--- title: 資訊科技產業專案設計課程作業 4 tags: 資產hw4 description: 資產hw4 --- ## 職缺(一) Mediatek: [AI researcher for computer vision and multimedia](https://careers.mediatek.com/eREC/JobSearch/JobDetail/MTK120210813002?langKey=en-US) :::info Job Description * Deep learning/AI algorithm development * Computer vision algorithm development * Image/Video processing algorithm development Requirement * Programming language: C/C++ or Python * Deep learning tools: tensorflow or pytorch ::: ### Self-assessment * Programming language: C/C++ or Python * 熟悉此兩種語言 * 能力評比: * C/C++: 3.5/5 * Python: 3.5/5 * Deep learning tools: tensorflow or pytorch * 經常使用pytorch,可修改API的source code並套用至自己的任務上。幾乎未使用tensorflow。 * 能力評比 * pytorch: 3.5/5 * tensorflow: 1/5 * 除Requirement要求外,工作內容還包含了影像處理、電腦視覺。 * 曾在課堂上實作過Sobel filter、Median filter、threshold等影像處理相關算法 * 也使用opencv操作過Camera Calibration、Stereo Disparity、SIFT等電腦視覺相關問題。 ## 職缺(二) Mediatek: [AI Algorithm Engineering-ChuPei](https://careers.mediatek.com/eREC/JobSearch/JobDetail/MTK120190523006?langKey=en-US) :::info Job Description * AI deep learning and machine learning algorithm development. * Training/fine tuning deep learning NN model parameters for image/video quality enhancement. * Deploy and optimize NN model on edge devices. Requirement * Familiar with deep learning and machine learning algorithm and model training process * Familiar with deep learning framework, like TensorFlow, PyTorch and programming language, like Python or C/C++ * It is nice to have experience in model optimization and deployment on edge devices. ::: ### Self-assessment * Familiar with deep learning and machine learning algorithm and model training process * 目前實驗室就是在做深度學習相關的題目,每周皆會聽兩次AI相關paper報告,常與同儕討論深度學習相關的問題。碩班幾乎都在train模型,對使用pytorch訓練模型的流程還算了解。 * Familiar with deep learning framework, like TensorFlow, PyTorch and programming language, like Python or C/C++ * 能力評比: * C/C++: 3.5/5 * Python: 3.5/5 * pytorch: 3.5/5 * tensorflow: 1/5 * It is nice to have experience in model optimization and deployment on edge devices. * 模型的優化應該可分成空間/時間的優化以及performance的優化,前者我未曾做過,但若要優化的話應可尋找合適的模型,或自行刪減一些層進行實驗;後者不外乎調整參數,察看結果後發現問題,搜尋相關文獻以解決問題,我已有類似經驗。 * 過去未曾部屬在邊緣裝置上,這部分可能需要自行購買開發板操作一次,積累經驗。 ### 面試題目 :::info **Quesion:** [CNN跟Fully connected誰能夠取到比較多特徵](https://www.ptt.cc/bbs/Soft_Job/M.1535716676.A.12B.html) * (個人理解)CNN應該是FC的一種特例,在CNN中特徵數量的計算應是使用filter的數量,但在FC中無法確定取到了多少特徵,因為我們不知道FC是以幾個參數決定一個特徵。 * 其他關於CNN vs FC的資料 * [Why are convolutional layers better than fully connected layers for images?](https://www.quora.com/Why-are-convolutional-layers-better-than-fully-connected-layers-for-images) * [Fully Connected vs Convolutional Neural Networks](https://medium.com/swlh/fully-connected-vs-convolutional-neural-networks-813ca7bc6ee5) ::: :::info **Quesion:** [業界賺錢需要壓成本,但同時performance不能降時,該怎麼辦?](https://www.ptt.cc/bbs/Soft_Job/M.1535716676.A.12B.html) * AI的成本大多集中在收集資料、標註資料上,收集資料的部分只能找找看有沒有開源的資料集,又或著評估是否有自動產生資料的方法,標註資料的部分可以使用Active Learning的方法降低需要標註的資料量,但會犧牲一點performance。(總不能又要馬兒好,又要馬兒不吃草) ::: :::info [AI面試題目彙整](https://github.com/DarLiner/Algorithm_Interview_Notes-Chinese) :::