paper_read

本文最后更新于:2024年2月11日 晚上

文章阅读

NOTE:数学公式堂堂支持

NOTE: 调试数学公式恼羞成怒调整至notion中完成

paper_read

学习代码

TODO list

nihao fdasfa

朗之万采样:从先验分布中随机采样一个初始样本,然后利用score逐渐将样本像数据分布的高概率密度区域靠近,同时为了生成结果的多样性,需要采用过程带有随机性。同时z服从正态分布,epsilon为每次移动的步长

Denoising Score Matching

固有: Missing argument for \mathbf q_\sigma(\tilde{x}) = \int q_\sigma(\tilde{x}|x)p_{data}(x)\mathbf

rec && llm

Prompting Large Language Models for Recommender Systems: A Comprehensive Framework and Empirical Analysis

some finding

  • Fine-tuning all parameters of LLMs for recommendation is more effective than parameter-efficient fine-tuning, but more training time is required
  • Increasing the number of historical items to represent users brings insignificant gains for LLM

problem

  • Existing literature has also found the position bias in LLM-based recommendations [32, 74, 133], and methods such as bootstrapping [32] and Bayesian probability framework [74] havebeen proposed to calibrate unstable results. However, the instability of LLMs is still an unresolved issue.

Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation

idea

recinterpret + llara

motivation

integrate language and collaborative semantics for recommender systems Our approach can directly generate items from the entire item set for recommendation, without relying on candidate items.

method

  1. item indexing
    1. These item indices are learned based on the text embeddings of items encoded by LLMs, enabling the learned IDs to capture the intrinsic similarity among items
    2. 按照语义相关度,分割成树状的id
  2. finetune
    1. 多种推荐任务的文本描述finetune

LlamaRec: Two-Stage Recommendation using Large Language Models for Ranking

idea

升级版本tallrec,tallrec的候选list的延伸,通过logits+softmax的形式来进行重新排序,但是训练的方式同alpaca,但是用了pytorch-lighting的形式。

motivation 召回 排序

  1. 序列推荐召回list
  2. 设计了 verbalizer approach,用计算logits的方式作为打分依据,解决了推理时延问题
  3. 分词的方式是item逐一处理再进行拼接。

novelty

method

llm

Toolformer: Language Models Can Teach Themselves to Use Tools

idea

无监督的方式教会语言模型使用工具

motivation

相当于多遍扫描的过程,第一遍text插入在哪使用工具,第二遍通过替换工具的内容,通过loss函数的计算决定是否替换,然后用finetune的手段进行生成。

method

  1. 添加在哪里使用工具
  2. 调用工具,同时计算loss,设置阀值决定是否需要替换。

vision && LLM

VAE

VAE采样

CLIP

CLIP

idea

our intuition is that CLIP models will need to learn to recognize a wide variety of visual concepts in images and associate them with their names

method

通过正负样本对进行训练,样本对通过编码器,然后同时对batch中多对样本进行点积处理,然后可以分出正负的样本对

clip-contrast

diffusion

把picture作为视频的id,然后用文本进行生成,然后用生成的picture作为检索的策略?

  1. 文字LLM增大,encoder 成对资料更好,但是diffusion model扩大效果不好
  2. FID 评估指标 small is better FID

NLP-diffusion

通过token转换成embedding上面添加noise,再进行还原 [ ] to be done

NLP
NLP2

stable diffusion

decode

decode

i am blueheloo i am red

hel

ablity

application

  1. adaptive attendsion span
  2. class-free guidance

week

  1. transformer相关模型与blog
  2. 探究多模态推荐,主要学习diffusion模型,并复原了diffusion(DDPM)的代码。然后阅读的论文有clip(文本段和图片的对比学习),dalle2(unclip模型)。这两篇工作都强调了增大数据规模后,性能获得了大幅提升,(不确定diffusion+rec是否可行局限于rec数据集和CV数据集的差异,以及rec的数据量是离散的,之后计划阅读两者结合后的论文diffurec,以及打算阅读NLP上的diffusion应用和class-free guidance论文)


paper_read
http://example.com/2024/01/25/paper-read/
作者
NGC6302
发布于
2024年1月25日
许可协议