In order to supply the seemingly endless demand for data labels to train ML algorithms, there has been a rise in AI software products designed to label raw data. In this article, we look at the difficulties facing the neural networks off which the models work, and how a system of HUMAN review can currently offer higher quality answers with less chance of attack, while also cooperating with the neural networks to improve ML algorithms and, therefore, the machines’s capacity to label.
Multimodal neurons occur in humans as the activation of one brain neuron to a certain thing across a multitude of classes. Famously, the ‘Halle Berry’ neuron activates in the human brain in the same manner regardless of whether the signifier is a photograph, a sketch, or the words ‘Halle Berry’.
Open AI discovered the same ‘multimodal neuron’ function in their vast neural network, CLIP.
There is, however, still a problem. For all the advancement this demonstrates, the neural network labelling software is incredibly vulnerable to simple attacks, as demonstrated below. It is important to note that CLIP is simply used for research purposes at this stage, yet the problem is typical of such computer-labelling softwares.
As Open AI explain in their post:
“By exploiting the model’s ability to read text robustly, we find that even photographs of hand-written text can often fool the model.”
The problem is: there has been an artificial abstraction of a complex human neurological function to the words ‘multimodal neuron’, which have then themselves been abstracted and relaid over a related, but not parallel, occurrence in artificial intelligence. It is undoubtedly worthwhile for ML to take inspiration from neuroscience; after all, AI is trying to reliably accomplish what a human brain does with ease. However, the ease with which such a system can be attacked is a reminder that we would do well not to overestimate neural networks too quickly and, meanwhile, to appreciate the value of human intuition.
While, one day, such simple attacks could be solved, there would no doubt be other difficulties, and other kinds of attack. The problem is in trying to apply linear abstract connections to how the human mind works.
This is because human derivation of essence is an innate quality of consciousness beyond the mere linearity of the languaging computers depend upon. It is not because a human knows the word dog, or the word apple, that it sees a dog or an apple; even an untrained mind that does not know how to read or write, and does not know those words, would understand the essence of the things to be different. It would know by simple intuition that there are two separate things in a picture, one which we distinguish as an apple and another as a post-it note, words which really have no bearing on the nature of the thing, but upon which, nevertheless, computer models depend.
We have no doubt that there will be an improvement in the computer’s ability to derive meaning. An increase in data and, importantly, an improvement in the way the algorithms are structured, will facilitate this improvement. It is vital to note, meanwhile:
There is no reason why a HUMAN labelling system could not provide the necessary human labels to improve computers’ capacity to do that very job.
No. While HUMAN technology currently utilises the world’s largest labor pools, the software could be applied on a human/machine-to-machine basis. A machine could know the data it needs, and request it; whether the request (in the form of a question) is sent to a human, as is currently the case, or to a sophisticated software labelling product, as may well be the case down the line, is irrelevant. The HUMAN Protocol will continue to power the creation of the marketplaces where data solutions are traded.
The HUMAN Protocol Foundation makes no representation, warranty, or undertaking, express or implied, as to the accuracy, reliability, completeness, or reasonableness of the information contained here. Any assumptions, opinions, and estimations expressed constitute the HUMAN Protocol Foundation’s judgment as of the time of publishing and are subject to change without notice. Any projection contained within the information presented here is based on a number of assumptions, and there can be no guarantee that any projected outcomes will be achieved.