Data Scientology
前往频道在 Telegram
Hot data science related posts every hour. Chat: https://telegram.me/r_channels Contacts: @lgyanf
显示更多1 138
订阅者
无数据24 小时
无数据7 天
-730 天
正在加载数据
标签云
进出提及
---
---
---
---
---
---
吸引订阅者
十一月 '2611月 '26
十一月 '26
+1
在2个频道中
十月 '26
+11
在6个频道中
Get PRO
九月 '26
+11
在0个频道中
Get PRO
八月 '26
+11
在0个频道中
Get PRO
七月 '26
+13
在0个频道中
Get PRO
六月 '26
+6
在0个频道中
Get PRO
五月 '26
+4
在0个频道中
Get PRO
四月 '26
+30
在0个频道中
Get PRO
三月 '26
+40
在0个频道中
Get PRO
二月 '26
+36
在0个频道中
Get PRO
一月 '26
+13
在1个频道中
Get PRO
十二月 '25
+10
在0个频道中
Get PRO
十一月 '25
+7
在0个频道中
Get PRO
十月 '25
+4
在0个频道中
Get PRO
九月 '25
+3
在0个频道中
Get PRO
八月 '25
+8
在0个频道中
Get PRO
七月 '25
+10
在0个频道中
Get PRO
六月 '25
+10
在0个频道中
Get PRO
五月 '25
+9
在0个频道中
Get PRO
四月 '25
+3
在0个频道中
Get PRO
三月 '25
+7
在0个频道中
Get PRO
二月 '25
+3
在0个频道中
Get PRO
一月 '25
+4
在0个频道中
Get PRO
十二月 '24
+6
在0个频道中
Get PRO
十一月 '24
+2
在0个频道中
Get PRO
十月 '24
+8
在0个频道中
Get PRO
九月 '24
+8
在0个频道中
Get PRO
八月 '24
+11
在0个频道中
Get PRO
七月 '24
+13
在0个频道中
Get PRO
六月 '24
+11
在0个频道中
Get PRO
五月 '24
+10
在0个频道中
Get PRO
四月 '24
+12
在0个频道中
Get PRO
三月 '24
+20
在0个频道中
Get PRO
二月 '24
+19
在0个频道中
Get PRO
一月 '24
+19
在0个频道中
Get PRO
十二月 '23
+17
在6个频道中
Get PRO
十一月 '23
+23
在3个频道中
Get PRO
十月 '23
+27
在3个频道中
Get PRO
九月 '23
+21
在0个频道中
Get PRO
八月 '23
+9
在0个频道中
Get PRO
七月 '23
+18
在0个频道中
Get PRO
六月 '23
+20
在0个频道中
Get PRO
五月 '23
+26
在0个频道中
Get PRO
四月 '23
+53
在0个频道中
Get PRO
三月 '23
+15
在0个频道中
Get PRO
二月 '23
+12
在0个频道中
Get PRO
一月 '23
+18
在0个频道中
Get PRO
十二月 '22
+31
在0个频道中
Get PRO
十一月 '22
+14
在0个频道中
Get PRO
十月 '22
+47
在0个频道中
Get PRO
九月 '22
+45
在0个频道中
Get PRO
八月 '22
+18
在0个频道中
Get PRO
七月 '22
+18
在0个频道中
Get PRO
六月 '22
+18
在0个频道中
Get PRO
五月 '22
+27
在0个频道中
Get PRO
四月 '22
+43
在0个频道中
Get PRO
三月 '22
+47
在0个频道中
Get PRO
二月 '22
+89
在0个频道中
Get PRO
一月 '22
+57
在0个频道中
Get PRO
十二月 '21
+36
在0个频道中
Get PRO
十一月 '21
+27
在0个频道中
Get PRO
十月 '21
+43
在0个频道中
Get PRO
九月 '21
+64
在0个频道中
Get PRO
八月 '21
+55
在0个频道中
Get PRO
七月 '21
+28
在0个频道中
Get PRO
六月 '21
+26
在0个频道中
Get PRO
五月 '21
+29
在0个频道中
Get PRO
四月 '21
+26
在0个频道中
Get PRO
三月 '21
+49
在0个频道中
Get PRO
二月 '21
+21
在0个频道中
Get PRO
一月 '21
+45
在0个频道中
Get PRO
十二月 '20
+882
在0个频道中
| 日期 | 订阅者增长 | 提及 | 频道 | |
| 03 十一月 | 0 | |||
| 02 十一月 | 0 | |||
| 01 十一月 | +1 |
频道帖子
| 2 | I trained YOLO26 on a 50,000 image AI-generated dataset and tested it on real COCO photos, it reached 59% of a real-photo model
https://redd.it/1xl0vv8
@datascientology | 31 |
| 3 | Freind of mine published a paper with fabricated results in a top journal and got accepted and i feel disgusted d
A friend of mine was doing gap analysis as a project , the result he got werent that big from the original papers, so he simply fabricated and increaaed the metrics results here and there, AND IT GOT ACCEPTED IN AAAI.
HOW DO THEY CHECK THE PAPERS DUDE, i was in a group with another two and worked soooo hard in an optimization paper, and was sooo happy to get accpetwd in acl, but seeing and knowing that someone fabricated and got to a better conference made me soooooo mad.
I wouldve expected something like aaai would check and know if a data is fabricaged somwhow, didnt expect this from a top conference like that
And then this type of people qill go and do their phd in one of the top unis because if this, i wouldve expected the system to be mkre advanced and professional in the academic field, specifically for somethinf like aaai
https://redd.it/1xj8tvc
@datascientology | 51 |
| 4 | Working through the Transformer forward pass by hand, one matrix at a time, is what finally made it click for me
When I read *Attention Is All You Need*, the equations made sense line by line, but I couldn't picture the whole forward pass. What helped was writing out every matrix for a tiny example, noting its shape at each step, and doing the multiplications myself.
A few things only clicked once I did it that way:
* **Positional encoding.** Sine and cosine at different frequencies give every position its own signature, and the model doesn't have to learn one vector per position. Working out a 6 × 512 table by hand made it obvious why the authors hoped it would extend to sequences longer than any seen in training.
* **Shapes barely change.** In the encoder, the matrix goes in as (tokens × d\_model) and comes out the same size after every layer. All the interesting work happens inside that fixed shape.
* **BERT is mostly the same picture.** Once the encoder was clear, BERT turned out to be a small edit to it: a packed sentence pair, segment embeddings, and two small heads on top for the masked word and the next sentence.
I wrote my notes up as step-by-step pages with practice exercises beside each step and a notebook that checks your answers, in case it's useful to anyone learning the same way: [https://guptaaditya95.github.io/llmbreakdown/](https://guptaaditya95.github.io/llmbreakdown/)
I'm curious how others built their intuition for these models. Did working the math by hand help you, or did something else make it click? And which architecture was hardest to picture?
https://redd.it/1xh9j50
@datascientology | 70 |
| 5 |  A free deep learning textbook + 68 Python notebooks, from neural network basics to transformers and diffusion models
https://redd.it/1xf8t1i
@datascientology | 65 |
| 6 |  +8没有文字... | 96 |
| 7 | Artistic application of monocular depth estimation – Vathography
https://redd.it/1xcs1nh
@datascientology | 82 |
| 8 |  Fully open source stereo depth stack with cheap cameras
https://redd.it/1xb47lf
@datascientology | 82 |
| 9 | Conference papers are cooked! D
Got three papers to review for ICLR and five for AAMAS, and believe me the quality of these papers is insanely bad.
I would even refrain from accepting them to a workshop, let alone be main venues.
A sincere request; Please don't submit the paper unless it is ready and if you are a first timer, please please get a feedback from a mentor out. Otherwise you are just wasting the time of so many reviewers out there
https://redd.it/1x9fyhd
@datascientology | 93 |
| 10 |  Detecting sewer pipe defects in CCTV inspection video: anyone working on something similar?
https://redd.it/1x6tzsq
@datascientology | 101 |
| 11 | PhD students, how ya feeling? D
All seems a bit pointless now. Either pump out AI papers and neglect your own development, or take it slow and fall behind in the publishing race. I already have two papers made largely without LLMs (still seeking publication), but it's hard to resist just vibing the third from start to finish with Claude.
https://redd.it/1x5fv1a
@datascientology | 111 |
| 12 |  Is this Gaussian Splatting, GenAI or some other technique?
https://redd.it/1x3tvpw
@datascientology | 105 |
| 13 | Nvidia’s erroneous paper accepted as ICML’s spotlight D
Here is the story:
Nvidia has published this work (with source code available) called dreamDojo which is a world model for robotics based of their prior work Cosmos 2.5 which is cited about 100 times and got ICML’s spotlight.
Authors are very well known and respected in the field with too many peer reviewed papers already published.
The work doesn’t have much novelty (which I don’t care) but it is yet another foundation model. The gist is that they collected about 44k hours of human data (data are NOT open sourced, which I don’t care) and used that for pre-training of DreamDojo initialized from Cosmos 2.5 .
Table 4 of the paper though shows a very marginal improvement over cosmos 2.5. Believe or not, just about 0.5 db PSNR improvement. Its fishy, is not it? You use 44000 hours of human data and few hundred hours of other kind of data and robot data and 256 H100 gpu and you got just a very marginal improvement.
Anyway I went to give it a try, started post training on their GR1 released data, and could produce their results. Meanwhile a colleague with the help of Claude found a bug in their post training code, which essentially make all the post training code wrong.
Then we checked their github issues, and noticed that two other bugs are reported which effect the whole pre-training! Especially this bug . So essentially based on the code their released everything in pre training phase, post training phase and evaluation is buggy and is wrong.
The code doesn’t seem AI written cause its (very) bad written. And now after seeing all those bugs the results in paper makes sense.
My question is:
WTF!
How it did not ring a bell for the authors. Did not they ask we spend a hell lot of resources and huge amount of data and essentially got nothing?
Did not it ring a bell for reviewers?
Openreview
https://redd.it/1x0i6b5
@datascientology | 128 |
| 14 | Saw this on Rednote, WTF D
https://preview.redd.it/t7wxguxqvzth1.png?width=1314&format=png&auto=webp&s=0ba382fb88894c6e687d5bb9a4b14a85334eae67
Guess we have a dataset for sycophancy and AI content detection
https://redd.it/1wzpzh3
@datascientology | 127 |
| 15 | Distilling Stockfish on a Billion Positions, Full 3.9B Dataset Available [P]
https://blog.lukesalamone.com/posts/distilling-stockfish
https://redd.it/1wxz5qq
@datascientology | 121 |
| 16 | arXiv now limits submitters to up to two submissions per calendar month [N]
https://blog.arxiv.org/2026/10/01/updated-rate-limit-policy/
https://redd.it/1wvg7yc
@datascientology | 130 |
| 17 | LessThink-Qwen3-4B: the same model, with far less thinking P
I post-trained Qwen3-4B to spend 44% fewer tokens on reasoning, keeping its knowledge and answer style. The whole pipeline ran on one GPU.
folks, you can check it out on : https://5ivatej.com/lessthink/
https://redd.it/1wtygav
@datascientology | 132 |
| 18 | Free, open-source AI engineering course where you build each algorithm by hand: 523 lessons, now as EPUB/PDF books P
AI Engineering from Scratch is an MIT-licensed curriculum: 523 lessons across 20 phases, from linear algebra and backprop to transformers, LLMs, agents, and production serving. The code is stdlib-first, so you see every step instead of calling a library.
This month's edition:
\- six EPUB and PDF volumes built from the lessons, attached to the release
\- the site interface and lessons in eight languages (Chinese, Hindi, Spanish, Arabic, French, Portuguese, Turkish, Vietnamese)
\- CI now runs each lesson's own tests, and a sweep fixed datasets, models, and links that had stopped working
Site: https://aiengineeringfromscratch.com
Books and release notes: https://github.com/rohitg00/ai-engineering-from-scratch/releases/tag/v2026.10
If you use a coding agent, npx skills add rohitg00/ai-engineering-from-scratch and then /start-learning gives you a placement quiz and a study plan.
https://redd.it/1ws6e9p
@datascientology | 145 |
| 19 | I made a computer vision tool for rock climbing analysis in 3D using iPhone LiDAR!
https://redd.it/1wphdyq
@datascientology | 158 |
| 20 |  Penalty kick analysis running fully in the browser at ~30 fps with WebGPU
https://redd.it/1wnafi6
@datascientology | 192 |
