How do we learn?

I just read the following article:

> A new model of learning centers on bursts of neural activity that act as teaching signals — approximating backpropagation, the algorithm behind learning in AI.


    https://www.quantamagazine.org/brain-bursts-can-mimic-famous-ai-learning-strategy-20211018/ <https://www.quantamagazine.org/brain-bursts-can-mimic-famous-ai-learning-strategy-20211018/> 

This seems promising and others have noted:

> The brain is the perfect place to look for inspiration to develop more efficient neural networks. The inner workings of our synapses and neurons provide a glimpse at what the future of deep learning might look like.


From a paper that surveys potential approaches:

https://www.researchgate.net/publication/354889734_Training_Spiking_Neural_Networks_Using_Lessons_From_Deep_Learning/link/6152e7c1d2ebba7be752a26c/download <https://www.researchgate.net/publication/354889734_Training_Spiking_Neural_Networks_Using_Lessons_From_Deep_Learning/link/6152e7c1d2ebba7be752a26c/download> 

The premise of back propagation is that error signals can be used to progressively improve the weights in multi-layer neural networks. However, children learn to recognise dogs and cats without being and presented with lots and lots of classified examples. Where does the error signal come from?

There is a lot of interest in self-supervised learning, however, this often involves techniques such as predicting masked items in a sequence, where the software decides what to mask and knows what the masked items are, e.g. words from documents found on the web.

Given lots of photos of dogs and cats, where only a few are classified, we need to find ways to generalise beyond the statistics in the classified examples. Very young children can make mistakes, but rapidly improve when their parents or carers correct them in what is essentially a single-shot process. That is qualitatively different from back propagation where learning takes place gradually over lots of examples.

My guess is that this relates to a notion of salience and constructed models rather than the black-box model of deep learning.

For an amusing take on that, see: “Motorist fined after CCTV confuses his number plate with woman’s T-shirt” 

https://www.theguardian.com/uk-news/2021/oct/18/motorist-fined-number-plate-t-shirt <https://www.theguardian.com/uk-news/2021/oct/18/motorist-fined-number-plate-t-shirt> 

The street camera spotted what it thought was a car number plate in a bus lane, but in fact was a woman with some text across her clothing.  This is an example where the deep learning algorithm was trained to recognise letters and digits without any understanding of salient details. In this case, a simple fix would be to combine two deep learning solutions one to recognise cars and another for number plates.  Bath and North East Somerset Council are operating an AI system that is clearly inadequate in respect to automated fines.

Other examples for vehicle recognition are good at recognising the presence of a vehicle, but lack any understanding of the components that make up the vehicle, e.g. wheels, headlights, windows, side mirrors, model of car, etc.

We can’t afford to ignore salience and part-whole relationships in scene understanding and I think we need AI regulations that prohibit deployment of dumb deep learning!

What do you think?

See also: https://ec.europa.eu/commission/presscorner/detail/en/IP_21_1682 <https://ec.europa.eu/commission/presscorner/detail/en/IP_21_1682> 

p.s. Horizon Europe is seeking proposals on improving AI with a submission deadline next April for work on a new AI-enabled Cloud-edge framework (Cognitive Cloud). Is anyone interested in collaborating on that?  See:

https://ec.europa.eu/info/funding-tenders/opportunities/portal/screen/opportunities/topic-details/horizon-cl4-2022-data-01-02 <https://ec.europa.eu/info/funding-tenders/opportunities/portal/screen/opportunities/topic-details/horizon-cl4-2022-data-01-02> 

Dave Raggett <dsr@w3.org> http://www.w3.org/People/Raggett
W3C Data Activity Lead & W3C champion for the Web of things