Regression(回歸)

3.Optimization（優(yōu)化）

1. step 1：function with unknown

2. step 2: define loss training data

3. step 3: optimization

y = 0.1k+0.97x1 achieves the smallest loss L = 0.48k on data of 2017-2020 (training data)

How about data of 2021 (unseen during training)?

feature（特征）

Linear models are too simple... we need more sophisticated modes.

也許顯示并不是這樣?？赡軙魏瘮?shù)。（非線性關(guān)系）

[圖片上傳失敗...(image-c9f176-1667375487472)]

Linear models have severe limitation. Model bias

red curve = constant + sum of a set of

[圖片上傳失敗...(image-7cb22c-1667375487472)]

藍(lán)色function 激活函數(shù)（sigmoid ReLU）

All Piecewise Linear Curves = constant + sum of a set of 曲折線

Beyond Piecewise Linear?

Approximate continuous curve by a piecewise linear curve.

[圖片上傳失敗...(image-c69ed8-1667375487472)]

How to represent this function?

[圖片上傳失敗...(image-c846f3-1667375487471)]

Sigmoid Function (激活函數(shù))

Hard Sigmoid

[圖片上傳失敗...(image-94f968-1667375487472)]

Different w b c

[圖片上傳失敗...(image-15df29-1667375487472)]

red curve = sum of a set of curve + constant

[圖片上傳失敗...(image-d1449f-1667375487472)]

[圖片上傳失敗...(image-1fe571-1667375487472)]

每一個式子代表不同的function。

New Model: More Features

y = b+wx1

[圖片上傳失敗...(image-7d7880-1667375487472)]

[圖片上傳失敗...(image-b99a92-1667375487472)]

[圖片上傳失敗...(image-b4f1a3-1667375487472)]

[圖片上傳失敗...(image-7fb4a8-1667375487472)]

Function with unknown parameters

[圖片上傳失敗...(image-2ba9d7-1667375487472)]

Back to ML Framework

[圖片上傳失敗...(image-311fad-1667375487472)]

[圖片上傳失敗...(image-d5e7a5-1667375487472)]

Loss:

1. Loss is a function of parameters L

2. Loss means how good a set of values is.

[圖片上傳失敗...(image-22596b-1667375487472)]

[圖片上傳失敗...(image-70f81b-1667375487472)]

Optimization of New MOdle

[圖片上傳失敗...(image-ab3b08-1667375487472)]

[圖片上傳失敗...(image-5c9ac5-1667375487472)]

Sigmoid - ReLU

[圖片上傳失敗...(image-cce81f-1667375487472)]

[圖片上傳失敗...(image-373681-1667375487472)]

Activation function

[圖片上傳失敗...(image-293ec4-1667375487472)]

Experimental Results

[圖片上傳失敗...(image-fcbf30-1667375487472)]

Back to ML Framework

[圖片上傳失敗...(image-b1c049-1667375487472)]

[圖片上傳失敗...(image-61e85b-1667375487472)]

Experimental Results

[圖片上傳失敗...(image-5b9e4b-1667375487471)]

3 Layers

1. Red: real no. of views

2. Blue: estimated no. of views

[圖片上傳失敗...(image-762e3c-1667375487471)]

[圖片上傳失敗...(image-a76acc-1667375487471)]

Neural Network - This mimics human brains...(???)

[圖片上傳失敗...(image-9422cf-1667375487471)]

Hidder layer

[圖片上傳失敗...(image-442c12-1667375487471)]

[圖片上傳失敗...(image-1db809-1667375487471)]

過擬合

[圖片上傳失敗...(image-1697cf-1667375487471)]

Let's predict no. of views today!

• If we want to select a model for predictiong no. of viwes today, which one will you use?

[圖片上傳失敗...(image-f900f2-1667375487471)]

[圖片上傳失敗...(image-7daef8-1667375487471)]

Park et al. created thin films based on oxygen-deficient tungsten oxide that can modulate the visible and infrared spectral ranges by balance the delicate composition of each constituent.

When different bias voltages are applied to the multilayer DAST films, their refractive index change accordingly to achieve selective transmission of electromagnetic waves.

In this work, we report and numerically investigate the novel multilayer film structure for electrochromic windows based on the electro-optic dielectric material of 4-dimethyl-amino-N-methyl-4-stilbazoliumtosylate (DAST) that they can modulate visible and IR light substantially and independently.

Transparent poly-crystalline spinels protect and defend

P9 2021-Pytorch 教學(xué) part 1

[圖片上傳失敗...(image-f92671-1667375487471)]