Mohamed el Nawawy It is common for ML applications to use quantization for neural net weights. Usually there are steps taken in the training process to compare the accuracy of the inference to the unquantized versions, so minimise the error between the two. Often it is possible to quantize the weights down to 8b, 4b or fewer, with some accuracy tradeoffs. For example, if the ML application is a classifier to recognise an image of a car, the number of misclassifications can be quantified (e.g. a bus is misclassified as a car), and the quantization adjusted to get the performance required. Some AI hardware chips use analog compute techniques too. The problem then is not quantization, but accuracy due to systematic errors or random errors (noise) in the analog implementation. I don't know neural biology, but it is possible that biological neural networks also exhibit errors, not quantization but systematic, or random errors due to noise. The brain could compensate for these errors as it learns.

Agree to guys who shared their knowledge but I just want to clarify, AI models process information defer from humans, It is using mathematical representation, and therefore reducing raw inputs. AI compressed all relevant information in a model to make superior decisions, so image recognition can ignore irrelevant details and the model will only focus to important one, this is for efficiency. But it does not mean that reducing it makes the model less efficient rather it means optimization of the decision. So the question is "Will AI Surpass human cognition ?" well if we define Intelligence as pattern recognition, then the answer is yes but if we define cognition as understanding and creativity then AI needs more improvement.

@Diosdado Maico

Thank you for your valuable contribution and interesting discussion. Optimization involves deriving the best possible output for a given system based on the available resources. As you kindly mentioned, AI uses compressed or reduced form of information than that used by humans, where this reduced form of information is a main resource for AI systems, so my proposed question is whether this limitation in resources for AI systems is an indication of its limited capabilities compared to human’s capabilities

AI as I mention process quantized or reduced information does not mean it reduced its capabilities, rather it reduce to find only the relevant information for the job making it more efficient than human in some areas where pattern recognition is involve and memory processing by mathematical representation. Human brain uses neural networks in parallel way, thus we don't remember every details while AI uses quantized numerical data but this does not mean its capability is less, it does not need huge information, it only needs the right one. As we compare these two, they have their own strength and weaknesses but combining these two, makes more effective and refer this thing as "HYBRID MODEL?

Diosdado Maico

You've clearly discussed your point concerning AI reducing its input to find only the relevant information necessary to achieve its task efficiently. However, what I'm referring to here is the quantization error due to rounding and truncation of the input analog signal while converting it into discrete values, and, the limitations which such approximation of the input data causes to the performance of an AI system, when compared to humans who handle data in its original continuous form.

For the benefit of the reader, we have to digest this, we will ask and define what is Quantization Error? This occur due to rounding and truncation of discrete representation, so it means that the lower the bit (resolution), the higher the error and therefore loss of information which cause the misinterpretation of AI due to missing details. Humans as I said operates in continues signals in the form of sound and light waves and doesn't do rounding errors and doesn't follow quantization due to our brains can infer details that are missing. So therefore the object that we see blurry can still infer details and we can still recognize due to previous knowledge, in this case AI should improve this limitation. At of this moment AI is improving its limitation, with the invention of neuromorphic chips that process like humans do in the future.