Hacker News new | past | comments | ask | show | jobs | submit login
Dendritic Computing: Branching Deeper into Machine Learning (sciencedirect.com)
45 points by rbanffy 3 months ago | hide | past | favorite | 5 comments

I'm glad this is finally being taken more seriously. I think machine learning is doomed to hit a plateau until we switch over to something like an oscillatory neural net. We already know that human visual perception works somewhat like a fourier transform optimized for projection from 3d to 2d [1,2]. There's evidence that consciousness itself happens at 40hz [3]. It is already known that from signal feedback loops you can get complex behavior like fractals [4] and perform sophisticated analysis like autocorrelation [5]. In the old days before computers we knew how to rig simple analog circuitry to do seemingly complex tasks like target seeking missiles [6]. Analog signals also solve the problem of connectivity in your neural network. It is known that certain neuron structures can act like band pass filters, which means the same neuron pathways can be used to send different signals just like old fashion telephones used frequency multiplexing [7]. Analog models of the brain give you wave superposition for free, so concepts that are composed of two or more simpler concepts emerge for free as a beat frequency with no additional training!

I postulate the set of tasks which are still human-easy, machine-hard have elegant solutions in analog signal processing. The reason they seem hard to solve is because we've been using the wrong tools and compensating with brute force (loads of training data and compute time). Its like taking a hammer to a screw. It sort of works on the small screws, but until someone notices the screwdriver we can only get so far.

[1]https://royalsocietypublishing.org/doi/pdf/10.1098/rstb.2000... [2]https://www.researchgate.net/publication/275770335_The_Confo... [3]https://citeseerx.ist.psu.edu/viewdoc/summary?doi= [4]https://www.youtube.com/watch?v=Pv26QAOcb6Q [5]https://en.wikipedia.org/wiki/Autocorrelation [6]https://en.wikipedia.org/wiki/Contrast_seeker [7]https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3214667/

As far as I can see this paper seems to be behind a paywall. Why is this posted here?

Yea, I checked sci-hub.se, and that DOI is not available (because it's new).

Let's hope someone with a university subscription will post a link to the PDF.

I even checked that, my university has no subscription.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact