Hello, I’m a computer science student studying artificial intelligence who has become interested in this project. Greetings from far across the sea in Korea.
I was deeply impressed by reading “A Thousand Brains” and, driven by additional curiosity, I’m now reading “A Brief History of Intelligence” listed on Further Reading with great interest.
I have so many questions I’d love to ask, but I’d like to pose just one or two here to further fuel my curiosity. In the book “A Thousand Brains,” there’s content like this. Please note that since this is a re-translation from the Korean translation, the translated content may not be perfectly accurate:
“Until then, only one type of spike was known. It was the (dendritic spike that occurs when enough input is received at synapses closer to the cell body) that starts from the cell body and travels along the axon to other neurons. Around 1990, it was discovered that there are spikes that travel along dendrites (originating from synapses farther away). … AI researchers who don’t know the purpose of dendritic spikes use neurons that lack dendritic spikes in their models. … Any theory or neural network that cannot explain 90% of the synapses that exist in the brain must be wrong.”
I’m curious whether the SNN (Spiking Neural Network) neuron model, known in the field under the keyword “neuromorphic,” can explain the 90% of spikes mentioned here, and whether this Monty codebase is modeling the insights contained in this content.
Additionally, for understanding the peripheral knowledge related to this project, beyond reading the content posted at Welcome to the Thousand Brains Project Documentation!, I would greatly appreciate any small advice you could give to a newbie who wants to join the community as quickly as possible. It would be an invaluable starting point.
Hi @ProtossDragoon , welcome to the forums! That’s great to hear about your growing interest in the Thousand Brains Project.
Re. your question about spiking neurons:
The typical models of spiking neurons are what are called “leaky integrate-and-fire neurons”. These model each neuron as a single point, but with the ability to integrate incoming spikes over time, as well as the “leakage” of the membrane potential to a resting potential.
These do not model dendritic spikes; to do so you typically need what is called a multi-compartment neuron. These are very complex to model, requiring much more compute than a simple point neuron. It is also complex to use these to build a network that can learn something.
Separately from this work, Numenta implemented HTM neurons, which are an abstract way of accounting for dendritic spikes, without closely modelling the physiology of how a membrane potential changes over time. If you are interested in dendritic spikes, I would recommend reading this paper.
In Monty, we do not model these details directly, i.e. we do not currently use HTM neurons. However, Monty’s representations (the hypotheses in the Evidence-based Learning Modules) do represent predictions, integrate information over time, and compete with one-another. In that sense, they model many of the benefits that dendritic spikes bring, without needing to model the biology in detail.
I appreciate its a complex topic, and the above might be confusing if you are new to this, but hopefully that helps.
Re. getting started on the TBP
As well as the link you shared, a great place to start is the videos we’ve posted to YouTube, in particular our Quickstart Playlist. Once you’ve watched those videos, there are many others you can find via the Playlists link.
Thank you so much for sharing such fascinating content. I’m really excited about the prospect of reading through these materials one by one over the next few years and building interesting things with them.