Ryoma Sato
@joisino_en
Followers
2K
Following
30
Media
15
Statuses
53
Assistant Professor at National Institute of Informatics, Japan. Machine Learning and Data Mining.
NII, Japan
Joined March 2021
My paper "Training-free Graph Neural Networks and the Power of Labels as Features" has been accepted to #TMLR 🎉 I proposed training-free (and optionally trained) GNNs. Paper📜： https://t.co/J6rOQrGejo Code📁： https://t.co/gEzmwu5N48
1
171
1K
🚀Just published: Why can LLMs sometimes reason about things they’ve never seen before? In this article, I explain how attention heads act like “little programs” inside the model—retrieving context, following grammar, and even running algorithms. https://t.co/nuNKCBIhnf
data-processing.club
Share This PostIt is now understood that the attention mechanism in large language models (LLMs) serves multiple functions. By analyzing attention, we gain insight into why LLMs succeed at in-context...
0
3
19
I think machine learning has too much accuracy already. What we really need is less accuracy and more fun. https://t.co/UDQI2huWZe
data-processing.club
Share This PostMost existing machine learning models aim to maximize predictive accuracy, but in this article, I will introduce classifiers that prioritize interestingness. What Does It Mean to...
0
8
168
🚀 Just published: "Deceptive and Lazy AIs: When Language Models Learn to Mislead!" https://t.co/pgTmhDdvNR
data-processing.club
Share This PostAs the capabilities of AI continue to grow, it is becoming increasingly difficult for humans to supervise them. In this article, we discuss this pressing issue through the lens of a...
0
10
46
Our paper "Influential Bandits: Pulling an Arm May Change the Environment" has been accepted to #TMLR 🎉 We proposed and analyzed the influential bandit problem, where an action affects the rewards of other actions. Paper📜： https://t.co/rBRTlb9gJ2
0
3
24
📢NEWS: The Word Tour problem with 40,000 words has been solved optimally — thanks to the collaboration of William Cook and Keld Helsgaun! Details are updated on https://t.co/kXZwFIxeK2 Solving such a large-scale NP-hard problem is truly remarkable!
data-processing.club
Share This PostIn the field of Natural Language Processing (NLP), a central theme has always been “how to make computers understand the meaning of words.” One fundamental technique for this is “Word...
🚀 Just published: "Word Tour: One-dimensional Word Embeddings via the Traveling Salesman Problem!" https://t.co/kXZwFIxMzA
1
5
26
🚀 Just published: "Word Tour: One-dimensional Word Embeddings via the Traveling Salesman Problem!" https://t.co/kXZwFIxMzA
data-processing.club
Share This PostIn the field of Natural Language Processing (NLP), a central theme has always been “how to make computers understand the meaning of words.” One fundamental technique for this is “Word...
1
45
419
🚀 Just published: "Hack Your Feed: Take Control of Your Recommendations!" https://t.co/bQ6a2eBr68
data-processing.club
While recommender systems are convenient, have you ever felt, "I keep getting recommended the same kinds of things?"
0
2
30
🚀 Just published: A Gentle Introduction to Machine Learning Theory! https://t.co/uhKQ9UHZTB
data-processing.club
Share This PostWhen I first began learning about machine learning, I struggled with the notion that lowering the loss on the training data does not necessarily guarantee performance on the test data,...
0
55
558
I just published “The Hidden Information AI Understands but Humans Miss.” Check it out here:
data-processing.club
Share This PostI came across an intriguing post while browsing X during New Year’s holidays. Historically, humanity has pursued the idea that truth is inherently simple, exemplified by equations like...
3
29
204
My paper "Making Translators Privacy-aware on the User’s Side" has been accepted to TMLR🎉 I proposed a method to guarantee privacy on the user's side when using an untrustworthy translator. Paper📜： https://t.co/ZR2SZplcdl
2
29
260
A preprint is out. I proposed training-free graph neural networks. Check it out‼️ https://t.co/J6rOQrGejo
arxiv.org
We propose training-free graph neural networks (TFGNNs), which can be used without training and can also be improved with optional training, for transductive node classification. We first advocate...
4
7
55
Lecture slides on graph neural networks at #MLSS2024 are available ‼️
speakerdeck.com
Lecture on Graph Neural Networks at Machine Learning Summer Seminar (MLSS2024@Okinawa) https://groups.oist.jp/mlss GCN Colab notebook (MIT License): …
0
41
226
Survey slides on speeding up deep learning models are available‼️
speakerdeck.com
I introduce various techniques to speed up deep learning models. Contact: @joisino_en (Twitter) / https://joisino.net/en/
0
7
33
My paper "Graph Neural Networks can Recover the Hidden Features Solely from the Graph Structure" has been accepted to #ICML2023 🎉 I showed GNNs can create new and useful node features even when the input node features are uninformative. Paper📜 https://t.co/mMqXOBNt6j
15
85
497
My paper entitled "Active Learning from the Web" has been accepted to #TheWebConf (WWW) 🎉 I proposed a method for acquiring useful data for model training by regarding the myriad data on the web as a huge pool of active learning. Paper 📜： https://t.co/ommLyOUKKL
2
10
81
(3) Twin Papers: A Simple Framework of Causal Inference for Citations via Coupling #CIKM2022 short We proposed a method to examine the effect of decisions on the number of citations. The idea is that papers that cite each other can be regarded as counterfactual twins.
0
2
9
(2) Towards Principled User-side Recommender Systems #CIKM2022 The information available in a user-side recommendation system is limited. we have investigated theoretically and experimentally how much can be done on the user side. The conclusion is that a great deal can be done.
2
2
15
(1) CLEAR: A Fully User-side Image Search System #CIKM2022 (demo) Paper📜: https://t.co/Ctr7DgDeVU GitHub📂: https://t.co/SZcRh6NIG4 This demo offers similar image searches from Flickr on the user side. You can search for different services and criteria by editing the code.
2
2
6
Three papers (one full, one short, and one demo) have been accepted to #CIKM2022 🎉
2
3
37