Zhenisbek Assylbekov

Rustem Takhanov, Zhenisbek Assylbekov

Conditionally positive definite (CPD) kernels are defined with respect to a function class $\mathcal{F}$. It is well known that such a kernel $K$ is associated with its native space (defined analogously to an RKHS), which in turn gives rise to a learning method -- called conditional kernel ridge regression (conditional KRR) due to its analogy with KRR -- where the estimated regression function is penalized by the square of its native space norm. This method is of interest because it can be viewed as classical linear regression, with features specified by $\mathcal{F}$, followed by the application of standard KRR to the residual (unexplained) component of the target variable. Methods of this type have recently attracted increasing attention. We study the statistical properties of this method by reducing its behavior to that of KRR with another fixed kernel, called the residual kernel. Our main theoretical result shows that such a reduction is indeed possible, at the cost of an additional term in the expected test risk, bounded by $\mathcal{O}(1/\sqrt{N})$, where $N$ is the sample size and the hidden constant depends on the class $\mathcal{F}$ and the input distribution. This reduction enables us to analyze conditional KRR in the case where $K$ is positive definite and $\mathcal{F}$ is given by the first $k$ principal eigenfunctions in the Mercer decomposition of $K$. We also consider the setting where $\mathcal{F}$ consists of $k$ random features from a random feature representation of $K$. It turns out that these two settings are closely related. Both our theoretical analysis and experiments confirm that conditional KRR outperforms standard KRR in these cases whenever the $\mathcal{F}$-component of the regression function is more pronounced than the residual part.

Arman Bolatov, Maxat Tezekbayev, Igor Melnykov et al.

We suggest a simple Gaussian mixture model for data generation that complies with Feldman's long tail theory (2020). We demonstrate that a linear classifier cannot decrease the generalization error below a certain level in the proposed model, whereas a nonlinear classifier with a memorization capacity can. This confirms that for long-tailed distributions, rare training examples must be considered for optimal generalization to new data. Finally, we show that the performance gap between linear and nonlinear models can be lessened as the tail becomes shorter in the subpopulation frequency distribution, as confirmed by experiments on synthetic and real data.

Sultan Nurmukhamedov, Thomas Mach, Arsen Sheverdin et al.

We choose random points in the hyperbolic disc and claim that these points are already word representations. However, it is yet to be uncovered which point corresponds to which word of the human language of interest. This correspondence can be approximately established using a pointwise mutual information between words and recent alignment techniques.

Rustem Takhanov, Maxat Tezekbayev, Artur Pak et al.

Classes of target functions containing a large number of approximately orthogonal elements are known to be hard to learn by the Statistical Query algorithms. Recently this classical fact re-emerged in a theory of gradient-based optimization of neural networks. In the novel framework, the hardness of a class is usually quantified by the variance of the gradient with respect to a random choice of a target function. A set of functions of the form $x\to ax \bmod p$, where $a$ is taken from ${\mathbb Z}_p$, has attracted some attention from deep learning theorists and cryptographers recently. This class can be understood as a subset of $p$-periodic functions on ${\mathbb Z}$ and is tightly connected with a class of high-frequency periodic functions on the real line. We present a mathematical analysis of limitations and challenges associated with using gradient-based learning techniques to train a high-frequency periodic function or modular multiplication from examples. We highlight that the variance of the gradient is negligibly small in both cases when either a frequency or the prime base $p$ is large. This in turn prevents such a learning algorithm from being successful.

Rustem Takhanov, Maxat Tezekbayev, Artur Pak et al.

The discrete logarithm problem is a fundamental challenge in number theory with significant implications for cryptographic protocols. In this paper, we investigate the limitations of gradient-based methods for learning the parity bit of the discrete logarithm in finite cyclic groups of prime order. Our main result, supported by theoretical analysis and empirical verification, reveals the concentration of the gradient of the loss function around a fixed point, independent of the logarithm's base used. This concentration property leads to a restricted ability to learn the parity bit efficiently using gradient-based methods, irrespective of the complexity of the network architecture being trained. Our proof relies on Boas-Bellman inequality in inner product spaces and it involves establishing approximate orthogonality of discrete logarithm's parity bit functions through the spectral norm of certain matrices. Empirical experiments using a neural network-based approach further verify the limitations of gradient-based learning, demonstrating the decreasing success rate in predicting the parity bit as the group order increases.

Arman Bolatov, Alan Legg, Igor Melnykov et al.

This study explores the classification error of Mixture Discriminant Analysis (MDA) in scenarios where the number of mixture components exceeds those present in the actual data distribution, a condition known as overspecification. We use a two-component Gaussian mixture model within each class to fit data generated from a single Gaussian, analyzing both the algorithmic convergence of the Expectation-Maximization (EM) algorithm and the statistical classification error. We demonstrate that, with suitable initialization, the EM algorithm converges exponentially fast to the Bayes risk at the population level. Further, we extend our results to finite samples, showing that the classification error converges to Bayes risk with a rate $n^{-1/2}$ under mild conditions on the initial parameter estimates and sample size. This work provides a rigorous theoretical framework for understanding the performance of overspecified MDA, which is often used empirically in complex data settings, such as image and text classification. To validate our theory, we conduct experiments on remote sensing datasets.

Maxat Tezekbayev, Arman Bolatov, Zhenisbek Assylbekov

We revisit Deep Linear Discriminant Analysis (Deep LDA) from a likelihood-based perspective. While classical LDA is a simple Gaussian model with linear decision boundaries, attaching an LDA head to a neural encoder raises the question of how to train the resulting deep classifier by maximum likelihood estimation (MLE). We first show that end-to-end MLE training of an unconstrained Deep LDA model ignores discrimination: when both the LDA parameters and the encoder parameters are learned jointly, the likelihood admits a degenerate solution in which some of the class clusters may heavily overlap or even collapse, and classification performance deteriorates. Batchwise moment re-estimation of the LDA parameters does not remove this failure mode. We then propose a constrained Deep LDA formulation that fixes the class means to the vertices of a regular simplex in the latent space and restricts the shared covariance to be spherical, leaving only the priors and a single variance parameter to be learned along with the encoder. Under these geometric constraints, MLE becomes stable and yields well-separated class clusters in the latent space. On images (Fashion-MNIST, CIFAR-10, CIFAR-100), the resulting Deep LDA models achieve accuracy competitive with softmax baselines while offering a simple, interpretable latent geometry that is clearly visible in two-dimensional projections.

Maxat Tezekbayev, Rustem Takhanov, Arman Bolatov et al.

We show that for unconstrained Deep Linear Discriminant Analysis (LDA) classifiers, maximum-likelihood training admits pathological solutions in which class means drift together, covariances collapse, and the learned representation becomes almost non-discriminative. Conversely, cross-entropy training yields excellent accuracy but decouples the head from the underlying generative model, leading to highly inconsistent parameter estimates. To reconcile generative structure with discriminative performance, we introduce the \emph{Discriminative Negative Log-Likelihood} (DNLL) loss, which augments the LDA log-likelihood with a simple penalty on the mixture density. DNLL can be interpreted as standard LDA NLL plus a term that explicitly discourages regions where several classes are simultaneously likely. Deep LDA trained with DNLL produces clean, well-separated latent spaces, matches the test accuracy of softmax classifiers on synthetic data and standard image benchmarks, and yields substantially better calibrated predictive probabilities, restoring a coherent probabilistic interpretation to deep discriminant models.

Zhenisbek Assylbekov, Alan Legg, Artur Pak

We investigate the convergence properties of the EM algorithm when applied to overspecified Gaussian mixture models -- that is, when the number of components in the fitted model exceeds that of the true underlying distribution. Focusing on a structured configuration where the component means are positioned at the vertices of a regular simplex and the mixture weights satisfy a non-degeneracy condition, we demonstrate that the population EM algorithm converges exponentially fast in terms of the Kullback-Leibler (KL) distance. Our analysis leverages the strong convexity of the negative log-likelihood function in a neighborhood around the optimum and utilizes the Polyak-Łojasiewicz inequality to establish that an $ε$-accurate approximation is achievable in $O(\log(1/ε))$ iterations. Furthermore, we extend these results to a finite-sample setting by deriving explicit statistical convergence guarantees. Numerical experiments on synthetic datasets corroborate our theoretical findings, highlighting the dramatic acceleration in convergence compared to conventional sublinear rates. This work not only deepens the understanding of EM's behavior in overspecified settings but also offers practical insights into initialization strategies and model design for high-dimensional clustering and density estimation tasks.

Maxat Tezekbayev, Vassilina Nikoulina, Matthias Gallé et al.

Softmax is the de facto standard in modern neural networks for language processing when it comes to normalizing logits. However, by producing a dense probability distribution each token in the vocabulary has a nonzero chance of being selected at each generation step, leading to a variety of reported problems in text generation. $α$-entmax of Peters et al. (2019, arXiv:1905.05702) solves this problem, but is considerably slower than softmax. In this paper, we propose an alternative to $α$-entmax, which keeps its virtuous characteristics, but is as fast as optimized softmax and achieves on par or better performance in machine translation task.

Vassilina Nikoulina, Maxat Tezekbayev, Nuradil Kozhakhmet et al.

There is an ongoing debate in the NLP community whether modern language models contain linguistic knowledge, recovered through so-called probes. In this paper, we study whether linguistic knowledge is a necessary condition for the good performance of modern language models, which we call the \textit{rediscovery hypothesis}. In the first place, we show that language models that are significantly compressed but perform well on their pretraining objectives retain good scores when probed for linguistic structures. This result supports the rediscovery hypothesis and leads to the second contribution of our paper: an information-theoretic framework that relates language modeling objectives with linguistic information. This framework also provides a metric to measure the impact of linguistic information on the word prediction task. We reinforce our analytical results with various experiments, both on synthetic and on real NLP tasks in English.

Zhenisbek Assylbekov, Alibi Jangeldin

We show that removing sigmoid transformation in the skip-gram with negative sampling (SGNS) objective does not harm the quality of word vectors significantly and at the same time is related to factorizing a squashed shifted PMI matrix which, in turn, can be treated as a connection probabilities matrix of a random graph. Empirically, such graph is a complex network, i.e. it has strong clustering and scale-free degree distribution, and is tightly connected with hyperbolic spaces. In short, we show the connection between static word embeddings and hyperbolic spaces through the squashed shifted PMI matrix using analytical and empirical methods.

Maxat Tezekbayev, Zhenisbek Assylbekov, Rustem Takhanov

We show that the skip-gram embedding of any word can be decomposed into two subvectors which roughly correspond to semantic and syntactic roles of the word.

Aidana Karipbayeva, Alena Sorokina, Zhenisbek Assylbekov

We critically review the smooth inverse frequency sentence embedding method of Arora, Liang, and Ma (2017), and show inconsistencies in its setup, derivation, and evaluation.

Alena Sorokina, Aidana Karipbayeva, Zhenisbek Assylbekov

We perform an empirical evaluation of several methods of low-rank approximation in the problem of obtaining PMI-based word embeddings. All word vectors were trained on parts of a large corpus extracted from English Wikipedia (enwik9) which was divided into two equal-sized datasets, from which PMI matrices were obtained. A repeated measures design was used in assigning a method of low-rank approximation (SVD, NMF, QR) and dimensionality of the vectors (250, 500) to each of the PMI matrix replicates. Our experiments show that word vectors obtained from the truncated SVD achieve the best performance on two downstream tasks, similarity and analogy, compare to the other two low-rank approximation methods.

Zhenisbek Assylbekov, Rustem Takhanov

This paper takes a step towards theoretical analysis of the relationship between word embeddings and context embeddings in models such as word2vec. We start from basic probabilistic assumptions on the nature of word vectors, context vectors, and text generation. These assumptions are well supported either empirically or theoretically by the existing literature. Next, we show that under these assumptions the widely-used word-word PMI matrix is approximately a random symmetric Gaussian ensemble. This, in turn, implies that context vectors are reflections of word vectors in approximately half the dimensions. As a direct application of our result, we suggest a theoretically grounded way of tying weights in the SGNS model.

Abylay Zhumekenov, Malika Uteuliyeva, Olzhas Kabdolov et al.

We review neural network architectures which were motivated by Fourier series and integrals and which are referred to as Fourier neural networks. These networks are empirically evaluated in synthetic and real-world tasks. Neither of them outperforms the standard neural network with sigmoid activation function in the real-world tasks. All neural networks, both Fourier and the standard one, empirically demonstrate lower approximation error than the truncated Fourier series when it comes to an approximation of a known function of multiple variables.

Zhenisbek Assylbekov, Rustem Takhanov

We propose several ways of reusing subword embeddings and other weights in subword-aware neural language models. The proposed techniques do not benefit a competitive character-aware model, but some of them improve the performance of syllable- and morpheme-aware models while showing significant reductions in model sizes. We discover a simple hands-on principle: in a multi-layer input embedding model, layers should be tied consecutively bottom-up if reused at output. Our best morpheme-aware model with properly reused weights beats the competitive word-level model by a large margin across multiple languages and has 20%-87% fewer parameters.

Rustem Takhanov, Zhenisbek Assylbekov

Words in some natural languages can have a composite structure. Elements of this structure include the root (that could also be composite), prefixes and suffixes with which various nuances and relations to other words can be expressed. Thus, in order to build a proper word representation one must take into account its internal structure. From a corpus of texts we extract a set of frequent subwords and from the latter set we select patterns, i.e. subwords which encapsulate information on character $n$-gram regularities. The selection is made using the pattern-based Conditional Random Field model with $l_1$ regularization. Further, for every word we construct a new sequence over an alphabet of patterns. The new alphabet's symbols confine a local statistical context stronger than the characters, therefore they allow better representations in ${\mathbb{R}}^n$ and are better building blocks for word representation. In the task of subword-aware language modeling, pattern-based models outperform character-based analogues by 2-20 perplexity points. Also, a recurrent neural network in which a word is represented as a sum of embeddings of its patterns is on par with a competitive and significantly more sophisticated character-based convolutional architecture.

Zhenisbek Assylbekov, Rustem Takhanov, Bagdat Myrzakhmetov et al.

Syllabification does not seem to improve word-level RNN language modeling quality when compared to character-based segmentation. However, our best syllable-aware language model, achieving performance comparable to the competitive character-aware model, has 18%-33% fewer parameters and is trained 1.2-2.2 times faster.