WORK

摘要： 本週對於加密貨幣可說是愁雲慘霧的一個禮拜，不僅以太幣跌至今年低點180美元，造成幣圈以及ICO公司之財務壓力。Sec也宣布暫停兩款加密貨幣ETN之交易。相較加密貨幣，區塊鏈技術在業界可說是蓬勃發展，IBM、Maersk等大型公司皆已投入其發展行列。另一方面，在機器學習的發展中，Google發展出無需編碼的TensorBoard應用程式工具What if Tool，讓使用者探索模型結果；發展多年的Adam優化演算法套件也有了新的發展¬– AdamW號稱是目前神經網絡最快的訓練演算法。

摘要： 新浪財經訊“第五屆金融科技外灘峰會”於2018年7月7日-8日在滬召開，SFI 理事單位代表、螞蟻金服副總裁韓歆毅發表演講。

若讀者有推薦的網站可以在討論串留下相關訊息，我們會立即更新內容!

Open Data

Economic and Financial Database

摘要： 關於機器學習的網站，在著名的reddit上獲得國外網友的熱情追捧。

摘要： 不論是AI也好，其他學科也好，學習、研究的過程中不斷反思學科的歷史，總結學科的發展現狀，找出最重要的理念，總能讓人能「吾道一以貫之」。軟體工程師James Le近期根據他研究的經驗總結出了AI研究必須要知道的十種深度學習方法。

摘要： 連假開始，你的論文存貨還夠嗎？對人工智慧感興趣的你，是否還停留在碎片閱讀階段？想要進行更深一步的學習，本篇推文中網羅了人工智慧領域15篇精選論文，讓你及時了解AI學科前沿成果。

摘要： 11節 麻省理工學院 的通用人工智慧課程。講師陣容超豪華：谷歌技術總監Kurzweil、特斯拉AI總監Andrej Karpathy、波士頓動力CEO Marc Raibert、OpenAI的聯合創始人Ilya Sutskever等。課程內容包括：通用人工智慧、計算認識科學、認知建模、深度學習等。文中附有課程講義下載和課程視頻地址。

摘要： 「贏者詛咒」：贏得拍賣品的中標者出價高於其他競標者，但他很可能對拍賣品估價過高，支付了超過其價值的價格，從而贏得的拍賣品的收益會低於正常收益甚至爲負。換句話說，就是當你一心想要贏得競標時，卻偏離了你原本的目的：當我們爲各種測評任務中取得的分數歡欣鼓舞時，可能我們已經受到了「贏者詛咒」。

Big data has both high volume and high velocity – one way this manifests is as silos of in-situ data representing departments in banks that are very difficult to move and integrate to obtain a single coherent customer view. Further, the ability to perform data analytics – dynamically and in near real-time – of rapidly changing customer and market data is increasingly critical for competitiveness. By considering the distributed nature of financial data storage and the velocity of financial markets, the objective of this RP is to develop distributed and real-time machine learning methods to identify decentralised and dynamic models for financial analysis, prediction, and risk management.

This project will develop (i) methods to identify cross-effects between different data resources, regions, sectors, and markets, (ii) distributed versions of methods to identify decentralised models that include individual local model components learned from local resources and cross-impact model components learned from data resources in other regions/sectors/markets, and (iii) real-time learning methods to update decentralised models and address financial market velocity.

Based on the distributed and cloud computing infrastructure, this approach should address the weakness of existing data-centralised and off-line machine learning methods, which fail to consider the cost of data transportation, storage, and fast timevarying characteristics of financial markets. The originality of this approach is its dynamic integration, by distributed and real-time mining, to maximise the effectiveness and efficiency of big data analysis.

Early Stage Resercher working on the project: Sergio Garcia Vega

轉貼自： Finance BigData.eu