Artificial intelligence could produce a human mind

Artificial Intelligence: What AI Teaches About Human Intelligence

In contrast to this, the human brain also has the ability to build mental models from our experience and thus systematically draw far-reaching conclusions beyond the current sensory impressions that even enable explanations. One form of explanatory structure is also known as a generative model. It is based on a multitude of experiences, and with its help our brain generates an internal hypothesis from a certain sensory perception, for example about the direction in which the car is currently driving and which other objects it is covering. Such hypotheses are much more robust than the classification-oriented systems on which deep learning is based. A car is never misclassified as an ostrich by humans because we have various and more general modeling strategies available to prevent such gross errors.

2. AI predictions with deep learning are difficult to explain

AI systems learn to make correct predictions by adjusting the set screws - that is, the weights with which they process information. This enables them to recognize patterns in the data. The resulting statistical rules are distributed over the entire network; we can hardly or not at all understand the rules and certainly not adapt them directly. The system is not trained to make comprehensible predictions, even if newer approaches are working on it: You can program the AI ​​system to make the relevant pixel patterns visible on which a classification is based. But these pixel patterns can seldom be assigned to a trait in the external world that we can understand. An example from face recognition: Do the pixel patterns correspond to an abstract line face or rather to another systematic contrast pattern that differs from our feature classes? That remains open.

So far, AI systems have been based on learned statistical regularities without recording them on a meta level and making them accessible in such a way that we could understand them. Even if an AI system learned this, it still lacks an understanding of relevant causal relationships. Because it only depicts the relationships between input and output data, but not cause and effect. Therefore, so far, artificial neural networks have not been able to develop simple world knowledge themselves.

3. Artificial intelligence is less flexible

Deep learning is only based on learning algorithms that evaluate based on similarity. We humans have many different learning strategies at our disposal. We can learn from a single experience, such as burning our fingers. We learn by observing others or mimicking their behavior. We can immediately record essential characteristics, for example cause-effect relationships, and in this way predict consequences and plan actions. We can ignore superficial similarities, such as that between dolphins and fish. Our theoretical knowledge saves us from making a mistake: the dolphin is not a fish, but a mammal. So we combine different learning strategies and knowledge to avoid mistakes.

A characteristic of human cognition is that on the one hand we can adapt to new situations, on the other hand we can evaluate the same situation in one way or another. Cognitive flexibility - and not high cognitive performance - is a central characteristic of human intelligence. The basic idea that artificial intelligence simply has to imitate individual human performance in order to reproduce and understand human cognition seems naive. In order for this to be successful, it has to integrate various forms of learning and interactions between individual learning modules into its network architecture.

The three levels of intelligent systems

Are there absolutely no similarities between human and artificial intelligence? The British neuroscientist David Marr describes intelligent systems on three levels. On the first computational level it is about what information processing can do, for example recognizing an object based on certain characteristics. Here, fundamental findings of AI can be transferred to human cognition, because the general principles of information processing apply to all systems, whether humans, animals or computers.