inefficiency.

As proved by Feng et al. (2021), the binary cross-entropy loss We include more results on teacher model and teacher model + {DRO (Hashimoto et al., 2018) /ARL (Lahoti et al., 2020) / FairRF (Zhao et al., 2022) /our knowledge distillation} in Tab. Effect of our label smoothing can be observed by comparing between "Teacher (with hard label)" and "Teacher (with softmax/linear label)" in the 6 tables. Here the capacity is the same, the only difference is the label smoothing. Here the training method is the same, only difference is capacity. Table 8: Results on COMP AS dataset with sensitive attribute race .

In scenarios where training data is limited due to observation costs or data scarcity, enriching the label information associated with each instance becomes crucial for building high-accuracy classification models. In such contexts, it is often feasible to obtain not only hard labels but also {\it additional supervision}, such as the confidences for the hard labels. This setting naturally raises fundamental questions: {\it What kinds of additional supervision are intrinsically beneficial?} And {\it how do they contribute to improved generalization performance?} To address these questions, we propose a theoretical framework that treats both hard labels and additional supervision as probability distributions, and constructs soft labels through their affine combination. Our theoretical analysis reveals that the essential component of additional supervision is not the confidence score of the assigned hard label, but rather the information of the distribution over the non-hard-labeled classes. Moreover, we demonstrate that the additional supervision and the mixing coefficient contribute to the refinement of soft labels in complementary roles. Intuitively, in the probability simplex, the additional supervision determines the direction in which the deterministic distribution representing the hard label should be adjusted toward the true label distribution, while the mixing coefficient controls the step size along that direction. Through generalization error analysis, we theoretically characterize how the additional supervision and its mixing coefficient affect both the convergence rate and asymptotic value of the error bound. Finally, we experimentally demonstrate that, based on our theory, designing additional supervision can lead to improved classification accuracy, even when utilized in a simple manner.

While the performance of machine learning systems has experienced significant improvement in recent years, relatively little attention has been paid to the fundamental question: to what extent can we improve our models? This paper provides a means of answering this question in the setting of binary classification, which is practical and theoretically supported. We extend a previous work that utilizes soft labels for estimating the Bayes error, the optimal error rate, in two important ways. First, we theoretically investigate the properties of the bias of the hard-label-based estimator discussed in the original work. We reveal that the decay rate of the bias is adaptive to how well the two class-conditional distributions are separated, and it can decay significantly faster than the previous result suggested as the number of hard labels per instance grows. Second, we tackle a more challenging problem setting: estimation with corrupted soft labels. One might be tempted to use calibrated soft labels instead of clean ones. However, we reveal that calibration guarantee is not enough, that is, even perfectly calibrated soft labels can result in a substantially inaccurate estimate. Then, we show that isotonic calibration can provide a statistically consistent estimator under an assumption weaker than that of the previous work. Our method is instance-free, i.e., we do not assume access to any input instances. This feature allows it to be adopted in practical scenarios where the instances are not available due to privacy issues. Experiments with synthetic and real-world datasets show the validity of our methods and theory.

In a model inversion (MI) attack, an adversary abuses access to a machine learning (ML) model to infer and reconstruct private training data. Remarkable progress has been made in the white-box and black-box setups, where the adversary has access to the complete model or the model's soft output respectively. However, there is very limited study in the most challenging but practically important setup: Label-only MI attacks, where the adversary only has access to the model's predicted label (hard label) without confidence scores nor any other model information. In this work, we propose LOKT, a novel approach for label-only MI attacks. Our idea is based on transfer of knowledge from the opaque target model to surrogate models.

The Myers-Briggs Type Indicator (MBTI) is one of the most influential personality theories reflecting individual differences in thinking, feeling, and behaving. MBTI personality detection has garnered considerable research interest and has evolved significantly over the years. However, this task tends to be overly optimistic, as it currently does not align well with the natural distribution of population personality traits. Specifically, (1) the self-reported labels in existing datasets result in incorrect labeling issues, and (2) the hard labels fail to capture the full range of population personality distributions. In this paper, we optimize the task by constructing MBTIBench, the first manually annotated high-quality MBTI personality detection dataset with soft labels, under the guidance of psychologists. As for the first challenge, MBTIBench effectively solves the incorrect labeling issues, which account for 29.58% of the data. As for the second challenge, we estimate soft labels by deriving the polarity tendency of samples. The obtained soft labels confirm that there are more people with non-extreme personality traits. Experimental results not only highlight the polarized predictions and biases in LLMs as key directions for future research, but also confirm that soft labels can provide more benefits to other psychological tasks than hard labels. The code and data are available at https://github.com/Personality-NLP/MbtiBench.

In supervised machine learning, models are typically trained using data with hard labels, i.e., definite assignments of class membership. This traditional approach, however, does not take the inherent uncertainty in these labels into account. We investigate whether incorporating label uncertainty, represented as discrete probability distributions over the class labels -- known as soft labels -- improves the predictive performance of classification models. We first demonstrate the potential value of soft label learning (SLL) for estimating model parameters in a simulation experiment, particularly for limited sample sizes and imbalanced data. Subsequently, we compare the performance of various wrapper methods for learning from both hard and soft labels using identical base classifiers. On real-world-inspired synthetic data with clean labels, the SLL methods consistently outperform hard label methods. Since real-world data is often noisy and precise soft labels are challenging to obtain, we study the effect that noisy probability estimates have on model performance. Alongside conventional noise models, our study examines four types of miscalibration that are known to affect human annotators. The results show that SLL methods outperform the hard label methods in the majority of settings. Finally, we evaluate the methods on a real-world dataset with confidence scores, where the SLL methods are shown to match the traditional methods for predicting the (noisy) hard labels while providing more accurate confidence estimates.