# Feature-Selection-Based Transfer Learning for Intracortical Brain–Machine Interface Decoding

###### tags: `論文摘要` `穩定性` `解碼器` `特徵選擇`

- Link: https://doi.org/10.1109/TNSRE.2020.3034234
- MLA: Zhang, Peng, et al. "Feature-Selection-Based Transfer Learning for Intracortical Brain–Machine Interface Decoding." IEEE Transactions on Neural Systems and Rehabilitation Engineering 29 (2020): 60-73.
- 年分: 2020
- 期刊: IEEE Transactions on Neural Systems and Rehabilitation Engineering 29

# 目標

透過feature selection降低calibration所需要的時間

# 背景

## finetune 原因

現階段的BCI decoder為了應付neural activity發生改變，因此會使用每天的一小部分資料做finetune。

## channel 數量重要性

對BCI decoder而言，訊號所含有的資訊量很重要，因此大多希望能夠取得越多的channel數月好。而實際上，隨著技術進展，可記錄的channel數量也是往越來越多的趨勢邁進。

這反而造成資料量的上升，直接影響到finetune以及訓練的時間。因此作者希望能夠透過選擇有用的特徵的方式，再不影響decoder performance的前提下，減少輸入資料的特徵數量，以加快每天的校正時間。

## 特徵選擇

大致上可以分為以下3種

### filter methods

是在資料前處理階段使用的方法

常見的方式是透過Distance measures、Information or uncertainty measures、Dependency measures...等方式進行

### wrapper methods

使用遍歷的方式從所有組合中挑選對結果有正向影響的組合

通常是用進化式算法進行

### embedded methods

The embedded methods combine the characteristics of both the filter methods and the wrapper methods and incorporate the feature selection as part of the classifier training. There are mainly three types of embedded methods, the pruning
methods, the model methods with a build-in mechanism and the regularization model methods.

## 在侵入式腦機介面中的使用

在iBCI之中通常使用neural selection來代替feature selection，並假設神經元的特徵是互相有關係的

Recently, several studies have provided tools for finding important neurons in iBMIs. Justin et al.[15] proposed three methods for quantitatively rating the importance of neurons: single neuron correlation analysis, sensitivity analysis using
a vector linear model, and a model-independent cellular directional tuning analysis. Remy et al.[33] used the individual removal error to measure each neurons contribution to the overall control. Xu et al.[16] proposed a local-learning-based
method to rank neurons by maximizing the distance between the local neuronal patterns while deploying L1 norm regularization. More related studies are discussed in the review[34]. Many neuron selection methods have been proposed; however, these studies mainly focus on neuron selection, and ignore feature selection; they also disregard the problems in decoder calibration.