SearcharxivSearch

arXiv subjects

Youngchae Kim

Publications and source records attributed to Youngchae Kim.

3 recordsLinked to original sources

Absence of the Lavrentiev phenomenon for a general class of parabolic double phase problems

In this paper, we prove the absence of the Lavrentiev phenomenon for a general class of parabolic double phase functionals with Orlicz growth. The energy density is given by $$ G(|Dw|)+a(x,t)H(|Dw|), $$ where $G$ and $H$ are Young functions satisfying the $\Delta_2$ and $\nabla_2$ conditions with $G\prec H$, and $a(\cdot)$ is a continuous nonnegative coefficient. Under suitable balance conditions between the growth gap of $G$ and $H$ and the modulus of continuity of $a(\cdot)$, we show that every finite-energy map can be approximated locally by smooth functions without loss of energy. The result extends the known parabolic double phase theory from power type growth to a broad Young function framework and identifies the natural space-time Orlicz energy class for the problem.

math.AP

Absence of the Lavrentiev phenomenon for degenerate parabolic double phase problems

We establish the absence of the Lavrentiev phenomenon for degenerate parabolic double phase problems. Any finite-energy function in the natural parabolic class admits smooth approximations with convergence in the parabolic Sobolev space and convergence of the corresponding energy. We provide explicit gap bound conditions and derive improved bounds under additional assumptions such as boundedness or stronger time regularity.

math.AP

IterL2Norm: Fast Iterative L2-Normalization

Transformer-based large language models are a memory-bound model whose operation is based on a large amount of data that are marginally reused. Thus, the data movement between a host and accelerator likely dictates the total wall-clock time. Layer normalization is one of the key workloads in the transformer model, following each of multi-head attention and feed-forward network blocks. To reduce data movement, layer normalization needs to be performed on the same chip as the matrix-matrix multiplication engine. To this end, we introduce an iterative L2-normalization method for 1D input (IterL2Norm), ensuring fast convergence to the steady-state solution within five iteration steps and high precision, outperforming the fast inverse square root algorithm in six out of nine cases for FP32 and five out of nine for BFloat16 across the embedding lengths used in the OPT models. Implemented in 32/28nm CMOS, the IterL2Norm macro normalizes $d$-dimensional vectors, where $64 \leq d \leq 1024$, with a latency of 116-227 cycles at 100MHz/1.05V.

cs.LG