It is time for Japan to shine in AI.
RESEARCH
-
Sakana AI Hiring Business Analyst for Tokyo R&D Powerhouse
By
–
We believe that Tokyo will become a Global R&D Powerhouse in AI, and we’re looking to hire a Business Analyst! For more details → https://
sakana.ai/careers/ The ideal candidate will have worked at a Global Investment Bank in Tokyo, as an Analyst or Associate in the Investment -
Unsupervised Text Embedding Space Alignment Research Problem
By
–
As an exercise in open science, gonna tweet the research problem I’m stuck on: i want to align two text embedding spaces in an unsupervised way. The motivation is that in my previous vec2text work, we have to know the embedding model and be able to query it. this is fine in
-
MathVista: Evaluating Math Reasoning in Visual Contexts
By
–
MathVista: Evaluating Math Reasoning in Visual Contexts with GPT-4V, Bard, and Other Large Multimodal Models Lu et al.: https://
arxiv.org/abs/2310.02255 #Artificialintelligence #DeepLearning #MachineLearning -
Random Seed Reset Issue During Model Training Epochs
By
–
every configuration of hyperparameters does this haha. i'm thinking maybe it resets the random seed every epoch? or somehow is shuffling poorly?
-
Praising Excellent AI Reporting by Khari Johnson
By
–
It's well worth rereading these fantastic stories by @kharijohnson
. AI reporting to be proud of. -
Improving Data Shuffling Strategy Across Training Epochs
By
–
can i just shuffle better somehow? I assumed this is what happens when you reuse the same shuffling bt epochs
-
HF Trainer Automatic Shuffling for Model Training
By
–
i’m just using HF trainer! it’s supposed to shuffle for me
-
12 Best Practices for Distilling Smaller Language Models from GPT
By
–
Ready to graduate from #OpenAI to #opensource? Check out our definitive list of 12 best practices for distilling smaller language models from #GPT.
-
Investigating Staircase Loss Curves in HuggingFace Trainer Training
By
–
has anyone seen these staircase loss curves before and really figured out what's going on? my two leading theories are that the loss is getting averaged somewhere behind the scenes (don't know where) and that the data isn't shuffled between epochs. i'm using huggingface trainer.