This document summarizes recent research on applying self-attention mechanisms from Transformers to domains other than language, such as computer vision. It discusses models that use self-attention for images, including ViT, DeiT, and T2T, which apply Transformers to divided image patches. It also covers more general attention modules like the Perceiver that aims to be domain-agnostic. Finally, it discusses work on transferring pretrained language Transformers to other modalities through frozen weights, showing they can function as universal computation engines.
PFN福田圭祐による東大大学院「融合情報学特別講義Ⅲ」(2022年10月19日)の講義資料です。
・Introduction to Preferred Networks
・Our developments to date
・Our research & platform
・Simulation ✕ AI
7月29日開催 July Tech Festa 2018基調講演スライドです。
大村伸吾「Preferred Networksの機械学習クラスタを支える技術」
https://2018.techfesta.jp/
Slides of Keynote in July Tech Festa 2018.
PFN福田圭祐による東大大学院「融合情報学特別講義Ⅲ」(2022年10月19日)の講義資料です。
・Introduction to Preferred Networks
・Our developments to date
・Our research & platform
・Simulation ✕ AI
7月29日開催 July Tech Festa 2018基調講演スライドです。
大村伸吾「Preferred Networksの機械学習クラスタを支える技術」
https://2018.techfesta.jp/
Slides of Keynote in July Tech Festa 2018.
DNN computes many multiply-accumulate operations, which generally use a GPU. I tried using M5StickV as a general-purpose DNN accelerator. / DNN(Deep Neural Network)を用いたアプリケーションは、その演算に大量の計算リソース(積和演算)が求められる。その為、一般的にGPUを用いる。今回、M5StickVをDNNのアクセラレータとして使えないか?試みた。
高性能データ処理プラットフォーム (Talk on July Tech Festa 2015)Yu Liu
Introduction to a Scalable, High-Performance Distributed Data Processing Platform (WorksApplications)
http://2016.techfesta.jp/
PFN Summer Internship 2021 / Kohei Shinohara: Charge Transfer Modeling in Neu...Preferred Networks
The document discusses techniques for modeling charge transfer in neural network potentials (NNPs) for materials simulation. It presents a graph neural network (GNN) baseline architecture called NequIP that predicts short-range atomic energies. Additional techniques are explored to model long-range electrostatic interactions, including adding an electrostatic correction term (Eele) using Ewald summation and using charge equilibration (Qeq) to predict atomic charges. Results show that while Qeq improves charge prediction accuracy, the baseline GNN achieves comparable or better overall accuracy in most datasets tested, possibly because the GNN can already learn electrostatic effects. The document also discusses PyTorch implementations of Ewald summation and Qeq for efficient evaluation.
5. 豊富な計算資源と高度な技術を基盤に複数の事業を創出
PFNを支える技術と事業内容
Computer Vision(コンピュータビジョン) Data Analytics(データ解析)
Navigation(ナビゲーション)
Visual Inspection(外観検査)
Pose(ポーズ推定)
Scene(シーン解析)
Image Segmentation
Anomaly Detection(異常検知)
Optimization(最適化)
Time series data(時系列データ)
Infrastructure (インフラ技術)
Machine Learning and Deep Learning(機械学習と深層学習)
Manufacturing Transportation Bio & Healthcare
Personal Robot Visual Inspection Entertainment
PFN
Technology
Business
Object Detection(物体検出)
7. 深層学習の要求計算量の増大
最先端の研究においてモデルサイズは増加
する傾向
● 画像→動画/立体、画像の高精細化
● 言語処理モデルの大規模化
● Foundation Model
NAS (Neural Architecture Search)
● アーキテクチャ探索の自動化
● 人ではなく、計算能力がボトルネック
→ 計算力の強化 = 競争力の強化
-> PFNが計算能力に対して投資する理由
The graph was excerptedfrom Shaden Smith et al. (2022)
Using DeepSpeed and Megatron to Train Megatron-TuringNLG 530B, A Large-Scale Generative Language Model