講者

頁數: 1 2 3 4 - 每頁 20 筆

共有 63 位講者

人工智慧演講:

Pascal Poupart

Pascal Poupart Structure Learning in Deep Learning

The goal of data science is to extract insights from unstructured and complex data. This often hinges on the use of a good representation where suitable features can simplify tremendously the extraction of knowledge. Traditionally, features were handcrafted based on domain knowledge. However, recent advances in deep learning have shown that it is often better to use a deep structure in which features are automatically learned from data. This has completely revolutionized computer vision, speech recognition and natural language. That being said, feature engineering is now replaced by architecture engineering since practitioners spend enormous time adjusting the architecture and hyperparameters by trial and error. Hence there is a need for techniques to automatically learn the structure and hyperparameters of networks. In this talk, I will show how to automatically learn the structure of a special class of deep neural networks known as sum-product networks from streaming data. This will be demonstrated in variety of domains where it is unclear what architecture might work well.
more
Romeo Kienzler

Romeo Kienzler Realtime-Cognitive IoT using Deep Learning and Online Learning

DeepLearning frameworks are popping up at very high frequency but only a few of them are suitable to run on clusters, use GPUs and supporting topologies beyond Feed-Forward at the same time. DeepLearning4J, ApacheSystemML and TensorSpark feature all this without forcing you to learn new exotic programming languages and in addition also scales-out on well established infrastructures like ApacheSpark. In this talk we will introduce DeepLearning4J and Apache SystemML on top of ApacheSpark with an example to create an anomaly detector for IoT sensor data with a LSTM auto encoder neural network. We’ll also explain how Apache SystemML uses cost-based optimisers for Neural Network training and how TensorSpark parallelises TensorFlow on ApacheSpark.
more
王淳恆

王淳恆 深度學習與 Kaggle 實戰

透過Kaggle的資料以及實戰經驗,分享如何快速累積深度學習的技能與經驗: 1. 由手寫辨識(Digit Recognizer)競賽入門 2. 入侵物種監測(Invasive Species Monitoring) 競賽 Top 1% (5/513) 經驗分享 3. 從太空了解亞馬遜流域 (Planet: Understanding the Amazon from Space) 競賽 Top 13%(119/938) 經驗分享 4. 紐約證交所 (New York Stock Exchange) 資料預測股價
more
朱威達

朱威達 深度學習於影像風格分類之應用

本演講將針對兩種類型的影像探討深度學習在影像風格上的應用。人們往往可感受到影像的風格(style)差異,卻難以用言語定義風格。我們將以深度學習技術描述影像的紋路、視覺內容、以及其他難以言明的視覺特性,針對油畫影像以及電影海報影像進行分類。這兩類影像分別在藝術史學以及電影類型(genre)上都有較明確的定義。我們提出以類神經網路自動學習風格特徵,建構端對端的網路架構進行風格分類。
more
朱柏憲

朱柏憲 使用少量標記資料訓練聊天機器人的語意模型

聊天機器人的組成模塊中,有一些需要借助機器學習或是深度學習模型來判別的部份,可藉由一些 Transfer Learning 的技巧,減少原本訓練所需的標記資料量。同樣的方法在影像或是聲音也能拓展出許多應用。
more
吳宏彬

吳宏彬 運用 Spark 與電腦視覺科技協助瀕臨絕種的雪豹

雪豹原產於亞洲中部山區,具有敏感、機警、喜歡獨行、夜間活動的習性,加上生活在高海拔及遠離人跡地區,科學家對雪豹的了解仍然十分有限。由於非法捕獵等多種人為因素,雪豹的數量急劇減少,個體總數估計只有4,080至6,590隻,現已成為瀕臨絕種物種。 雪豹研究人員透過野外熱感應攝影機拍攝了將近一百三十萬張照片。本講次介紹如何透過MMLSpark開源工具,運用電腦視覺技術與Spark,協助研究員人員過濾出成功捕捉到雪豹的照片。並介紹專案中如何透過Transfer Learning, Data Augmentation的技巧提升辨識率。
more
吳亭範

吳亭範 從雛形到千台連網相機的挑戰

從零到壹拼湊出雛形是創業的開始。從一到一千是系統擴充能力(Scalability)與演算法通用性(Generalizability)的考驗。兩歲的盾心科技,跌跌撞撞地走過這段日子。我們見證過平時表現好的模型在資料不足的地方嚴重失常。我們有很多未標記的影片,但我們沒有財力大量標注它們。因此我們透過研究與學術合作,把少量的資源做到最大的發揮 :聰明的選擇關鍵的資料標注(Active Learning);用我們對相機資料分佈的了解,做跨視覺領域適應(Cross Domain Adaptation) 用資料充足的領域去提升資料不足領域的表現,例如黑夜或是少見相機視角等等。希望這些研究結果對於各位手上的問題也能有幫助。
more
吳尚鴻

吳尚鴻 小 App 背後的大數據與人工智慧

App 市場是全球性的,但是非常競爭。開發者採取了許多不同的策略以獲取先期使用者(early adopters)以及產品市場契合(product-market fit)。多樣的策略、高度的競爭、以及許多不確定因素往往讓人以為 app 開發與執行必須要有大量的資源才能獲得成功。在這個演講中,講者基於多年深入研究機器學習以及參與並指導 app 開發團隊的經驗,帶領大家進行個案探討,從中一窺 app 成功的要素,以及如何有效利用巨量資料與深度學習創造競爭優勢。
more
吳毅成

吳毅成 CGI and CGI

首先,我將先談第一個CGI,那就是我們CGI(Computer Games and Intelligence)實驗室,過去在人工智慧遊戲的研發成果。除了早期發明六子棋外,也研發出各種遊戲的人工智慧程式,如六子棋、圍棋、象棋、禁圍棋、暗棋、麻將、Nonogram、2048等。將會探討一些機器學習方法,用來改善這些程式強度。目前這些程式獲得超過50面國際競賽金牌;其中所研發出的2048人工智慧程式,是全世界第一支程式能打到65536磚塊的2048程式。 我將談的第二個CGI,那就是我們CGI(CGI Go Intelligence)圍棋程式之設計,包括我們所發展出新的類神經網路。該程式,在2017年八月世界智能圍棋公開賽中,獲得預賽全勝冠軍、決賽亞軍,其中曾擊敗騰訊公司的絕藝、DeepZenGo;也是第一個學界程式在正式的人機賽中,打敗職業九段棋士。
more
李宏毅

李宏毅 GAN (不是髒話)

為什麼最近在做深度學習的人都一直在 GAN 來 GAN 去的呢?那是因為有一個新的技術叫生成式對抗網路 (Generative Adversarial Network, GAN),這個技術顯然是深度學習的下一個熱點。在機器學習領域,回歸 (regression) 和分類 (classification) 這兩項任務的解法人們已經不再陌生,但是如何讓機器更進一步創造出有結構的複雜物件 (例如:圖片、文句) 仍是一大挑戰。用生成式對抗網路,機器不只可以畫出以假亂真的人臉,也可以根據一段敘述文字,自己畫出對應的圖案。本課程希望能帶大家認識生成式對抗網路這個深度學習最前沿的技術。
more
李祈均

李祈均 機器智能與人類行為: 跨領域決策分析於醫療應用

長年來領域專家透過解析感測器訊號、歸納有用資訊、並進行適切處理以完成各種與人息息相關的決策應用(decision analytics)。在(非)結構化資料快速累積、機器學習演算法高速進展,並配合深度跨領域知識整合下,依機器智開發嶄新的訊息決策分析架構正蓬勃發展中 。在此一演講中,我們會以多個醫療應用為主軸,分享我們近期透過機器智能並以人本運算角度為出發而開發出「高可信度(reliability)」、「可規模化(scalable)」、「客觀且一致(objective)」 的人類訊息行為分析決策於醫療應用(AI-in-Health Behavior Analytics)。實例分享包涵:臨床血癌診斷輔助、未來中風風險評估、可計算式自閉症失調行為、及內在情緒感受辨識等。此一技術發展更仰賴深度跨領域合作,以提供專家現實實務上有效且全新變革式的分析工具。
more
李育杰

李育杰 Smart Sensing and Continuous Monitoring

IoT applications are expected to have huge impact to us in a very near future. People imagine Intelligent Transportation Systems, Intelligent Care Systems, smart buildings, or even smart cities may become reality. With smart sensing technology and machine learning algorithms, we are able to understand the environment states and monitor any particular anomalous conditions. Data driven approach becomes a key corner stone for the success of most IoT applications. However, compared to traditional data analytics, data analysis in IoT applications seems to be more challenging simply due to the huge amount of data that can be easily generated by IoT devices in a small period and we have to deal with them using very limited computational resources. In this talk, we introduce our envelope representation for IoT time series data which can be considered as a sparse coding for the time series. With this representation, we are able to deal with IoT data and develop anomaly detection algorithm under the hardware limitations. We will show its applications in monitoring the running machine status and user identification.
more
宋政隆

宋政隆 深度學習環境建置與模型訓練實務

近年來,計算規模的提升激發深度學習在大量資料下的成效,快速地訓練模型成為深度學習應用之成功關鍵。平行化計算是加速訓練大型資料集或是複雜網路的主要手段,其涵蓋了軟硬體配置、深度學習框架設定、資料的預處理、到 hyper parameters 調整優化,過程中有許多環節可能影響計算速度或正確性。此外,平行化系統可能是配置有多台機器分散式環境,管理與監控計算資源也是一項重要的課題。在這個演講,我們將分享校調深度學習訓練參數的經驗以及管理監控深度學習運算資源的實務。
more
林國銘

林國銘 從手解演算法看 AI,搶錢搶糧搶未來

1. 高薪人才需求: 談談中國的中高階人才需求,不可不知的市場現況 2. 搶錢搶糧搶未來: 學習沒有懶人包,AI 教育的重要性 3. 動手學 AI: 從算法著手,從應用回眸 4. 手撕演算法: 把技術貢獻給全世界,力量留給臺灣 5. 加碼不小心分享: 超低調手刻 AI 程式交易演算法經驗談
more
林守德

林守德 人工智慧第三波革命

人工智慧演進至今已經邁入第三波。本演講將闡述第三波人工智慧在看什麼樣的問題?尋求什麼樣的技術?產生什麼樣的影響?講者也將討論未來人工智慧的走向,以及何謂目前人工智慧的藍海。
more
林軒田

林軒田 The Interplay of Big Data, Machine Learning, and Artificial Intelligence

Big data has been widely recognized as crucial assets to enterprises across the globe while artificial intelligence has become one of the most transformative forces this time, and also brought fame to machine learning overnight. In this talk, Hsuan-Tien will share his perspective of the interplay among the three concepts according to his substantial research results and solid experience in the field of applications.
more
林宗瑤

林宗瑤 AI Technologies for Embedded Devices

There are more and more intelligence demands on the embedded system such as smart phone, smart phone, drone…. The challenge for embedded AI is limited power budget and computing resource. Challenges also mean opportunities. This speech will introduce the technologies to enable AI on the embeded devices.
more
周志成

周志成 Learn Data Science by Doing a Kaggle Competition

Once sequenced, a cancer tumor can have thousands of genetic mutations. But the challenge is to distinguish the mutations that contribute to tumor growth from other neutral mutations. Currently the interpretation of genetic mutations is done manually, which is time-consuming and knowledge-demanding. Therefore, Classifying Clinically Actionable Genetic Mutations challenged the Kaggle community to develop algorithms that automatically classify genetic variations based on evidence from text-based clinical literature. As a problem of natural language processing (NLP) and machine learning, this Kaggle competition is not a trivial task. The main difficulties are three fold. First of all, interpreting clinical evidence from literature is very challenging even for human specialists, since it takes expertise of domain knowledge and lengthy time of reading to understand key information in the literature and make classification accordingly. Secondly, only 3321 training data is given, which is far less compared with other Kaggle challenges and will increase the risk of overfitting. Moreover, much of the test data is machine-generated, which boosts the complexities of this task. To tackle this challenge, cooperation between teams of data science and clinical medicine was built up. Algorithms with insights from both fields were developed. To extract the key information of genetic mutation from text, hand crafted feature engineering was done with several state-of-the-art NLP methods. To obtain effective representations of the texts, we also consulted medical experts about how specialists read and classify genetic mutations, and adjusted our approaches accordingly. Furthermore, efforts were spent on classifiers to prevent issue of overfitting resulting from small training dataset. From the experience of participating in this competition, we demonstrated how cooperation between different expertise can bring in further insights to deal with challenging data science problem as this one.
more
周恩誌

周恩誌 以深度學習估算車牌競標成交價

764/5000 因著社會迷信,擁有吉祥數字的車輛牌照在拍賣會中往往能賣得非常高的價錢。 與其他常見的拍賣不同的是,現時車輛牌照在拍賣前並沒有估價。是項研究的目標是構建一個準確的模型用來估算香港車牌的拍賣成交價。因為車牌的價值取決於牌上的數字及其語義,車牌競標成交價的估算可以被看成為自然語言處理的一種,並就此構建相應的深度神經網絡模型。
more
紀懷新

紀懷新 Optimizing for User Experience with Data Science and Machine Learning

Understanding users and optimizing for user experience are critical parts of building successful apps and services. While there had been a tremendous amount of past work studying user and social interactions, in practice, it wasn’t until quite recently that researchers are able to study these interaction mechanisms at scale easily. In this talk, I will illustrate data-driven approaches to understand what are happy engaged users, and present case studies of how we utilize novel machine learning techniques to optimize for long-term user engagements in practice.
more
頁數: 1 2 3 4 - 每頁 20 筆

歡迎在此登錄您的大名及電子郵件地址,日後任何台灣資料科學協會舉辦的相關活動,我們將會以電子郵件通知您。謝謝。