Machine learning and computational neuronscience intersection

https://www.quora.com/Is-there-strong-research-being-done-at-the-intersection-of-machine-learning-and-computational-neuroscience

Is there strong research being done at the intersection of machine learning and computational neuroscience?

 

Is there strong research being done at the intersection of machine learning and computational neuroscience?

I would say yes. Certainly people in both areas show increasing interests in each other and start to organize conferences and events to encourage sharing the knowledge. For example UC Berkeley will organize the Brain and Computation program that brings many active machine learning researchers and neuroscientists together (see Simons Institute for the Theory of Computing). If you look at those people’s research pages you will realize there are indeed many works on this topic.

One famous work I particularly like is the sparse coding research by Olshausen and Field (1996, 1997). They show that the encoding mechanism of early visual cortex can be hypothesized as a sparse coding algorithm. Artificial “neurons” (basis functions) trained using this algorithm similar functionality to the biological neurons, e.g., edge detection, orientations sensitivity, etc. Following this line, machine learning researchers start to form many biological plausible hypotheses and models.

Due to the popularity of deep learning, neural networks starts to draw more attention. If you look at early researchers of neural networks, such as Yann LeCun, Geoffrey Hinton, and Terry Sejnowski, many of them have very strong background in computational neuroscience. Important algorithms and models, such as backpropagation, convolutional neural network, dropout, are more or less inspired by brain mechanisms. On the other hand, new techniques developed by computer vision, signal processing,

Another line of research is the reinforcement learning, the algorithm that can play Atari games, beat human champions in poker and Go, and many more. Reinforcement learning (as a machine learning research field) itself has come from an large research area of psychology and behavioral neuroscience, known as the operant conditioning. Neuroscientists has been trying to understand how reinforcement learning is done in brain. Important mechanisms such as reward, temporal discounting (how much do I down-weight future reward?), state encoding (where am I in some decision space?). Neurotransmitters such as Dopamine and Serotonic systems seem to be closely related to these mechanism, but we are far from understanding their full functionality.

There are many, many more stories like sparse coding, neural networks, and reinforcement learning. From my personal standpoint, AI & Neuroscience are essentially one field: to build smarter AIs we will need to understand intelligence itself. Certainly AIs and machine learning algorithms do not have to be exactly like biological brains, but brains provide helpful directions and inspirations we are lost in the maze of mathematics.