Sirin Schariac

Thinking will not overcome fear but action will.

UMich DL for CV

Generative Models-2

Generative Adversarial Networks Generative Models Review Autoregressive Models directly maximize likelihood of training data \(p_{\theta}(x)=\prod_{i=1}^N p_{\theta}(x_i|x_1,\dots,x_{i-1})\) Var...

Posted by Sirin on April 20, 2025

UMich DL for CV

Generative Models-1

判别模型与生成模型 Component Supervised Learning Unsupervised Learning Data (x: data, y: label) (x: data) Goal learn a fun...

Posted by Sirin on April 17, 2025

UMich DL for CV

Attention

Attention Machine Translation (seq2seq) Here $a_{ij}$ represents the weights of input sequence predicted by the attention. e.g., if the word ‘estamos’ = ‘eating’, then maybe $a_{23}=0.8\;a_{21...

Posted by Sirin on March 31, 2025

UMich DL for CV

Recurrent Neural Networks

Recurrent Neural Networks Key idea RNNs have an “internal state” that is updated as a sequence is processed, the recurrence formula looks like $h_t = f_W(h_{t-1},x_t)$. Here $h_t$ is the state a...

Posted by Sirin on March 18, 2025

UMich DL for CV

Convoluntional Network

Convolutional Network Why Convolutional Network? So far our classifiers don’t respect the spatial structure of images (merely stretch pixels into column). Thus new operators is needed to han...

Posted by Sirin on February 28, 2025

Continual Learning

An Intro to continual learning

Continual Learning What is continual learning 一般来说，ML是在一个固定的数据集上针对特定的下流任务来训练一个相应的模型。但对于现实世界的任务而言，环境是在不断变化着的，人的智能是在不断学习的，而之前学过的知识也会保留。因此，Continual Learning被提出，旨在让模型能够从新任务中学习，完成新的目标的同时，保持其在原有数据...

Posted by Sirin on January 16, 2025

UMich DL for CV

Neural Network

Neural Network Feature Transformation for example, the original space is a Cartesian coordinate system. After mathematic transformation, we can turn it into Polar coordinate system, called featur...

Posted by Sirin on December 12, 2024

GNU Makefile

make and cmake

Makefile 1. Makefile规则 target ... : prerequisites command ... ... target是这个makefile指定的操作所要产生的目标，可以是一个目标文件object file，也可以是一个可执行文件，也可以是一个标签label prerequisites是生成target所需要的依赖文件 command是生...

Posted by Sirin on December 7, 2024

UMich DL for CV

Image Classification

Challenges 细粒度划分 Fine-grained Categories 例如猫的具体种类，缅因猫，布偶猫 背景杂乱 Background Clutter 动物可能具有的保护色，黑猫的深色背景，雪地里的北极狐，落叶堆里的橘猫 光照变换 Illumination Changes 光暗场景的Robustness 物体变形 Deformation...

Posted by Sirin on November 28, 2024

PPoPP'24 RoDe Approach

A Row Decomposition-based Approach for SparseMatrix Multiplication on GPUs

RoDe-based approach for SpMM and SDDMM 概述 本文是针对SpMM和SDDMM两个kernels给出了一种基于行分解的方法，采用的是Compressed Sparse Row格式来存储稀疏矩阵。通过将稀疏矩阵的行分解为Regular part和Residual part来分别进行计算优化，并设计了相应的负载均衡和细粒度Pipelining技术。 相关...

Posted by Sirin on October 31, 2024

FEATURED TAGS
Tools Lecture CV DL
ABOUT ME

Like Tears in The Rain

✉️ None