Self-distillation (SD), a technique where a model improves itself using its own predictions, has attracted attention as a simple yet powerful approach in machine learning. Despite its widespread use, the mechanisms underlying its effectiveness remain unclear. In this study, we investigate the efficacy of hyperparameter-tuned multi-stage SD with a linear classifier for binary classification on noisy Gaussian mixture data. For the analysis, we employ the replica method from statistical physics. Our findings reveal that the primary driver of SD's performance improvement is denoising through hard pseudo-labels, with the most notable gains observed in moderately sized datasets. We also identify two practical heuristics to enhance SD: early stopping that limits the number of stages, which is broadly effective, and bias parameter fixing, which helps under label imbalance. To empirically validate our theoretical findings derived from our toy model, we conduct additional experiments on CIFAR-10 classification using pretrained ResNet backbone. These results provide both theoretical and practical insights, advancing our understanding and application of SD in noisy settings.

This paper examines the quantization methods used in large-scale data analysis models and their hyperparameter choices. The recent surge in data analysis scale has significantly increased computational resource requirements. To address this, quantizing model weights has become a prevalent practice in data analysis applications such as deep learning. Quantization is particularly vital for deploying large models on devices with limited computational resources. However, the selection of quantization hyperparameters, like the number of bits and value range for weight quantization, remains an underexplored area. In this study, we employ the typical case analysis from statistical physics, specifically the replica method, to explore the impact of hyperparameters on the quantization of simple learning models. Our analysis yields three key findings: (i) an unstable hyperparameter phase, known as replica symmetry breaking, occurs with a small number of bits and a large quantization width; (ii) there is an optimal quantization width that minimizes error; and (iii) quantization delays the onset of overparameterization, which mitigate overfitting as indicated by the double descent phenomenon. We also discover that non-uniform quantization can enhance stability. Additionally, we develop an approximate message-passing algorithm to validate our theoretical results.

We study the inference problem in the noisy group testing to identify defective items from the perspective of the decision theory. We introduce Bayesian inference and consider the Bayesian optimal setting in which the true generative process of the test results is known. We demonstrate the adequacy of the posterior marginal probability in the Bayesian optimal setting as a diagnostic variable based on the area under the curve (AUC). Using the posterior marginal probability, we derive the general expression of the optimal cutoff value that yields the minimum expected risk function. Furthermore, we evaluate the performance of the Bayesian group testing without knowing the true states of the items: defective or non-defective. By introducing an analytical method from statistical physics, we derive the receiver operating characteristics curve, and quantify the corresponding AUC under the Bayesian optimal setting. The obtained analytical results precisely describes the actual performance of the belief propagation algorithm defined for single samples when the number of items is sufficiently large.

We analyze the matrix factorization problem. Given a noisy measurement of a product of two matrices, the problem is to estimate back the original matrices. It arises in many applications, such as dictionary learning, blind matrix calibration, sparse principal component analysis, blind source separation, low rank matrix completion, robust principal component analysis, or factor analysis. It is also important in machine learning: unsupervised representation learning can often be studied through matrix factorization. We use the tools of statistical mechanics-the cavity and replica methods-to analyze the achievability and computational tractability of the inference problems in the setting of Bayes-optimal inference, which amounts to assuming that the two matrices have random-independent elements generated from some known distribution, and this information is available to the inference algorithm. In this setting, we compute the minimal mean-squared-error achievable, in principle, in any computational time, and the error that can be achieved by an efficient approximate message passing algorithm. The computation is based on the asymptotic state-evolution analysis of the algorithm. The performance that our analysis predicts, both in terms of the achieved mean-squared-error and in terms of sample complexity, is extremely promising and motivating for a further development of the algorithm.

We develop a method for evaluating restricted isometry constants (RICs). This evaluation is reduced to the identification of the zero-points of entropy density which is defined for submatrices that are composed of columns selected from a given measurement matrix. Using the replica method developed in statistical mechanics, we assess RICs for Gaussian random matrices under the replica symmetric (RS) assumption. In order to numerically validate the adequacy of our analysis, we employ the exchange Monte Carlo (EMC) method, which has been empirically demonstrated to achieve much higher numerical accuracy than naive Monte Carlo methods. The EMC method suggests that our theoretical estimation of an RIC corresponds to an upper bound that is tighter than in preceding studies. Physical consideration indicates that our assessment of the RIC could be improved by taking into account the replica symmetry breaking.

We consider a learning problem of identifying a dictionary matrix D ∈ RM×Nfrom a sample set of M dimensional vectors Y ∈ RM×P= N-1/2DX ∈ RM×P, where X ∈ RN×pis a sparse matrix in which the density of non-zero entries is 0c(sample complexity) necessary for perfectly identifying D of the optimal learning scheme when D and X are independently generated from certain distributions. By using the replica method of statistical mechanics, we show that Pc~ O(N) holds as long as α = M/N > ρ is satisfied in the limit of N → ∞. Our analysis also implies that the posterior distribution given Y is condensed only at the correct dictionary D when the compression rate α is greater than a certain critical value αM(p). This suggests that belief propagation may allow us to learn D with a low computational complexity using O(N) samples.