我一直對於學習感到非常大的興趣,但學校的教育並不能滿足我,我在學校求學的不同階段,其實都有著不同的掙扎。我任性的用我自己想要的方式學習,一路走來雖然跌跌撞撞,因此對於學習有了深刻的體會。
我的父母非常重視我的教育,因此在國小國中的時候,幾乎每天下課後都還有才藝班的課要上。我曾經上過的才藝班有書法、作文、英文、畫畫、音樂、鋼琴。這些課主要是爸媽覺得我應該去上而幫我做出選擇的,我的個性也算是順從,都乖乖的去上這些課,也盡量花時間練習。小時候心裡難免會羨慕有很多時間可以玩耍的同學,因為相較之下,我在做的事情不是那麼有趣。除了這些才藝班的課之外,我爸媽還想盡辦法鼓勵我念書。我爸媽不太給我買玩具,倒是書買了一大堆,牛頓,科學小百科,偉人傳記,大英百科全書,世界美術館全集等等。每唸完一本課外書,我會得到五塊錢的零用錢。
因為國小國中的時候,都被規畫好要學甚麼了,在進到建中這麼自由的地方,我覺得我好像頓時失去了目標。我不清楚人生的方向,也無法認同學校教育以及考試的價值,我無法理解為什麼考試考高分就代表成功,就代表人生的成就。在高二的時候,我常翹課去打撞球,鬼混,也花了非常多的時間在社團活動上,但對於學校的課業就是提不起興趣。我記得我高三有一次模擬考,排名是全校九百多名,我們一整個年級也不過比一千人多一些些。雖然學校的課程無法引起我的興趣,但我卻對生物學深深地著迷。我考進了中研院的高中生物人才培育計畫,每兩個禮拜的周日到中研院去,早上上課,下午到實驗室做實驗。參與這個計畫對我有非常深遠的影響。從動物生理學、植物生理學、細胞生理學、到分子生物學,我廣泛學習生物的知識。下午的實驗課也讓充滿好奇心的我大開眼界。我看到跟臉盆一樣大的椰子蟹,解剖跟手掌一樣大的文蛤觀察內臟,學習PCR技術,學習萃取葉綠素,也看到費洛蒙如何讓一整箱蟑螂瘋狂振翅。這個培育計畫讓我深深的愛上生物學。我從中研院借了很多英文教科書回家,也常到台大醫學院的圖書館去印研究論文來看。我高三的狀態,應該用著魔來形容。我因為想要參加生物奧林匹亞競賽,瘋狂的請公假到重慶南路的漢堡王去K書。當年我的導師對我的寬容,現在想起來簡直不可思議。我記得有一次她把我叫到辦公室去,她問我知不知道我自己在做什麼,我說我知道(回頭看,其實我並不是這麼肯定…),她就相信我,讓我依照我的心意去學習了。我花很多時間研究Kreb’s Cycle, Signaling pathways, 粒線體以及葉綠體的電子傳遞鏈,DNA複製的分子機制 (我甚至買了Arthur Kornberg著作的DNA Replication)。好笑的是,我雖然花了這麼多時間學習生物學,但高中生物教材無法引起我的興趣,我在學校的生物學考試成績並不算頂尖,許多對於生物學並沒有特別的熱情的同學考試分數是比我高的。後來我考進了生物奧林匹亞國手培訓營,但是因為我平常沒有機會練習生物實驗,實驗的技術不是很好,最後沒有選上國手。這件事情讓我耿耿於懷許多年,因為我把我高中升大學的希望都寄託在奧林匹亞競賽上,我花了這麼多時間學習卻沒選上,我覺得我的人生失敗了。因為我把高三的時間大多都花在學習生物學上了,國手培訓營後已經沒有多少時間再準備聯考,而且我還處在挫敗的陰影當中,我第一次大學聯考的成績很差。後來我在重考班再準備了一年,才考進了台大醫科。
當年的我以為高中這段的學習經驗是失敗的,覺得這是不自量力的行為。但我現在意識到,高中那段學習歷程對我來說,是最重要的,第一次自我啟蒙的經驗。建中自由的風氣,再加上導師的信任以及家人的包容,讓我有機會充分的探索我的興趣並且積極的去追求。我被考試的框架侷限了,以為沒有選上國手,沒有得到金牌就是失敗。但其實我在高中時就養成了查找以及閱讀科學研究論文的習慣,並且開始培養獲取陌生學術領域知識的能力。我在高中所培養的閱讀英文教科書的習慣,也讓我日後受用無窮。
我在醫學系的學習歷程比較沒有這麼劇烈的掙扎,頂多在前四年的時間懷疑為何要死背大量的知識。醫學系要背誦的知識量實在是太大,常常讓我感到心煩,有種唸到想吐的感覺。但到了大五進醫院見習之後,我的態度有了一百八十度的大轉變。當病人以及家屬問起病情時,我必須要能清楚地回答他們的問題。當病人有狀況需要處理時,我沒有時間再去查閱書籍,我需要的知識必須要像反射一樣快速的被提取。如果我沒有學好,沒有記清楚,就無法對病患做出最好的處置。因為有了責任感,我充分的理解也認同熟記醫學知識的重要性,因此開始認份認真的念書。我在醫學系唯一一次得到書卷獎,是在臨床工作最繁重的大七的時候。我清楚的感受到,我被賦予越多照顧患的責任,我學得越認真。
我的博士班的學習歷程是最痛苦的一段。博士班雖然是高等教育的一環,但幾乎沒有一個既定的框架。博士班還有一些學分要修,但是並不多。我在博士班修過的課印象都不深,許多課程其實大學都上過了,但偏偏學校不給抵免,非得要我再修一次不可。倒是我自己選的程式設計概論對我的人生造成了非常深遠的影響。博士班絕大部分的時間都是在實驗室度過的,因此博士班導師是影響博士班學習歷程當最重要的一個關鍵。回頭檢視起來,我會離開博士班一個很重要的原因是,我和我的指導教授對於科學研究的一些基本想法並不契合,對我個人來說,發展新技術是我比較喜歡的方向,但我的指導教授認為生物研究應該著重在探討現象背後的原理。簡單來說,我們兩個人重視的點不一樣,我不適合在他的實驗室做研究,我比較適合在生醫工程的博士班就讀。我處在一個不適合的環境裡其實給了我一個意料之外的機會,為了追求我理想中的科學研究,我必須要仰賴自我學習的能力去獲取許多我的實驗室並沒有的知識與技術。
這幾年下來,我對於學習新的知識和技術有了比較清楚的領會。我要學習一個新的領域時,我會先確認學習的目的。對我個人來說,清楚學習的目的之後才能找到學習的動機。純粹為了好玩,一時興起而學習很難持久,也很難達到好的效果。比如說,我學習深度神經網路是為了實現數位病理輔助診斷系統,進而彰顯數位病理系統的價值。確立學習的目的之後,我會開始嘗試了解這個知識領域的大架構,接著試著找到這個知識領域裡面關鍵的基礎假設與原理。比如說,深度神經網路是機器學習的一個分支,我會先學習機器學習領域的大架構,接著了解演算法學習的原理。我會先從最簡單的基礎原理開始,確認完全了解之後,就開始朝我需要學習的子領域學習。在學習的時候可以動手實作非常重要,我很認同諾貝爾獎得主Richard Feynman所說的 “What I cannot create, I don’t understand.” 我動手實做了Perceptron, Multi-layer perceptron, fully connected deep neural network, convolutional network,在這過程中,透過程式的實踐印證我在理論上的學習。總結來說,我對於學習新領域的心得是,要達到有效的學習,首先要清楚學習的目的,接著了解知識體系大架構以及涵蓋的範疇,理解關鍵原理,接著要明確界定學習的目標及範圍,最後是深入的學習、並且反覆透過實踐來檢驗學習成效。同樣重要的是熟悉學習的工具,找到關鍵的資訊管道,這對於提高學習效率至為關鍵。
創業,對我來說又是另外一個自我學習的巨大挑戰。創業是個無邊無際的問題,我目前遇到的最大的挑戰在於,我還無法清楚的界定我所需要學習的知識及技術的領域,更遑論去了解其架構。當然,這問題的核心在於,創業的過程本身就無法清楚界定。試著要去找出一個清楚的脈絡,我認為是極其困難的。我認為創業比較像是一個戰爭。我們可以清楚觀察並理解的,是每個小戰役的樣貌。試著贏得一個戰役,帶著在這個戰役學習到的經驗,前往下一個戰場,才有機會逐漸拼湊出這個戰爭的大局。在這過程當中,創業家必須要既短視又有遠見,要可以讓公司贏得一個戰役,但是又要預見下一個戰場並且做好準備,這樣長短程思考的轉換以及平衡,是個非常艱鉅的挑戰。我想,三年之後我再回頭看,應該會對於這個學習的歷程有更清楚的理解。現在,也就只能帶著我所學到的勇敢前進了。
同時也有10000部Youtube影片,追蹤數超過2,910的網紅コバにゃんチャンネル,也在其Youtube影片中提到,...
「fully connected layer」的推薦目錄:
- 關於fully connected layer 在 謝銘元:失敗並不可恥但要有用 Facebook 的最讚貼文
- 關於fully connected layer 在 國立陽明交通大學電子工程學系及電子研究所 Facebook 的最讚貼文
- 關於fully connected layer 在 Lee388 Hi Fi 發燒專頁 Facebook 的最佳解答
- 關於fully connected layer 在 コバにゃんチャンネル Youtube 的最佳解答
- 關於fully connected layer 在 大象中醫 Youtube 的最佳解答
- 關於fully connected layer 在 大象中醫 Youtube 的最讚貼文
- 關於fully connected layer 在 CS231n: Convolutional Neural Networks (CNNs / ConvNets) 的評價
- 關於fully connected layer 在 What do the fully connected layers do in CNNs? - Cross ... 的評價
- 關於fully connected layer 在 How to implement a neural network with a not-fully-connected ... 的評價
- 關於fully connected layer 在 Perceptual Loss Github Keras 的評價
fully connected layer 在 國立陽明交通大學電子工程學系及電子研究所 Facebook 的最讚貼文
【演講】2019/11/19 (二) @工四816 (智易空間),邀請到Prof. Geoffrey Li(Georgia Tech, USA)與Prof. Li-Chun Wang(NCTU, Taiwan) 演講「Deep Learning based Wireless Resource Allocation/Deep Learning in Physical Layer Communications/Machine Learning Interference Management」
IBM中心特別邀請到Prof. Geoffrey Li(Georgia Tech, USA)與Prof. Li-Chun Wang(NCTU, Taiwan)前來為我們演講,歡迎有興趣的老師與同學報名參加!
演講標題:Deep Learning based Wireless Resource Allocation/Deep Learning in Physical Layer Communications/Machine Learning Interference Management
演 講 者:Prof. Geoffrey Li與Prof. Li-Chun Wang
時 間:2019/11/19(二) 9:00 ~ 12:00
地 點:交大工程四館816 (智易空間)
活動報名網址:https://forms.gle/vUr3kYBDB2vvKtca6
報名方式:
費用:(費用含講義、午餐及茶水)
1.費用:(1) 校內學生免費,校外學生300元/人 (2) 業界人士與老師1500/人
2.人數:60人,依完成報名順序錄取(完成繳費者始完成報名程序)
※報名及繳費方式:
1.報名:請至報名網址填寫資料
2.繳費:
(1)親至交大工程四館813室完成繳費(前來繳費者請先致電)
(2)匯款資訊如下:
戶名: 曾紫玲(國泰世華銀行 竹科分行013)
帳號: 075506235774 (國泰世華銀行 竹科分行013)
匯款後請提供姓名、匯款時間以及匯款帳號後五碼以便對帳
※將於上課日發放課程繳費領據
聯絡方式:曾紫玲 Tel:03-5712121分機54599 Email:tzuling@nctu.edu.tw
Abstract:
1.Deep Learning based Wireless Resource Allocation
【Abstract】
Judicious resource allocation is critical to mitigating interference, improving network efficiency, and ultimately optimizing wireless network performance. The traditional wisdom is to explicitly formulate resource allocation as an optimization problem and then exploit mathematical programming to solve it to a certain level of optimality. However, as wireless networks become increasingly diverse and complex, such as high-mobility vehicular networks, the current design methodologies face significant challenges and thus call for rethinking of the traditional design philosophy. Meanwhile, deep learning represents a promising alternative due to its remarkable power to leverage data for problem solving. In this talk, I will present our research progress in deep learning based wireless resource allocation. Deep learning can help solve optimization problems for resource allocation or can be directly used for resource allocation. We will first present our research results in using deep learning to solve linear sum assignment problems (LSAP) and reduce the complexity of mixed integer non-linear programming (MINLP), and introduce graph embedding for wireless link scheduling. We will then discuss how to use deep reinforcement learning directly for wireless resource allocation with application in vehicular networks.
2.Deep Learning in Physical Layer Communications
【Abstract】
It has been demonstrated recently that deep learning (DL) has great potentials to break the bottleneck of the conventional communication systems. In this talk, we present our recent work in DL in physical layer communications. DL can improve the performance of each individual (traditional) block in the conventional communication systems or jointly optimize the whole transmitter or receiver. Therefore, we can categorize the applications of DL in physical layer communications into with and without block processing structures. For DL based communication systems with block structures, we present joint channel estimation and signal detection based on a fully connected deep neural network, model-drive DL for signal detection, and some experimental results. For those without block structures, we provide our recent endeavors in developing end-to-end learning communication systems with the help of deep reinforcement learning (DRL) and generative adversarial net (GAN). At the end of the talk, we provide some potential research topics in the area.
3.Machine Learning Interference Management
【Abstract】
In this talk, we discuss how machine learning algorithms can address the performance issues of high-capacity ultra-dense small cells in an environment with dynamical traffic patterns and time-varying channel conditions. We introduce a bi adaptive self-organizing network (Bi-SON) to exploit the power of data-driven resource management in ultra-dense small cells (UDSC). On top of the Bi-SON framework, we further develop an affinity propagation unsupervised learning algorithm to improve energy efficiency and reduce interference of the operator deployed and the plug-and-play small cells, respectively. Finally, we discuss the opportunities and challenges of reinforcement learning and deep reinforcement learning (DRL) in more decentralized, ad-hoc, and autonomous modern networks, such as Internet of things (IoT), vehicle -to-vehicle networks, and unmanned aerial vehicle (UAV) networks.
Bio:
Dr. Geoffrey Li is a Professor with the School of Electrical and Computer Engineering at Georgia Institute of Technology. He was with AT&T Labs – Research for five years before joining Georgia Tech in 2000. His general research interests include statistical signal processing and machine learning for wireless communications. In these areas, he has published around 500 referred journal and conference papers in addition to over 40 granted patents. His publications have cited by 37,000 times and he has been listed as the World’s Most Influential Scientific Mind, also known as a Highly-Cited Researcher, by Thomson Reuters almost every year since 2001. He has been an IEEE Fellow since 2006. He received 2010 IEEE ComSoc Stephen O. Rice Prize Paper Award, 2013 IEEE VTS James Evans Avant Garde Award, 2014 IEEE VTS Jack Neubauer Memorial Award, 2017 IEEE ComSoc Award for Advances in Communication, and 2017 IEEE SPS Donald G. Fink Overview Paper Award. He also won the 2015 Distinguished Faculty Achievement Award from the School of Electrical and Computer Engineering, Georgia Tech.
Li-Chun Wang (M'96 -- SM'06 -- F'11) received Ph. D. degree from the Georgia Institute of Technology, Atlanta, in 1996. From 1996 to 2000, he was with AT&T Laboratories, where he was a Senior Technical Staff Member in the Wireless Communications Research Department. Currently, he is the Chair Professor of the Department of Electrical and Computer Engineering and the Director of Big Data Research Center of of National Chiao Tung University in Taiwan. Dr. Wang was elected to the IEEE Fellow in 2011 for his contributions to cellular architectures and radio resource management in wireless networks. He was the co-recipients of IEEE Communications Society Asia-Pacific Board Best Award (2015), Y. Z. Hsu Scientific Paper Award (2013), and IEEE Jack Neubauer Best Paper Award (1997). He won the Distinguished Research Award of Ministry of Science and Technology in Taiwan twice (2012 and 2016). He is currently the associate editor of IEEE Transaction on Cognitive Communications and Networks. His current research interests are in the areas of software-defined mobile networks, heterogeneous networks, and data-driven intelligent wireless communications. He holds 23 US patents, and have published over 300 journal and conference papers, and co-edited a book, “Key Technologies for 5G Wireless Systems,” (Cambridge University Press 2017).
fully connected layer 在 Lee388 Hi Fi 發燒專頁 Facebook 的最佳解答
GRANDIOSO K1X
"Two revolutions integrated. Two revolutions connecting people to music... The VRDS-ATLAS and Master Sound Discrete DAC - now come to fruition in an integrated player that carries the name… Grandioso ( Grandioso (music): Grand or noble)."
An unprecedented transport mechanism that invites you to the furthest reaches of sound.
VRDS (Vibration-Free Rigid Disc-Clamping System) is Esoteric's unique CD/Super Audio CD transport technology, which has been specially designed to clamp the disc to a same-diameter turntable in order to eliminate the disc's rotational vibration, as well as any extraneous vibration emanating from the mechanism itself. It also corrects any warpage of the disc, dramatically improving the accuracy of both the optical pickup and the disc pit layer's corresponding optical axis, while also minimizing servo current. This all adds up to dramatically reduced disc read error for exceptional audio quality.
The new VRDS-ATLAS platform boasts unprecedented mechanical perfection and superlative audio quality, with heavyweight construction that is fully 127% heavier than previous iterations (6.6kg for the mechanism alone, and 13.5kg including its base). This highest rigidity and weight in the history of VRDS mechanisms has the effect of greatly reducing any and all vibration that could adversely affect audio quality. The transport's larger new side panels and bridge are all formed of SS400 steel, and its turntable is made from duralumin, which is renowned for its excellent sound quality. Also, its spindle features a newly designed thrust bearing system, for a friction-free one-point spindle support system that ensures smooth, noiseless turntable rotation.
Esoteric's impressive VRDS-ATLAS is the quietest and most elegant transport in the history of VRDS. The key to achieving this was a resolute focus on mechanical grounding technology that more effectively reduces vibration.
Adopting a wide & low-profile design for the entire mechanism helped achieve a lower center-of-gravity, and relocating the turntable motor to below the turntable from its previous position above the bridge dramatically shortened the route for grounding vibration and reducing mechanical noise. Tray hollowing has also been minimized for increased rigidity, and special stoppers made from vibration-absorbing elastomer cancel out resonance when the tray is stored.
The Grandioso K1's mighty power supply was the key to its powerful and profound three-dimensional sound. With the introduction of its new X version, this power supply has been dramatically enhanced to add further depth to its exceptional audio tone. Its built-in D/A converter features independent power transformers for left and right channels, and the K1X includes a total of four independent toroidal power supply transformers. Furthermore, technologies cultivated in the development of the Grandioso P1X/D1X have also been adopted, including a new low-feedback DC power supply regulator featuring a discrete circuit configuration which contributes to its powerful open sound. The K1X also contains a total of 76 EDLC* supercapacitors (with a total capacity of 2,050,000µF or 2.05F). This greatly increased power supply capacity provides a tremendously improved sound quality for exceptional resolution in the lower frequency ranges.
*EDLC (Electric Double-Layer Capacitor):A special type of capacitor boasting an astoundingly high capacity compared to the electrolytic capacitors used in conventional audio equipment.
The VRDS-ATLAS transport mechanism is centrally located within the chassis, where it is fixed to a 5mm-thick steel bottom plate and supported by four proprietary pinpoint feet (patent no. 4075477 and 3778108). This effectively isolates the rotational mechanism from vibration. The internal structure of the chassis features a double deck design, with the audio boards on the upper level and power supply circuits and transformers on the lower level to minimize magnetic flux leakage and vibration, while also shortening the power supply wiring.
Semi-Floating Top Panel
The enclosure's top panel utilizes a screwless semi-floating structure that further contributes to an open and expansive sound.
Master Sound Discrete DAC
Bringing out all the dynamics and energy of the original master recordings.
Master Sound Discrete DAC Bringing out all the dynamics and energy of music.
Unattainable by integrated chips, our goal was to assemble carefully selected and tested discrete components into a complete circuit that could perfectly reproduce all the dynamics and energy of music. Our top engineering teams put their pride on the line to design and produce a quality of sound that can only be found in the Master Sound Discrete DAC discrete D/A converter.
The K1X's Master Sound Discrete DAC is the most revolutionary two-channel stereo DAC circuit in Esoteric's history, and is based on the original circuit developed for the Grandioso D1X monaural D/A converter.
As an extension of the D1X's design philosophy, a luxurious volume of materials have been invested in the construction of K1X's Master Sound Discrete DAC. For example, it features 32 separate elements for each channel. Key components such as a clock driver, logic circuit, and capacitors and resistors, are kept independent for each of these 32 elements to ensure the purest output without no loss of musical energy.
The K1X's independently developed Δ∑ modulator supports 64-bit/512Fs and the latest high-end digital formats, including of 22.5MHz DSD and 768kHz PCM signals. The FPGA's* dedicated digital processing algorithm was developed exclusively for the Master Sound Discrete DAC, and has been fully optimized for outstanding playback of both DSD and PCM digital audio data.
Advanced Quality Control
With a discrete DAC, where component tolerances are directly linked to arithmetic precision, highly advanced quality control is also required for the manufacture of electronic circuit boards. Esoteric's own in-house factory boasts some of the world's leading board mounting technologies, such as soldering performed in an oxygen-free furnace, which is located in a clean room featuring the same level of cleanliness found in a hospital operating room. Technologies cultivated in the production of electronic circuit boards for audio, medical, aerospace, and defense industries support the high production quality of the Master Sound Discrete DAC.
The ESOTERIC-HCLD output buffer amplifier boasts an amazingly high-speed slew rate (response speed) of 2,000V/µs. Current transmission and speed—the most important factors for an analog output circuit—have been pushed to the limits to reproduce the reality of music with a dynamic range that is truly breathtaking.
In addition to its XLR and RCA amplifier line connections, the K1X also features the ES-LINK Analog method of current transmission. Taking full advantage of the HCLD buffer circuit's powerful current supply makes the K1X less susceptible to the detrimental effects of impedance on the signal route. It also enables a more powerful signal transmission, fully driving connected compatible devices.
A high-accuracy clock circuit is the key to achieving superlative sound quality in digital audio playback. A top-quality clock circuit is truly the heart of any high-end digital player. The K1X now features the superb Grandioso Custom VCXO II clock originally developed for the Grandioso P1X/D1X. This clock was developed with carefully considered modifications made to the earlier model's internal circuitry and components to more vividly emphasize the K1X's ultimate audio quality while further boasting remarkably low phase noise and excellent center precision (±0.5ppm).
The K1X can also be connected to the Grandioso G1 Master Clock Generator for playback, with internal circuits synchronized to a 10MHz clock that features the same high precision as an atomic clock, enabling further pursuit of high sound quality.
D/A Converter Functions for External Input, D/D Conversion, MQA Support & USB Input
In addition to its coaxial and optical inputs, the K1X also features a USB type-B port that is compatible with both 22.5MHz DSD and 768kHz/32-bit PCM asynchronous signal transmission, thus enabling it to also be used as a stand-alone D/A converter. Functions are provided for upsampling a PCM digital signal to 2/4/8/16× (max. 768kHz) and converting PCM to DSD. The K1X is also compatible with MQA-CD decode playback and full MQA decoding for playback of various digital inputs including USB.*
*MQA certification is pending as of August, 2019, and support is scheduled to be added via a software update as soon as certification is complete.
(Source: Esoteric Japan)
fully connected layer 在 コバにゃんチャンネル Youtube 的最佳解答
fully connected layer 在 大象中醫 Youtube 的最佳解答
fully connected layer 在 大象中醫 Youtube 的最讚貼文
fully connected layer 在 What do the fully connected layers do in CNNs? - Cross ... 的推薦與評價
The output from the convolutional layers represents high-level features in the data. While that output could be flattened and connected to the output layer, ... ... <看更多>
fully connected layer 在 How to implement a neural network with a not-fully-connected ... 的推薦與評價
... <看更多>
fully connected layer 在 CS231n: Convolutional Neural Networks (CNNs / ConvNets) 的推薦與評價
Neurons in a fully connected layer have full connections to all activations in the previous layer, as seen in regular Neural Networks. Their activations can ... ... <看更多>