# 7/8 Paper #8 ## iCaRL: Incremental Classifier and Representation Learning [toc] ---- ## Abstract (2016, Nov 23) New Idea: class-incremental learning. 1. Be trainable from a stream of data on different batch of classes training time. 2. multi-class classifier 3. memory should be bound so far or grow slowly. ---- ## 1st Author  --- ## Briefly illustraction  High Level interpretation: On Classify: $\theta_s$ -> feature extractor $\theta_s: X \rightarrow \mathbb{R}^{d}$ $\theta_o$ -> output layer with sigmoid function ---- Assume we have 6 classes on dataset: We split 6 class into two class each time. After training on 5,6 class: we want to classify $X^{y1}, X^{y2},X^{y3}, X^{y4},X^{y5}, X^{y6}$. ---- 1. sample data on all class. $X = \{X^{y1},X^{y2}, X^{y3}, X^{y4} X^{y5}, X^{y6}\}$ where $X^{y1} = \{ x_{1}^{y1}, x_{2}^{y1}, x_{3}^{y1} \}$ three data point on a set $X_{1}^{y1}$ 2. Forward $X$ to model and get $\theta_o(X)$ and $\theta_s(X)$ 3. Calculate each class mean representaion from each $\theta_s(X^{y_n})$ 4. find the most closed mean label representation to each data and assign $y^{*}$ --- ## Detail ----  ----  ----  ----  ----  --- ## Experiment Setting RESNET = Feature extractor one output layer with sigmoid layer for each classes task ---- hybric setup: hybrid 1: use classification accuracy to calculate accuracy hybrid 2: the exemplars for classification but not add distillation hybrid 3: training use exemplar representation learning and not distillation but testing not use the exemplars for classification different k setup: memory capacity --- ## Result ----  ----  ----  ----  ---- --- <style> img { width: 80%; height: auto; } </style> ---  ----  --- <style> .reveal { font-size: 24px; } img { width: 60%; height: auto; } div { resize: both; } </style> ---