Re: unboxed structure fields (or class slots)

> On 14 Jan 2026, at 20:41, David McClain (as dbm at refined-audiometrics dot com) <[email protected]> wrote:
> 
> I don’t see a path to feedback training of Perceptron networks, with any layer depth.

Well, that has been happening a lot in the last 15 or so years, and as you can judge by the current LLM boom, not necessarily unsuccessfully.

> Training is by back propagation of training answers versus training inputs,

No, this is not how it works. You do not propagate answers back to the input: 

1. you propagate inputs forward first, 
2. measure the error between the expected output and the result that the network produced, 
3. and then feed the error back to the inputs, 
4. adjusting the weights in the process. 

You typically run a batch of several thousand inputs first to obtain the average of the aggregate error, then scale it down further by something called learning rate, calculate the error delta at each input and adjust the weight accordingly.

> and requires gobs of iterations,

Yes, this is correct. You operate on the aggregate error, not error from each individual sample and each batch mixed from random inputs adjusts the weights by a minuscule amount.

> since the training weight of any one iteration remains very small.

You could make the delta that is used to adjust the weight larger but it makes the network unstable, it overshoots the local maxima/minima. So you have to get by with using only small adjustment. Not unlike writing lisp code: write a bit, check it out in the listener. Make a change or two, run it in the listener again. An iterative process of discovering an error, and correcting it.

> I have always seen Perceptron NNs as very weak solvers/classifiers.

They are not as good as humans, but there have been giant leaps of progress in the recent years. China has built surveillance  systems that can identify individuals from crowds in real-time, track people's movements across different locations, and build profiles of their activities and associations. The technology works even with partial face visibility and can function in various lighting conditions.

> Each Perceptron becomes gradually trained to see a hyperplane partitioning its space into two subspaces.

There are several ways to describe it:

1. A point in a hyperspace.
2. A vector extending from the hyperspace origin to the point.
3. A volume in hyperspace.
4. Your idea of two subspaces, which is essentially the volume and (hyperspace - volume).

> Overlapping (conjuntive / disjunctive ) combinations of these subspaces gradually isolate useful regions of information.

Overlapping is a measure of similarity of certain concepts, but in this case it would be more beneficial to think of a concept as a constellation of points (model N1 from above), which form something called clusters. When an input point lands within one of those constellations, we have a successful classification. We can also measure similarity of the point to other constellations/concepts by measuring the distance from this input point to the centres of the clusters we are interested in.

> But this iterative training is very inefficient

And very expensive in case of modern commercial LLMs.

> and would essentially ignore any single correction fed back into the system.

Hmmm, no, this is exactly how the error backpropagation algorithm works, feeding corrections back into the system during the training phase.

>  So combining a reasoning engine with a NN leaves you at the mercy of the initial training of the NN.

Yes, it is the same as a rule based engine leaving you at the mercy of the algorithm of the expert system written by someone else, as well as the data it contains.

> The NN itself cannot learn from the output of the combined system.

Not in principle, however neural networks do operate in two distinct phases:

1. Training.
2. Inference.

Those typically do not overlap, you can think of training as the sleep phase in humans, where we consolidate the information the system received during the day. However it does not necessarily means that the error data generated from the inference activity cannot be used. This is how LLM companies improve their offering, they gather feedback from user interaction with the system, and use that data to tune the weights further by new training bouts.