用人脑反应来引导机器学习

原文：Using human brain activity to guide machine learning
作者： Ruth C. Fong（牛津大学），Walter J. Scheirer（哈佛大学）

摘要：近年来，机器学习在各个领域获得了巨大的成功。机器学习向人脑中获得了很多启发，但是却很少有研究关于人脑反应如何引导机器学习。本文演示了一个新的‘neurally-weighted’机器学习范例（paradigm）。该范例通过把参与者fMRI对每一个图片的神经反应灌输到原来的目标识别的训练过程中，使得学习过程越来越趋向人脑。

方法

1. 从fMRI中获得激活的权重

• A. 收集stimulus的激活向量
  • 1386彩色图片。
  • ROIs： EBA，FFA，LO，OFA，PPA，RSC and TOS
  • extrastriate body area (EBA), fusiform face area (FFA), lateral occipital cortex (LO), occiptal face area (OFA), parahippocampal place area (PPA), retrosple- nial cortex (RSC), transverse occipital sulcus (TOS). 1,
• B. 训练分类器
  • SVM
• C. 获取激活权重为到决策线的距离

2. 训练图像分类器

• D. 常用的分类器训练
• E. 利用激活权重重新调整权重

Experimental workflow for biologically-informed machine learning using fMRI data.

结果

通过将传统SVM的hinge loss变为Activity 权重（AWL）的Loss，

image.png

Results showing the effect of conditioning classification models for four visual classes on fMRI data during supervised training.

• C到E分别表示用AWL的EBA，FFA，PPA分别改变了某一种或某几种特定物体的识别。

Statistical influence of each ROI in object classification models using AWL and HOG features

结论

Our results provide strong evidence that information measure directly from the human brain can help a machine learning algorithm make better， more human-like decisions and suggest the potential of a new class of hybrid algorithms.

Our paradigm currently requires access to biological data that corresponds to some training examples. Future work should extend this method to other kinds and combinations of biological data, sensory modalities, and machine learning algorithms. Additionally, our current technique discards meaningful information by using scalar activity weight; thus, further research should investigate the development and incorporation of low-dimensional activity weights.