- Neuron
Action potenitals: spiking or firing
- Anatomy
Recordings
invasive(most accurate): it could be single unit or multi-unit
ECoG: very rare
- neuroimaging: MEG/EEG/fMRI(good spatial resolution but bad time accuracy)
Why it works: nice properties that neurons in the same orientation
- Visual system
- Ventral Stream: what is the object
- Dorsal Stream: where is the object
feed-forward, feed-back connection
真正的神经网络可能也是层级的, 一层处理某一项功能, 比如浅层的来识别物体种类, 深层的识别物体位置.
如果要通过计算机实现上述的模型, 会用到一个分层的前馈模型
CNN
简单的cell: 映射
复杂的cell: 类似于卷积
Max Pooling: 池化作用于图像中不重合的区域, 所以pooling之后会变小