Design Choice for Mapping → Colors Rods and Cones Rods 막대세포 active at low light levels one wavelength sensitivity function 100 million rod receptors Cone...

1. Attention Bottleneck & Solutions Attention bottleneck $\hookrightarrow$ computationally expensive. masked self-attention & autoregress...

1. Introduction RNN and LSTM era Fixed size context vector forced to summarize long sequences. Vanishing/Exploding gradient. ...

1. Attention Attention $\hookrightarrow$ Iterative process. Backprop trainable (differentiable). Ex) NLP: ...

1. RNN (Recurrent Neural Networks) RNN $\hookrightarrow$ A family of neural networks for processing sequential data. $x^{(1)}, \dots, x^{(\tau)}$...

1. CNN (Convolutional Neural Networks) CNN Neural net uses convolution in their layers. $\hookrightarrow$ Time-series, image, video. ...

1. Regularization Noise, … we cannot model $\hookrightarrow$ overfitting, less generalization Overfitting $\Rightarrow$ fitting the data more. $E_{train}$ alone is not a good guide. ...

1. Gradient-based Optimization (1) Slope 기반 (Slope-based) Gradient 모으기 (Collect gradients) (vector) $\rightarrow$ Gradient Descent (2) Curvature 기반 (Curvature-based) Hessian (matrix) $\r...

Gestalt Laws the whole is different from the sum of its part leaving a set of descriptive principles, without a model of perceptual processing. The Principle of Pragnanz Fundamental pri...

Why marks and channels? Every complex visual idiom can be broken into two orthogonal(?) components → Marks and Channels Marks Basic graphical element. Classified with their spatial dimensi...