Latin cubes of order 4 and 4320-point quaternionic constructions in ℝ¹⁶
Constructs a family of 4,320-point kissing arrangements in sixteen dimensions using Latin cubes of order four and permutations of quaternionic levels.
Research & writing
Selected papers on kissing numbers, kernel methods, constraint satisfaction, language modeling, and other machine learning topics. Explore each paper for a short overview and accompanying resources.
Constructs a family of 4,320-point kissing arrangements in sixteen dimensions using Latin cubes of order four and permutations of quaternionic levels.
New constructions improve lower bounds for kissing numbers in dimensions 25, 26, 27, 29, and 31.
A quaternionic viewpoint reveals the structure behind a numerical 841-point kissing arrangement in twelve dimensions.
Uses flexibility in 840-point arrangements to guide optimization toward an 841-point kissing arrangement in twelve dimensions.
Classifies maximum independent sets in signed Johnson graphs and reveals a rich family of kissing arrangements.
Combines unpenalized features with kernel ridge regression, giving a statistical framework for learning the residual component.
How does the geometry of data influence kernel learning? Two notions of intrinsic dimension offer complementary answers.
Relates the information in a learning gradient to pairwise independence and the collision entropy of the input distribution.
Analyzes why gradient-based methods struggle with high-frequency periodic functions and modular multiplication.
Spectral bounds on the ε-entropy of kernel classes, linking approximation complexity to kernel eigenvalues.
Studies how almost pairwise independence limits the information available to gradient-based learning attacks.
Connects neural network approximation guarantees to the RKHS of the neural network Gaussian process kernel.
Investigates how algebraic structure in a CSP solution space can be encoded as another fixed-template problem.
Uses Jacobian rank constraints to encourage autoencoders to learn low-dimensional manifolds.
Algorithms for partition functions and minimization in generalized pattern-based sequence models over semirings.
Quantifies how quickly truncated Mercer expansions converge uniformly to a positive definite kernel.
Reformulates dimension reduction through tempered distributions supported on low-dimensional subspaces.
An algebraic study of adding implied constraints and finding compact equivalent descriptions of CSP instances.
Explains a geometric relationship between word and context embeddings: reflection in approximately half the dimensions.
Reproduces the Structurally Constrained Recurrent Network and studies regularization for neural language modeling.
Explores weight sharing in subword-aware language models to reduce model size while improving selected architectures.
Compares syllable-based and character-based word representations for recurrent neural language models.
Uses statistically informative subword patterns as building blocks for word representations and neural language modeling.
Studies hybrid constraint satisfaction through crisp and conservative valued templates and the notion of wide tractability.
Characterizes when structural restrictions can make an otherwise hard constraint satisfaction problem tractable.
Combines local pattern-based sequence models with the longer-range structure of weighted context-free grammars.
Efficient inference for sequence models with pattern-based potentials: partition functions, marginals, MAP inference, and sampling.
Extends minimum cost homomorphisms to a broader optimization framework with conservative constraint languages.
An algebraic classification of the computational complexity of the general minimum cost homomorphism problem.
Studies the sample monotonization problem, connecting ordered data with algorithmic questions in learning.
Try another title, author, or research topic.