What's knowledge useful for?

In a previous blog I discussed the use of spaced repetition for retaining knowledge. But why do this? After all why can’t we just Google things now, and with the astonishing developments in AI, just ask AI?

I discussed this somewhat in a blog on England’s educational reforms but I would like to expand here. The capacity to acquire knowledge is a function of the knowledge you already have. The quality of your question, the capability to comprehend the answer, and the likelihood that the new information will be retained are all influenced by your current state of knowledge at the moment you ask the question.

the capability to comprehend the answer

Say you wish to develop some knowledge of neural networks. To even begin you must have some concept of the meaning of “neural” and of “networks”. If you wish to grasp backpropogation you’ll need to understand derivatives and the chain rule. If you wish to read many topics explaining neural networks you’ll need to know mathematical notation and what a vector or a matrix is.

Attempting to learn what a neural network is without some knowledge of the above, is akin to attempting to read a paragraph but having to decode every word along the way. Consider a child who has the skill of word decoding, and here we assume the phoneme in their mind, but lacks the grapheme-phoneme mapping in memory of every word and so to build comprehension of the paragraph step by step first decoding and then re-evaluating this new knowledge in the light of the previously decoded words. They are literally going through the process of learning the requisite knowledge so that they can comprehend the higher order knowledge.

No AI answer or Google result will explain neural networks in any way that will click without the base knowledge.

quality of your question

Suppose you don’t understand something about neural networks, maybe it’s a nagging feeling that after carrying out a training run your not sure it’s a given the minimisation of the loss function would have produced the optimal outcome. You can picture a 3 dimensional function and you can see that it’s possible the process might have caused the final parameter set to reach a trough, but not the lowest trough. You wonder if there is someway neural network training can mitigate this risk. How do you find out what those options are? I would wager if you know the vocabulary of global minima, and local minima your questions will be so much better you are likely to zero in on the answer far more efficiently. But this again is knowledge.

the likelihood that the new information will be retained

New knowledge on arrival to your mind is most likely to stick if it is able to hang or build off of the scaffold of knowledge you already have. Random historical dates are a good example of this. Most mean almost nothing unless the event being described can be mapped in some way to other knowledge. This could be anything. An already known event in a similar time period. Knowledge of an influential figure near the time of the event. Causes or Effects of said event and the dates they took place. Building those connections between new knowledge and knowledge you have is far easier when you have a lot more knowledge to start with. A friend here had a far better intuition for two events happening in the same year via time spent playing games.

The widespread explosion of search engines did not result in a wave of ignoramuses dominating cognitive fields just googling everything as they encountered it. It simply enabled those who knew things to develop more quickly as their capacity to learn was super charged.

One example I’d like to share is that of my own experiences in IT, where at a relatively late stage in life, i.e. after Uni, I had to pick up a fairly large body of very unfamilar knowledge. I would regularly cover a topic, and due to the erratic nature of when that knowledge might get used, then forget it. But when the time came round i knew at least of a concept which could guide my question. What is that number after the forward slash following an IPv4 address? It has something to do with cider…

So are all intellectual pursuits a waste of time if I don’t use spaced repetition to retain the knowledge?

While learning should be a joy in and of itself and continually forgetting things does leave lots of low hanging fruit regularly available to learn I don’t think despair should be the response to a broken anki streak.

The act of learning while developing specific points of knowledge also let’s you know that knowledge is “there”. When you go through a learning experience and then forget you often still retain the awareness that answers to X type of questions are in the Y space of knowledge. I can’t remember which command checks open ports on a remote machine, but I know there is one. I can’t remember how many protons are in beryllium, but I know it is an element and so has a defined proton number. The act of learning leaves residue in your mind, and even where there are gaps you at least