# Deep Learning in JS - ashi
###### tags: `jsdc2019`

{%hackmd @JSDC-tw/2019-announcement %}

Train in the sky and run in the browser
## TensorFlow

TF.py | [TF.js](https://www.tensorflow.org/js)
---|----
CPU, GPU, TPU | CPU(Node), GPU(WebGL)
Giant ML ecosystem | not yet

---

## What are models?

> learning rate
> loss landscape -> visualizing
> transfer learning
> Stochastic Gradient Descent (比喻：隨機畫線→趨近一條蛇?)
> Relu Activation Function

> mobilenet：pretrained model
> 用mobilenet的output當作ensemble 的input，train一個smaller network
> Loss function: categoricalCrossentropy →計算[1, 0, 0] 和 [0.4, 0.5, 0.1] 的差異
https://github.com/tensorflow/tfjs-examples/tree/master/webcam-transfer-learning

example model: https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json

convolutions
https://github.com/tensorflow/tfjs-examples/tree/master/visualize-convnet

---

- Tensorflow
- JS
- Python
- Modeling
- [example model](https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json)
- [Loss: 損失函數 (How badly we did)](https://medium.com/@chih.sheng.huang821/%E6%A9%9F%E5%99%A8-%E6%B7%B1%E5%BA%A6%E5%AD%B8%E7%BF%92-%E5%9F%BA%E7%A4%8E%E4%BB%8B%E7%B4%B9-%E6%90%8D%E5%A4%B1%E5%87%BD%E6%95%B8-loss-function-2dcac5ebb6cb)
- Mean Square Error
- [Loss Landscape](https://github.com/tomgoldstein/loss-landscape)
- mobilenet：pretrained model
> 用mobilenet的output當作ensemble 的input，train一個smaller network
- [Gradient Decent](https://www.wikiwand.com/zh-tw/%E6%A2%AF%E5%BA%A6%E4%B8%8B%E9%99%8D%E6%B3%95)
- [stochastic_gradient_descent](https://www.wikiwand.com/en/Stochastic_gradient_descent)
- [ReLu 線性整流函數](https://www.wikiwand.com/zh-tw/线性整流函数)
- [Transfer Training](https://www.wikiwand.com/zh-tw/%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0)
- [tfjs-examples/webcam-transfer-learning](https://github.com/tensorflow/tfjs-examples/tree/master/webcam-transfer-learning)
- [Adam Optimizer](https://www.tensorflow.org/api_docs/python/tf/keras/optimizers/Adam)
- [Cross Entropy](https://medium.com/@chih.sheng.huang821/%E6%A9%9F%E5%99%A8-%E6%B7%B1%E5%BA%A6%E5%AD%B8%E7%BF%92-%E5%9F%BA%E7%A4%8E%E4%BB%8B%E7%B4%B9-%E6%90%8D%E5%A4%B1%E5%87%BD%E6%95%B8-loss-function-2dcac5ebb6cb)
- categoricalCrossentropy → 計算[1, 0, 0] 和 [0.4, 0.5, 0.1] 的差異
- [Convolutions](https://www.wikiwand.com/zh-tw/%E5%8D%B7%E7%A7%AF) 卷積
- [tfjs-examples/visualize-convnet](https://github.com/tensorflow/tfjs-examples/tree/master/visualize-convnet)
- 結語
- ML 可以被用來計算、產生模擬人講話的影片，未來的接收到的訊息是否可以被信任?