In my opinion, AI should me more sophisticated in edge environments, when a new data is coming, the deep neuronal architecture should change his weights entirely, but safely, not just about adding blob data to local databases, the system should interpret the data itself and the dissimulated semantic meanings behind it.
Like the humans, when a student learn by heart some texts or paragraphs, after he sleeps, his brain has assimilated new information, and under the hood he have a personal opinion behind each word he learn. And surely, when the student learn by heart a non-comprehensible text it will be automatically more difficult for him to learn the text by heart.
So edge AI models have to be systems, not just weights, evaluating the safety level of any kind of information, text, ASR outputs, videos, images, and the impact of human alteration of data. So a text data like: "The earth is flat" is an opinionated sentence, it should be interpreted as a logic/opinion data when the system read it, not as a raw data, but on the other hand, a satellite image of the earth is a safe data, a raw data, so even the sentence "The earth is approximately spherical" is a non-safe data, it is a human deduction. So we can define the safety degree of the data as below:
- Raw Data (Images, videos, audios...): it is the most safe represented data.
- Context-aware Human Data (ASR, texts...): because an article of a scientist in a book dedicated about science of earth is more sure than a simple discussion between two humans about the form of the earth.
- Non Context-aware Human Data (ASR, texts...): Just discussions or some texts about the earth's form in a math book where the subject is math and not particularly the earth.
- AI Extracted Data (ASR, texts...): So, AI can extract some data that is not definitely corresponding to the context, and AI can alter data.
- AI Generated Data (ASR, texts...): because the generated data can be generated from non neutral data. And the rules of AI may differ from one to another.
In all of the cases expect the first one (raw data), the data is extrapolated or extracted or even generated from the raw data, so even AI generated data is coming from the raw data, because the raw data was used to train the model to generate an another similar data. So although we go in deep from raw data to AI Generated Data, the data may suffer a lot of modification or interpretation or even deduction from the mathematical model or the human, so in global cases the AI generated data is even more affected because it is trained by raw dataset and human deduction also, it depend on how the human data is present in the training dataset split.
There is also an another factor to take in the mind, it is how the data is collected, so:
- Raw video: it is completely safe because the video encapsulate the entire scene with visual arguments.
- Raw image: it is a bit non safe because image do not take care about the context. For example, a video of a human eating an apple is sufficient to prove that the human eat the apple but an image can be taken before the human enter the kitchen when the apple is on the table, so the image is not sufficient to prove the entire context, but it is sufficient to prove that the apple is red in an instant T.
- Raw audio: it is more safe than ASR extracted data but less precise and less meaningful, so an audio can prove that a person is talking, or prove that a music is played in the room, and so in a context aware encapsulated data.
- Raw text: it is the last type of data we will say about it, this type of data it is more unsafe, because it is the more non natural, it is a deduction of a human in a book, or an AI generated or extracted text, so we can say about it as the more non safe type of data. A raw text is not found in the nature, there is no texts graved on the plants, so when I say Ahmed is eating the apple, the text do not prove anything, it just say it but without arguments. But, when I have a raw video non AI generated of Ahmed eating the apple it is a lot more evident that Ahmed is eating an apple.
- ASR extracted speech (a subclass of text data): So the safety level depend on the ASR quality and precision and the background sounds in the raw audio. It can be classed as an AI extracted data. So we cannot talk about it as a directly founded data, it is the nature of the data, because it is AI extracted, do here the source of the data is raw audio altered by AI model.
So the safety level of a new data should be evaluted by the system depending of two factors:
- the safety degree of the nature of the data : Raw Data (100% safe), Context-aware Human Data (65% safe), Non Context-aware Human Data (45% safe), AI Extracted Data (35% safe), AI Generated Data (20% safe)
- the safety degree of the extraction of the data: Raw video (90% safe), Raw image (65% safe), Raw audio (50% safe), Raw text (20% safe)
We can do the percentage of the percentage, if the safety degree of the extraction of the data is 0% the data have no sense and it is completely unheeded by the system, and even if the safety degree of the nature of the data is 100%.
So we have to do the percentage of the percentage to take look at the importance of the data as evaluated by the system, if the importance of a specific data is 100%, the system do not revoke it at any time, but if the importance percentage is zero percent, so the data is just completely unheeded. If the system storage is too full, the system can suppress the most non important data. It is just an eventuality, but at most the time the probability to the importance of a specific data to be zero is just nearly impossible, it is recommended to save the data.
From this point we have defined the global guideline to establish the Weights Of The Data (WotD), mathematically, it is just a set of ordered pairs of raw data and his importance, we can also add the date of the data so it will be an array of three indexed nested arrays, or even a SQL database.
We have now to determine the processes of selection of the data, we have the data but we now have to interact with the data in a dynamic way where the principal deep neuronal network of the system should to have dynamic weights impacted and altered over the time, more and more data is coming depending on the safety of the data and its repetition, so if I have "Ahmed eat the apple" a hundred of times it do not prove that Ahmed eat the apple even a thousand of times, it is like just one time.
But now, we have a big problem, is how we will do that, and how the system should interpret the data itself and the dissimulated semantic meanings behind it? not just copying the data or training the principal deep neuronal network on the data. So this is the subject of my next article. In conclusion, at this time I write this article, our system is now like a sophisticated composition of a database, and a deep neuronal surely based subsystem that is sorting and putting the incoming data in the database in a sure way, and an another subsystem that have the unique goal of improving in an intelligent way and dynamic way the principal neuronal network with the sorted dataset's information.
Please write me in the comments about your opinion, do not miss that your comment is a human deduction and not a raw data πβπβ
Top comments (0)