Newly Blog

Credential ticket for principles without need to type in password, from MIT Kerberos.
A ticket has ticket lifetime and renewable lifetime.
Ticket lifetime is shorter than renewable lifetime.
For liniu@ANT.AMAZON.COM, the default ticket lifetime is 10h (resp., 6h40m) when using kinit -l (resp., -r), why?

From KDC server side:

1. modify the max_life in /etc/krb5kdc/kdc.conf and restart the KDC daemon /var/kerberos/krb5kdc/kdc.conf
2. Via “kadmin”, changed the “maxlife” for a test principal via “modprinc -maxlife 14hours “

From Kerberos client side:
modify in /etc/krb5.conf

In fact, the ticket lifetime is the minimum of the following values:

• max_life in kdc.conf on the KDC servers.

• ticket_lifetime in krb5.conf on the client machine.

• maxlife for the user principal.

• maxlife for the service principal “krbtgt/[REALM_in_CAPS]”

• requested lifetime in the ticket request. For example: kinit -l 14h

• maxlife for the AFS service principal “afs/[realm_in_lower_case]”, if you want to increase the lifetime of your AFS token.

commonly used commands:

1
2

klist
kdestroy

An example:

Ticket cache: FILE:/tmp/krb5cc_4126574_GM19Ct
Default principal: liniu@ANT.AMAZON.COM

Valid starting Expires Service principal
08/18/16 08:13:20 08/18/16 18:13:20 krbtgt/ANT.AMAZON.COM@ANT.AMAZON.COM

time format is like 4d5h30m

1
2
3
4
5
6

kinit -l lifetime //request a ticket with ticket lifetime of lifetime
-r renewable-life //request renewable ticket with a total lifetime of renewable-life
//I'm still unclear about the difference between -l and -r
-f //forwardable
-F //non-forwardable
-R //requests renewal of the ticket-granting ticket. No need for password but must be within ticket lifetime instead of renewable lifetime.

Automatically renew tickets: Since you need to renew a ticket before its ticket lifetime expires, the easiest way to renew tickets is to put it in a cron job since renewing a ticket is non-interactive.

Run ‘crontab -e’ to edit the file in /var/spool/cron/. Use ‘crontab -l’ to see the file.

1
2
3
4
5
6
7
8
9

# Renew the kerberos ticket every 8 hours, this will extend the lifetime of 
# the ticket until the renew lifetime expiers, after that this command will 
# fail to renew the ticket and you will need to interactively 
# run `kinit -f -l 86400 -r 2592000`
#
# minute  hour  day_of_month  month  weekday  command
59 00,08,16 * * * /usr/kerberos/bin/kinit -R
//59 minute, 0 or 8 or 16 o'clock, any, any, any, the command to be executed is '/usr/kerberos/bin/kinit -R'
//some short notations: 1-3 means 1,2,3; */15 means every 15 

Key Notes:

1. do not use sudo kinit

2. when no credential ticket can be found, add -c $KRB5CCNAME, where KRB5CCNAME is the environment variable recording the path of credential ticket.

1. Preinstallation

   • install Node.js https://nodejs.org/en/download/
   • install git https://git-scm.com/download/win
   • install hexo: right click “git bash here”, $npm install hexo-cli -g

2. SSH keys

   • check whether ssh exists $cd ~/.ssh. If not, $ssh-keygen -t rsa -C "your_email@example.com" to generate key file
   • get the SSH key $cat ~/.ssh/id_rsa.pub
   • create the key in github: account setting->SSH
   • check the SSH key $ ssh -T git@github.com

3. create an empty folder as Hexo folder

   1 2 3

   $hexo init $hexo generate $hexo server

   local test: http://localhost:4000/

4. deploy local Hexo folder

   • in Hexo folder, modify the _config.yml file as follows,

     1 2 3 4

     deploy: type: git repository: $(SSH address from github) branch: master

   • in Hexo folder, install the deployer by $npm install hexo-deployer-git --save
   • type $hexo g and $hexo d, or $hexo d -g.

5. change theme: go to the folder “/themes” and git clone https://github.com/iissnan/hexo-theme-next. config the theme in the file “/themes/XXXX/_config.yml”.

6. local search

   • $npm install hexo-generator-searchdb --save

   • in the site _config.yml file, set local_search: enable: true. paste the following lines anywhere.

     1 2 3 4 5

     search: path: search.xml field: post #post, page, all format: html limit: 10000

7. add new page for some subcategory (e.g., write): in the theme _config.yml file

   1 2	menu: write: /categories/write

8. add social links: in the theme _config.yml file

   1 2 3 4

   # Social links social: GitHub: https://github.com/ustcnewly Linkedin: https://www.linkedin.com/in/li-niu-b0905133/

9. latex:

   • install $npm install hexo-math --save and restart Hexo

   • in the theme _config.yml file, modify as follows (notice that cdn may change)

     1 2 3 4 5

     # MathJax Support mathjax: enable: true per_page: false cdn: //cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.1/MathJax.js?config=TeX-AMS-MML_HTMLorMML

   • to address the conflict between MathJax and Hexo

     1 2	$npm uninstall hexo-renderer-marked --save $npm install hexo-renderer-kramed --save

10. push posts to the top

    • $ npm install hexo-generator-index-pin-top --save.
    • for the target post, add top: true in Front-matter, or top: 10 with larger number indicating higher priority.

11. insert code block

    1 2 3

    {% codeblock lang:python %} code snippet {% endcodeblock %}

12. insert image

    1	<img src="http://bcmi.sjtu.edu.cn/~niuli/github_images/t0IXoZq.jpg" width="50%" height="50%">

Tips:

1. If something is wrong with the representation and hard to tune, you can try deleting extra spaces or adopting an alternative format (e.g., two code block formats).

2. Case sensitive: sometimes you switch between capital letter and small letter, which may lead to 404 not found errors.

   • set ignorecase as false in the file .deploy_git/.git/config
   • clean the folder .deploy_git
   • hexo clean and hexo d -g

• interaction and active reasoning: [2] (Sokoban, Maze, and Taxi)

• performs both policy and world model learning: [3]

Survey: [1]

Reference

[1] Zhao, Changyuan, et al. “Edge general intelligence through world models and agentic AI: Fundamentals, solutions, and challenges.” arXiv preprint arXiv:2508.09561 (2025).
[2] Shu, Bao, et al. “Thinking by Doing: Building Efficient World Model Reasoning in LLMs via Multi-turn Interaction.” arXiv preprint arXiv:2511.23476 (2025).
[3] Yu, Xiao, et al. “Dyna-Think: Synergizing Reasoning, Acting, and World Model Simulation in AI Agents.” arXiv preprint arXiv:2506.00320 (2025).

Survey: [1]

Reference

[1] Tong, Jingqi, et al. “Thinking with video: Video generation as a promising multimodal reasoning paradigm.” arXiv preprint arXiv:2511.04570 (2025).

Survey: [1]

Reference

[1] Su, Zhaochen, et al. “Thinking with images for multimodal reasoning: Foundations, methods, and future frontiers.” arXiv preprint arXiv:2506.23918 (2025).

We use the Dawid-Skene vote aggregation algorithm to obtain the ground truth label for each snippet, since this is often considered ‘gold standard’ for aggregation in practice. DawidSkene is an unsupervised inference algorithm that gives the Maximum Likelihood Estimate of observer error rates using the EM algorithm.

1) Using the labels given by multiple annotators, estimate the most likely “correct” label for each video snippet.

2) Based on the estimated correct answer for each object, compute the error rates for each annotator.

3) Taking into consideration the error rates for each annotator, recompute the most likely “correct” label for each object.

4) Repeat steps 2 and 3 until one of the termination criteria is met (error rates are below a pre-specified threshold or a pre-specified number of iterations are completed).

Vector Quantization: VQVAE [1],VQVAE2 [2], VQGAN [6].

Residual Quantization: RQVAE [3]

Accelerate auto-regression: [4] [5]

Hierarchical residual quantization: VAR [7]

References

[1] Oord, Aaron van den, Oriol Vinyals, and Koray Kavukcuoglu. “Neural discrete representation learning.” arXiv preprint arXiv:1711.00937 (2017).

[2] Razavi, Ali, Aaron van den Oord, and Oriol Vinyals. “Generating diverse high-fidelity images with vq-vae-2.” Advances in neural information processing systems. 2019.

[3] Lee, Doyup, et al. “Autoregressive Image Generation using Residual Quantization.” arXiv preprint arXiv:2203.01941 (2022).

[4] Bond-Taylor, Sam, et al. “Unleashing Transformers: Parallel Token Prediction with Discrete Absorbing Diffusion for Fast High-Resolution Image Generation from Vector-Quantized Codes.” arXiv preprint arXiv:2111.12701 (2021).

[5] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, William T. Freeman, “MaskGIT: Masked Generative Image Transformer”, arXiv preprint arXiv:2202.04200.

[6] Patrick Esser, Robin Rombach, Björn Ommer, “Taming Transformers for High-Resolution Image Synthesis”.

[7] Tian, Keyu, et al. “Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction.” arXiv preprint arXiv:2404.02905 (2024).

1. learn attribute vector based on the relation and difference between different categories (each dimension if uninterpretable): [1] (Laplacian matrix), [2] (triplet loss)

2. exploit local information and encode them into attribute vector (each dimension is interpretable): [3] (discriminative cluster, doublets), [4] (joint attribute learning and feature learning)

3. learn attention map for each latent attribute [5]

Reference

1. Yu, Felix X., et al. “Designing category-level attributes for discriminative visual recognition.” CVPR, 2013.

2. Li, Yan, et al. “Discriminative learning of latent features for zero-shot recognition.” CVPR, 2018.

3. Singh, Saurabh, Abhinav Gupta, and Alexei A. Efros. “Unsupervised discovery of mid-level discriminative patches.” ECCV, 2012.

4. Huang, Chen, Chen Change Loy, and Xiaoou Tang. “Unsupervised learning of discriminative attributes and visual representations.” CVPR, 2016.

5. Yang, Wenjie, et al. “Towards rich feature discovery with class activation maps augmentation for person re-identification.” CVPR, 2019.

Let us use $S$ to denote the set of training categories and $T$ to denote the set of testing categories.

• $S=T$: the most common case
• $S\cap T=\emptyset$: zero-shot learning
• $S\subset T$: generalized zero-shot learning
• $S\supset T$: pretrained model

1. Blend text and background images.

   • text image (font, color, border, blending): [1]
   • scene-text image (font, color, border, blending, geometry): [2] [3]

Reference:

1. Jaderberg, Max, et al. “Synthetic data and artificial neural networks for natural scene text recognition.” arXiv preprint arXiv:1406.2227 (2014).

2. Gupta, Ankush, Andrea Vedaldi, and Andrew Zisserman. “Synthetic data for text localisation in natural images.” CVPR, 2016.

3. Zhan, Fangneng, Shijian Lu, and Chuhui Xue. “Verisimilar image synthesis for accurate detection and recognition of texts in scenes.” ECCV, 2018.