Deep Learning on iOS #360iDev

This document discusses Rhebok, a high performance Rack handler written in Ruby. Rhebok uses a prefork architecture for concurrency and achieves 1.5-2x better performance than Unicorn. It implements efficient network I/O using techniques like IO timeouts, TCP_NODELAY, and writev(). Rhebok also uses the ultra-fast PicoHTTPParser for HTTP request parsing. The document provides an overview of Rhebok, benchmarks showing its performance, and details on its internals and architecture.

Slide for Shibuya.pm Tech Talk #17 LT

Mackerel & Norikra mackerel meetup #4 LT

Norikraで作るPHPの例外検知システム YAPC::Asia Tokyo 2015 LT

ISUCONの勝ち方 YAPC::Asia Tokyo 2015

メルカリのデータベース戦略 ／ PHPとMySQLの怖い話 MyNA会2015年8月

slide for norikra meetup #2

This document discusses strategies for optimizing access to large "master data" files in PHP applications. It describes converting master data files from PHP arrays to tab-separated value (TSV) files to reduce loading time. Benchmark tests show the TSV format reduces file size by over 50% and loading time from 70 milliseconds to 7 milliseconds without OPcache. Accessing rows as arrays by splitting on tabs is 3 times slower but still very fast at over 350,000 gets per second. The TSV optimization has been used successfully in production applications.

This document summarizes the 2nd place solution to an Instacart market basket analysis competition. The approach involved feature engineering using user, item, user-item interaction, and datetime features. Feature importance analysis identified key predictive features. Important findings provided insights like frequent reorders for fruits and a user's previous order predicting their next order. The solution maximized the F1 evaluation metric by simulating predictions and thresholds to optimize recall and precision.

第7回Security-JAWSにて発表させていただいた負荷試験の失敗事例の紹介

LINE Ads Platform の CTR を2倍にした開発手法 /小川 拡 (LINE株式会社 サービス開発1室) LINE Developer Meetup in Tokyo #22 -Ads Platform-の登壇資料です https://line.connpass.com/event/69277/

The feature we always hear about whenever Java 9 is in the news is Jigsaw, modularity. But this doesn't scratch the same developer itch that Java 8's lambdas and streams did, and we're left with a vague sensation that the next version might not be that interesting. Java 9 actually has a lot of great additions and changes to make development a bit nicer. These features can't be lumped under an umbrella term like Java 8's lambdas and streams, the changes are scattered throughout the APIs and language features that we regularly use. In this presentation Trisha will show, via live coding: - What the Java Platform Module System is and how to make your code modular - How we can use the new Flow API to utilise Reactive Programming - The improvements to the Streams API that make it easier to control infinite streams - How to the Collections convenience methods simplify code Along the way we'll bump into other Java 9 features, including some of the additions to interfaces and changes to deprecation.

2017/6/17に行われたDeNA主催のICLR2017読み会でお話させて頂きました。 https://connpass.com/event/57631/?utm_campaign=event_reminder&utm_source=notifications&utm_medium=email&utm_content=detail_btn

This document describes research on semi-supervised learning on graph-structured data using graph convolutional networks. It proposes a layer-wise propagation model for graph convolutions that is more efficient than previous methods. The model is tested on several datasets, achieving state-of-the-art results for semi-supervised node classification while training faster than alternative methods. Future work to address limitations regarding memory requirements, directed graphs, and locality assumptions is also discussed.

ICLR2017読み会 @ DeNA ICLR2017紹介

"Tying Word Vectors and Word Classifiers: A Loss Framework for Language Modeling"の解説

ICLR2017読み会 Data Noising as Smoothing in Neural Network Language Models @Dena

This document summarizes a presentation on deep image processing and computer vision. It introduces common deep learning techniques like CNNs, autoencoders, variational autoencoders and generative adversarial networks. It then discusses applications including image classification using models like LeNet, AlexNet and VGG. It also covers face detection, segmentation, object detection algorithms like R-CNN, Fast R-CNN and Faster R-CNN. Additional topics include document automation using character recognition and graphical element analysis, as well as identity recognition using face detection. Real-world examples are provided for document processing, handwritten letter recognition and event pass verification.

The document summarizes an OpenCV based image processing attendance system. It discusses using OpenCV to detect faces in images and recognize faces by comparing features to a database. The key steps are face detection using Viola-Jones detection, face recognition using eigenfaces generated by principal component analysis to project faces into "face space", and measuring similarity by distance between projections.

A practical talk by Anirudh Koul aimed at how to run Deep Neural Networks to run on memory and energy constrained devices like smartphones. Highlights some frameworks and best practices.

The document provides an overview of machine learning use cases. It begins with an agenda that will discuss the basic framework for ML projects, model deployment options, and various ML use cases like text classification, image classification, object detection, etc. It then covers the basic 5 step framework for ML projects - defining the problem, planning the solution, acquiring and preparing data, designing and training a model, and deploying the solution. Next, it discusses popular methods for various tasks like image classification, object detection, pose estimation. Finally, it shares several use cases for each task to demonstrate real-world applications.

Apple makes it really easy to get started with Machine Learning as a developer. See how you can easily use Create ML and Turi Create to train Machine Learning models and use them in your iOS apps.

In this guide, we will explore how to perform face detection in Python using popular libraries and tools.

1. The document discusses optical character recognition (OCR), including its applications, how it works, and the platform used. 2. OCR involves using software to convert scanned images of text into machine-encoded text by recognizing glyphs and classifying characters through feature extraction and neural networks. 3. The authors explore using OCR for tasks like digitization and security monitoring to reduce human error, and discuss future enhancements like recognizing multiple characters and improving accuracy.

A practical talk by Anirudh Koul aimed at how to run Deep Neural Networks to run on memory and energy constrained devices like smart phones. Highlights some frameworks and best practices.

This document describes an automatic attendance system using face recognition. It discusses the following: 1. The system was created by a team of 4 students to automate attendance taking and save faculty time. It uses OpenCV and face detection algorithms. 2. The system works by training on images of students taken from the class and stored in the database. It then detects faces in new images or video and identifies the students to mark attendance. 3. Key technologies used include OpenCV for image processing, Tkinter for the GUI, Pandas and NumPy for data handling, and algorithms like Haar Cascade and LBPH for face detection and recognition.

This document provides an overview of computer vision and OpenCV. It defines computer vision as using algorithms to identify patterns in image data. It describes how images are represented digitally as arrays of pixels and how features like edges and corners are important concepts. It introduces OpenCV as an open source library for computer vision with over 2500 algorithms. It supports languages like C++ and Python. OpenCV has modules for tasks like image processing, video analysis, and object detection. The document provides details on OpenCV data structures like Mat and how to get started with OpenCV in Android Studio by importing the module and adding the native libraries.

Slides for the AI community meetup organized by Deltatre Innovation Lab, in Turin, November 19th (OGR Tech, Talent Garden).

This document provides an overview of deep learning on mobile. It discusses why deep learning is important for mobile, how to build and run models on mobile, factors like hardware that impact performance, benchmarking models, example applications, techniques for increasing model efficiency like quantization and pruning, and federated learning. The document is a guide for practitioners to develop deep learning applications for mobile.

Currently, Siddha is an Architect at Nvidia focusing on the Self-Driving initiative. She works towards stable and scalable training of neural networks on very large data centers, and utilizes simulation to validate the neural networks. In 2017 Siddha led NASA’s Long-Period Comets team within their AI accelerator, called Frontier Development Lab, where she used machine learning to develop meteor detectors. Recently this project was able to provide the first-ever instrumental evidence of an outburst of 5 meteors coming from a previously known comet, called C/1907 G1 (Grigg-Mellish). As a member of the NASA FDL AI Technical Committee, Siddha is working towards incorporating AI in many space science projects! Previously Siddha was a Deep Learning Data Scientist at Deep Vision where she worked on developing and deploying deep learning models on resource constraint edge devices. Siddha graduated from Carnegie Mellon University with a Master’s in Computational Data Science and a Bachelor’s in Computer Science and Technology from the National Institute of Technology (NIT), Hamirpur, India. She has also authored a book on Practical Deep Learning for Cloud, Mobile & Edge – O’Reilly Publishers Speech Overview: Over the last few years, convolutional neural networks (CNN) have risen in popularity, especially in the area of computer vision. Many mobile applications running on smartphones and wearable devices would potentially benefit from the new opportunities enabled by deep learning techniques. However, CNN’s are by nature computationally and memory intensive, making them challenging to deploy on a mobile device. We explain how to practically bring the power of convolutional neural networks and deep learning to memory and power-constrained devices like smartphones and web browsers.

Unstructured data are a fast growing area and a source for many innovative Big Data & Analytics solutions. Often the first idea of unstructured data seems to be that it's probably text data, even though that is just a small part. A lot of that "new data" is sensor data and especially multimedia (audio, video). Even though this part is growing extremly fast, it is very rarely used in analytics today. And even less in a real time context. In order to experience what does it mean and how does it feel (and if it is possible to make sense of it) to work with this new data in real time, Wilfried Hoge and I have created a demo that shows our own experience and explains important concepts & implementation. approaches. The demo we created shows a drill equipment as it is used to build tunnels and how to analyze the output on the conveyor belt visually with machine learning approaches.

This document describes a system for generating image captions using neural networks with attention mechanisms. It involves using a convolutional neural network (CNN) as an encoder to extract image features, and a long short-term memory (LSTM) network as a decoder to generate words describing the image. An attention mechanism is used to provide more focus on important regions of the image. Optimal beam search is employed to construct optimal captions from the generated words. The system was developed to generate more descriptive captions and help visually impaired people understand images. It was evaluated on the Flickr8K dataset using BLEU score metrics.

This document provides an overview of deep learning concepts and techniques for computer vision applications using MATLAB. It discusses traditional machine learning versus deep learning, popular pretrained deep learning models, building and training convolutional neural networks (CNNs), and using transfer learning to fine-tune pretrained models on new datasets with fewer samples. The key techniques covered are loading pretrained networks, replacing the final layers for a new task, training the modified network on a smaller labeled dataset, and evaluating the trained model on test data. The document aims to explain deep learning workflows and enable readers to implement techniques like transfer learning using MATLAB.

The document provides an overview of a presentation about Google Cloud developer tools and an easier path to machine learning. It introduces the speaker and their background and experience. It then outlines the agenda which includes introductions to machine learning and Google Cloud, Google APIs, Cloud ML APIs, and other APIs to consider. It provides examples of using various Cloud ML APIs like Vision, Natural Language, and Speech for tasks like image labeling, text analysis, and speech recognition. The goal is to demonstrate how APIs powered by machine learning can help ease the burden of learning machine learning by allowing users to leverage pre-built models if they can call APIs.

This document summarizes the DawnScience Eclipse project, which is an open source not-for-profit project on GitHub. It aims to provide APIs and reference implementations for loading, describing, slicing, transforming, and plotting multidimensional scientific data. Phase 1 from 2014-2015 defined long-term APIs and a reference implementation for HDF5 loading, data description, plotting, and slicing interfaces. Phase 2 in 2016 will release concrete implementations. The project utilizes Eclipse technologies and collaborates with scientific facilities.

Talk given at PYCON Stockholm 2015 Intro to Deep Learning + taking pretrained imagenet network, extracting features, and RBM on top = 97 Accuracy after 1 hour (!) of training (in top 10% of kaggle cat vs dog competition)

JSR 381 is a standard Java API for Machine learning, design with the goal to simplify machine learning in Java for non-experts, and make easier intergration with existing Java applications.

Core ML 3 was announced at WWDC with new features for the Core ML API. The Core ML framework allows importing machine learning models in the .mlmodel format, including updates to supported model types and operations defined in protobuf files. Core ML Tools was also updated with over 3500 new models and operations supported for conversion between Core ML and other frameworks like TensorFlow and PyTorch.

2018.12.19、フェンリル社で行った講演スライドです。

iOSDC Japan 2018でのプレゼンテーションスライドです。 ## 概要 原始のiPhoneからカメラは搭載されていましたが、深���センサが搭載されたのは比較的最近のことです。カメラやGPSが、デジタルの世界と我々が生きる現実世界を繋ぐ重要な役割を担い、アプリ開発者に多くの創造性を与えてくれたのと同様に、「奥行き」がわかるようになったというのはアプリ開発の次元がひとつ増えたようなものです。本トークではiOSでの深度の扱いについて、取得方法だけではなく、細かいパラメータの意味やMetalでの処理まで詳解します。 ## アジェンダ ## アジェンダ - 深度の種類とセンサのしくみ - 深度データを取得する - 深度データを使ってみる - できること・できないこと事例集

try! Swift Tokyo 2018 での発表資料です。Metalの使い方の話「ではなく」、Metalを通じて**普段意識する機会の少ないGPUレイヤに目を向けてみる**という内容となっておりますので、Metal使わないし興味ないという方々もぜひ。 英語版は https://www.slideshare.net/t26v0748/uiimageview-vs-metal-89418399/ にあります。 補足記事を書きました： http://shu223.hatenablog.com/entry/2018/03/05/124639 【概要】 MetalはGPUへのアクセスを提供するAPIで、OpenGLより10倍速いという謳い文句で登場しました。本セッションではMetalの基礎を解説しつつ、iOSにおけるグラフィックス描画性能をUIImageViewと比較してみます。 MetalのAPIを直接利用する機会がなくても、Metalはあなたのアプリの裏で暗躍しています。身近なクラスとの比較を通じて、普段我々が意識することのないGPUのレイヤで何が起きているのか、目を向けてみるきっかけになればと思います。

1. The presenter compared the graphics rendering performance of Metal to UIImageView to learn about GPU usage. 2. Metal was initially 10-20x faster than UIImageView for rendering images but was found to be slower after further analysis and optimization of the measurement code. 3. Two key problems were identified with the Metal implementation: processing on the CPU was blocking the GPU, and texture loading was a bottleneck. 4. Optimizations including combining operations, caching textures, and ensuring resources were in GPU memory improved the Metal performance.

I gave this talk in Jerusalem, Israel, and Palestine in 2016 with following schedule: - July 25, 2016 Azrieli College, Jerusalem - July 26, 2016 Google Campus - Tel Aviv, Israel - July 27, 2016 SigmaLabs - Tel Aviv, Israel - July 28, 2016 Birzeit University - Palestine These events were hosted by Embassy of Japan in Israel. [Description] While introducing Japanese technologies (products) such as WHILL, Moff, BONX, and etc. which Mr. Tsutsumi was involved in inventing the applications, he will talk about how BLE, a key technology of IoT, is utilized in those products.

n recent years, "IoT" or "Wearable" are one of buzzwords, so you might have interests in building hardware products. But learning how to develop electric circuits, mechanical systems or embedded systems etc. from zero is so difficult. However, iOS developers can contribute to projects of hardware products with the knowledge of Core Bluetooth / Bluetooth Low Energy (BLE), even if they are not familiar with hardware layer. In this session, you can learn the basics of Core Bluetooth / BLE (what it is, why we use it, and how it works), and practical knowledges to build apps for hardware products (how to design the apps, how to test without actual hardware prototypes, troubleshooting tips, and how the apps can be reviewed by Apple) which I learned through actual IoT/Wearable projects. This would be interesting & understandable even if you are not familiar with or have no interests in Core Bluetooth because of the actual examples.

In recent years, "IoT" or "Wearable" are one of buzzwords, so many people might have interests in building hardware products. But learning how to develop electric circuits, mechanical systems or embedded systems etc. from zero is so difficult. However, iOS developers can contribute to projects of hardware products with the knowledge of Core Bluetooth / Bluetooth Low Energy (BTLE), even if they are not familiar with hardware layer. In this session, he will introduce BTLE, show easy examples of Core Bluetooth, and share knowledges with his experiences developing more than 10 apps for IoT and Wearable products. What is Bluetooth Low Energy? Why use this? Very easy examples of how to communicate using Core Bluetooth What part was my responsibility in the projects? Communication with firmware engineer. Designing GATT Designing the behavior of the app in background Limitations in background. What are possible and impossible? State Preservation and Restoration Develop without prototypes of the hardware BTLE Module's Developer Kit Prototyping tools Build emulator apps Trouble Shootings Debugging tools Each cases: Can't find / connect / send or receive information

2015.12.13 Demoday.Tokyo #0 登壇資料 http://demoday.tokyo/

iOS 9 における Core Image の新機能について。Apple も多用しているブラーを利用した画面遷移アニメーションの実装方法等を紹���しています。

watchOS 2 でウォッチ側でも使えるようになった Core Graphics について、何ができて何ができないのか、検証しました。

iOS 9 の新機能「Audio Unit Extensions」について、そのメリットや実装方法を紹介しています。

iOS 9 の Core Image の新機能を紹介します。 - CIDetector: テキスト領域を抽出する CIDetectorTypeText / CIFeatureText - CIFilter: 43の新フィルタが追加

watchOS 2 の数ある新機能の中で、UI/UX に影響の大きそうな機能を3つ抜粋して紹介します。

watchOS-2-Sampler 実装にあたって気付いた細かい諸々について（実装上の注意点、調べてわかったこと etc..）

Core Image や vImage、GPUImage 等々、便利な画像/映像処理フレームワークが存在する昨今のiOS開発環境においても、OpenCVも依然として魅力的ですよ、という話。