What are examples of machine learning techniques (i.e. models, algorithms, etc.) inspired (to different extents) by neuroscience?

Particularly, I'm interested in recent developments, say less than 10 years old, that have their basis in neuroscience to some degree.

  • $\begingroup$ This question is related to Which artificial neural network can mimic biological neurons the most?, but it isn't an exact duplicate, because here the OP is asking about machine learning techniques (so, in this sense, this question is more general) and, in particular, developed in the last years (so, in this sense, more specific). $\endgroup$ – nbro Apr 22 at 13:37

There is a category of neural networks that more closely attempt to mimic biological neural networks by incorporating also time (i.e. not all neurons fire at the same time). They are called spiking neural networks (SNNs) and their name comes from the fact that they use spiking neurons (i.e. neurons that fire discrete signals and affect other neurons at possibly different times).

SNNs are mainly used in neuroscience, and aren't commonly used in machine learning because they currently have some apparent limitations (e.g. non-differentiability, so gradient descent and back-propagation can't be applied, but GD and BP aren't really biologically realistic anyway, although some people already tried to apply GD to SNNs) and their performance isn't still as good as the performance of traditional deep learning models, which make them not so appealing to the deep learning community (which is currently mainly driven by performance and utility). Nevertheless, the performance gap between traditional neural networks and spiking neural networks is decreasing. See Deep Learning in Spiking Neural Networks (2019) by Amirhossein Tavanaei et al. for more details.

There are already commercial implementations of a hardware-accelerated SNNs (e.g. BrainChip provides this service). These hardware-accelerators are often called neuromorphic chips (or processors) and all computing based on SNNs or processors that attempt to implement biological neural networks is known as neuromorphic computing.

There's also the related area called reservoir computing, which studies neural networks (such as liquid-state machines or echo state machines) that make use of reservoirs (which are fixed during learning) to attempt e.g. to improve training efficiency. See An overview of reservoir computing: Theory, applications and implementations (2007) by Benjamin Schrauwen et al. for an overview.

Numenta (and, in particular, Jeff Hawkins, the founder of Numenta and author of an interesting book called On Intelligence) has also been studying neuroscience for a long time in order to develop models and theories of human intelligence. They call their new theory The Thousand Brains Theory of Intelligence, which is inspired by biological grid cells. This is also related and similar to capsule networks (often associated with Hinton).

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.