[ 人工智能 ]簡介下現代AI

154 回覆
180 Like 4 Dislike
2017-12-01 19:09:55
有無非image processing 類既deep learning 講解下?
定係其實都只係當d data 係一幅n dimension 既graph?
2017-12-01 19:28:35
識好多concept野, 但tensorflow認真難用
而家就好似學左好多內功
冇武功咁
都唔知做咩春.....

pytorch

感覺tensorflow強大D

橫掂都係

你玩得掂TF先得架

唔好聽到google個名就覺得好勁(雖然真係好勁),但真係elephant in a room

pytorch一出就好多人用,因為佢夠簡潔效率又高,development時間都係成本黎架

單計performance pytorch定tf好?
2017-12-01 21:32:09
有無非image processing 類既deep learning 講解下?
定係其實都只係當d data 係一幅n dimension 既graph?

Bayesian deep learning
做probabilistic modeling
2017-12-01 22:04:11
識好多concept野, 但tensorflow認真難用
而家就好似學左好多內功
冇武功咁
都唔知做咩春.....

pytorch

感覺tensorflow強大D

橫掂都係

你玩得掂TF先得架

唔好聽到google個名就覺得好勁(雖然真係好勁),但真係elephant in a room

pytorch一出就好多人用,因為佢夠簡潔效率又高,development時間都係成本黎架

單計performance pytorch定tf好?

pytorch
2017-12-01 22:14:42
有無非image processing 類既deep learning 講解下?
定係其實都只係當d data 係一幅n dimension 既graph?

我諗應該係倒返轉,image classification本身就係將image睇成一個n-dimensional 既data
個model事實上你唔止可以fit個image落去,fit咩都冇問題架喎
反而我想知好似CNN呢d直接inspired by human vision system既model,係image processing以外有d咩應用

CNN係專為影像設計的AI

其實仲有RNN for 聲音/natural language processing(NLP) 處理, 呢個係語言類的ai

運動都係用RNN訓練

人腦/動物腦強大之處就係集成曬以上所有系統仲有機制互相協調

目前最強嘅AI CNN network只需用550萬個conncection就分到數字,大約同一隻烏蠅嘅腦效能相近,但明顯地烏蠅係唔識認數字

用最多connection的CNN而家大約一億五千萬個connection左右,但係呢個network係sequential, 樓上個550萬認字network升階左上n維感知,所以效果比一億個connection仲好好多

人腦嘅neuron connection係呢個最先進的CNN 的2億倍左右
2017-12-01 22:26:25
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM
2017-12-01 22:41:49
Lm
2017-12-01 22:59:15
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout
2017-12-01 23:42:11
2017-12-01 23:47:37
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer
2017-12-02 00:14:35
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight
2017-12-02 00:53:36
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight

姐係CNN 可以做到time series ?
2017-12-02 08:12:15
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight

姐係CNN 可以做到time series ?

CNN做time series classification 係有的,有人拎來做鯨魚聲音分類,亦見有人拎淺層CNN來做dna prediction,但係唔太多人咁做,只係2D array轉返去1D

單純CNN結構應用係algo trading可以拎來搵上升/下降信號,或者估交易信號/regession

反而cnn+rnn hybrid model就理性一d

當係股票market,你有當刻高位,低位,交易量之類,變n 個input x 1D time series的input, 其實就等價於一張圖片,一樣可以做返樓上講的野

用CNN 單做估升跌/regression作用不大,因為time dimension extrapolation 一定效果唔好。有人拎CNN來做signal filter, 過走曬d noise先放入RNN/LSTM, 好似話會準d

做time series 用LSTM為主幹啦,而家做stock prediction好似得60%準確度,對比做lung cancer detection準確度去到99.97%果d強大的classifer, 股票應用真係對AI 無乜貢獻,有d finance professor夠膽死出paper淨係寫用左NN去train唔講個training model係乜野樣,淨係output d kernel俾你睇,然後話自己好準,估都估到係d淺層NN
2017-12-02 12:29:04
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight

姐係CNN 可以做到time series ?

CNN做time series classification 係有的,有人拎來做鯨魚聲音分類,亦見有人拎淺層CNN來做dna prediction,但係唔太多人咁做,只係2D array轉返去1D

單純CNN結構應用係algo trading可以拎來搵上升/下降信號,或者估交易信號/regession

反而cnn+rnn hybrid model就理性一d

當係股票market,你有當刻高位,低位,交易量之類,變n 個input x 1D time series的input, 其實就等價於一張圖片,一樣可以做返樓上講的野

用CNN 單做估升跌/regression作用不大,因為time dimension extrapolation 一定效果唔好。有人拎CNN來做signal filter, 過走曬d noise先放入RNN/LSTM, 好似話會準d

做time series 用LSTM為主幹啦,而家做stock prediction好似得60%準確度,對比做lung cancer detection準確度去到99.97%果d強大的classifer, 股票應用真係對AI 無乜貢獻,有d finance professor夠膽死出paper淨係寫用左NN去train唔講個training model係乜野樣,淨係output d kernel俾你睇,然後話自己好準,估都估到係d淺層NN

我好似睇過你講果篇paper
係咪stanford既?
2017-12-02 12:34:03
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight

姐係CNN 可以做到time series ?

CNN做time series classification 係有的,有人拎來做鯨魚聲音分類,亦見有人拎淺層CNN來做dna prediction,但係唔太多人咁做,只係2D array轉返去1D

單純CNN結構應用係algo trading可以拎來搵上升/下降信號,或者估交易信號/regession

反而cnn+rnn hybrid model就理性一d

當係股票market,你有當刻高位,低位,交易量之類,變n 個input x 1D time series的input, 其實就等價於一張圖片,一樣可以做返樓上講的野

用CNN 單做估升跌/regression作用不大,因為time dimension extrapolation 一定效果唔好。有人拎CNN來做signal filter, 過走曬d noise先放入RNN/LSTM, 好似話會準d

做time series 用LSTM為主幹啦,而家做stock prediction好似得60%準確度,對比做lung cancer detection準確度去到99.97%果d強大的classifer, 股票應用真係對AI 無乜貢獻,有d finance professor夠膽死出paper淨係寫用左NN去train唔講個training model係乜野樣,淨係output d kernel俾你睇,然後話自己好準,估都估到係d淺層NN

Genomic多數d人想知mechanism多 所以少用nn
2017-12-02 12:58:58
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight

姐係CNN 可以做到time series ?

CNN做time series classification 係有的,有人拎來做鯨魚聲音分類,亦見有人拎淺層CNN來做dna prediction,但係唔太多人咁做,只係2D array轉返去1D

單純CNN結構應用係algo trading可以拎來搵上升/下降信號,或者估交易信號/regession

反而cnn+rnn hybrid model就理性一d

當係股票market,你有當刻高位,低位,交易量之類,變n 個input x 1D time series的input, 其實就等價於一張圖片,一樣可以做返樓上講的野

用CNN 單做估升跌/regression作用不大,因為time dimension extrapolation 一定效果唔好。有人拎CNN來做signal filter, 過走曬d noise先放入RNN/LSTM, 好似話會準d

做time series 用LSTM為主幹啦,而家做stock prediction好似得60%準確度,對比做lung cancer detection準確度去到99.97%果d強大的classifer, 股票應用真係對AI 無乜貢獻,有d finance professor夠膽死出paper淨係寫用左NN去train唔講個training model係乜野樣,淨係output d kernel俾你睇,然後話自己好準,估都估到係d淺層NN

我好似睇過你講果篇paper
係咪stanford既?

係standford果篇

似係undergrad既文來,CS人寫
2017-12-02 13:19:45
堅係要留一留名
2017-12-02 13:32:29

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight

姐係CNN 可以做到time series ?

CNN做time series classification 係有的,有人拎來做鯨魚聲音分類,亦見有人拎淺層CNN來做dna prediction,但係唔太多人咁做,只係2D array轉返去1D

單純CNN結構應用係algo trading可以拎來搵上升/下降信號,或者估交易信號/regession

反而cnn+rnn hybrid model就理性一d

當係股票market,你有當刻高位,低位,交易量之類,變n 個input x 1D time series的input, 其實就等價於一張圖片,一樣可以做返樓上講的野

用CNN 單做估升跌/regression作用不大,因為time dimension extrapolation 一定效果唔好。有人拎CNN來做signal filter, 過走曬d noise先放入RNN/LSTM, 好似話會準d

做time series 用LSTM為主幹啦,而家做stock prediction好似得60%準確度,對比做lung cancer detection準確度去到99.97%果d強大的classifer, 股票應用真係對AI 無乜貢獻,有d finance professor夠膽死出paper淨係寫用左NN去train唔講個training model係乜野樣,淨係output d kernel俾你睇,然後話自己好準,估都估到係d淺層NN

我好似睇過你講果篇paper
係咪stanford既?

係standford果篇

似係undergrad既文來,CS人寫

篇文打哂J
實際點做又冇咩點講
2017-12-02 14:51:02
最近學緊cnn, 發現有篇文講到fcn, 全部只用convolutional layer, 做image segmentation更準, 比cnn compute更快, 請問樓主點睇?
2017-12-02 15:32:01
最近學緊cnn, 發現有篇文講到fcn, 全部只用convolutional layer, 做image segmentation更準, 比cnn compute更快, 請問樓主點睇?

resnet係fcnn的一種(唔計最後一層FC layer做classification)

fcnn之前都覺得幾ok, 但hinton最新個capsule network一出,樓上堆野都好似好小兒科
2017-12-03 04:18:11
樓主會唔會講LSTM同dropout rate? Time series 一般都會用LSTM

dropout 新一代network唔會用了,原理係randomly停左某d neuron 唔做update去防止overfitting

去到ResNet 改左用residual,無左dropout機制, 有人話shortcut呢樣野本質上就係0.5機率的dropout

Fc layer都唔dropout 定係冇FC Layer


應該dropout問題仲研究緊,FC layer係last果層做,去到呢個位dropout唔知仲有幾重要

我睇左hinton最新(?)篇capsule net, 好似完全無講dropout,類似行為改左做dynamic routing,呢篇文好重要,出左3星期超多人discuss, 好有可能revolutionize成個CNN體系,我同個fd一邊研究一邊o 曬咀

佢有篇2018年(?)的論文加深左caspule net同改左routing algorithm, 呢兩日得閒d會睇下又有無咩新insight

姐係CNN 可以做到time series ?

CNN做time series classification 係有的,有人拎來做鯨魚聲音分類,亦見有人拎淺層CNN來做dna prediction,但係唔太多人咁做,只係2D array轉返去1D

單純CNN結構應用係algo trading可以拎來搵上升/下降信號,或者估交易信號/regession

反而cnn+rnn hybrid model就理性一d

當係股票market,你有當刻高位,低位,交易量之類,變n 個input x 1D time series的input, 其實就等價於一張圖片,一樣可以做返樓上講的野

用CNN 單做估升跌/regression作用不大,因為time dimension extrapolation 一定效果唔好。有人拎CNN來做signal filter, 過走曬d noise先放入RNN/LSTM, 好似話會準d

做time series 用LSTM為主幹啦,而家做stock prediction好似得60%準確度,對比做lung cancer detection準確度去到99.97%果d強大的classifer, 股票應用真係對AI 無乜貢獻,有d finance professor夠膽死出paper淨係寫用左NN去train唔講個training model係乜野樣,淨係output d kernel俾你睇,然後話自己好準,估都估到係d淺層NN

我就係用lstm 做prediction ,都想睇下可以點樣做好dd
2017-12-04 11:59:30
https://t.me/HKAIG

如果有人想識多d香港玩緊AI既朋友,有個tg group可以俾大家一齊再深入share下
2017-12-04 21:54:54
吹水台自選台熱 門最 新手機台時事台政事台World體育台娛樂台動漫台Apps台遊戲台影視台講故台健康台感情台家庭台潮流台美容台上班台財經台房屋台飲食台旅遊台學術台校園台汽車台音樂台創意台硬件台電器台攝影台玩具台寵物台軟件台活動台電訊台直播台站務台黑 洞