Why it’s unreasonable to use word Learning in relation to AI_post 29

Company

Company

Why it’s unreasonable to use word Learning in relation to AI

Memory is an integral part of computer processes and computer hardware. In the classical von Neumann architecture, the computation and memory blocks are separated. This creates a bottleneck effect as you have to store data in one place and perform computation in another, wasting time and resources each time. To create the next generation of computer hardware, we need an approach to process architecture that could help avoid such waste.

01

02

There is no uniform opinion among scientists concerning the structure and mechanisms of human memory, but it is commonly understood to be fundamentally different from that of a computer. They tend to refer to short-term and long-term memory or to storage phases: the short-term memory contains a lot of information, but most of it is lost, and the remaining 20-30% of information is retained in long-term memory.

Scientists believe that the original model of molecular consolidation of memory looks like this. Learning leads to protein and RNA synthesis in nerve cells, which occurs within 1-2 hours after learning. Once it is completed, the memory moves into a stable, consolidated form and cannot be disrupted by the influence on the nervous system.

04

Biologists believe that retrieving "information" from memory is always dynamic, not repetitive: it does not follow a single "path" of re-calculated connections, which is unlike the way it is organised in AI models with their weights.

07

06

After being retrieved from long-term memory, the meanings are actively transformed: the categories that have been assigned to them are re-distributed in a new way. At the same time, the "old" memory is reconsolidated: it does not return to the previous variant of organisation, but is built in accordance with the new distribution of categories

The transition from the first phase of memory to the second one is an active process called consolidation. There is no equivalent to this process in computer technology. Consolidation is not a total transfer of everything that has been collected over a period of time into slow (long-term) memory. It is more like extracting meanings and packaging them for long-term storage.

03

This would quite possibly help to overcome the bottleneck. Thinking is not a process that consists of computation and memory, and therefore divisible into computation and memory. Instead, memory is an integral part of the (overall) thinking process along with computation. It is necessary to keep "peeping" at the manifestations of the mechanisms of thinking if we want to obtain data about its laws.

13

Insights, intuition and deep observations can outpace objective scientific research. On the other hand, objective scientific research brings in new material and opens up new possibilities. There is something in human thinking itself, which could be "peeped" at to observe manifestations, work of thinking mechanisms — as a man "peeped" at a bird's wing or a falling apple to make discoveries.

12

But even a hundred years ago Sigmund Freud used his psychoanalysis and his methods of observation to get the same data about memory and build theories about its workings as modern science has done nowadays. Freud gave his theory the name of memory retranscription. The content of memory traces undergoes occasional reorganisation under the influence of fresh circumstances. Memory is set not once, but throughout a sequence of recollection episodes.

10

All of the above represents current data on memory collected and researched by scientists. Memory in a broad sense is both a term, a brain process, and a scientific problem. Problems with its research/understanding/use lead to von Neumann's bottleneck or AI weights, for example. That said, neurobiologists know much more about memory, but this knowledge is not used in computers at all.

09

During the retrieval, brain cells respond by reactivating processes similar or identical to the molecular mechanisms that are needed for memorising new things, namely reconsolidation processes. In other words, the process of remembering can be as demanding as the process of recalling. AI's processes of remembering and retrieving from memory work in a radically different way: once built connections are reproduced over and over again, and reconsolidation does not occur.

08

11

It turns out that with the help of observations, reasoning and intuition Freud, who had no objective scientific research, "got" to the theory saying that memory is periodically reorganised, changed, and this happens not only when new information is received from outside, but also while "receiving" memories — "extracting", "pulling out" meanings.

05

Computer technology has no counterpart for the synthesis and repackaging of meanings for long-term storage, which is the most important part of learning. Therefore, it is unreasonable to use the term "Learning" in relation to AI. More likely, AI "learning" boils down to limitless memorisation and automatic hypothesis generation.