engleski

BirdbrAIn Genius – the Reach and limitations of Machine Intelligence

Creating machines that think and act like human beings has puzzled scholars from the dawn of time. While pondering on this concept, ancient Greek philosophers have established the rules that govern correct thought, consequently laying out the foundation for contemporary theorem provers and deduction systems. For quite some time, we believed that computation could be used to mimic reasoning, which in turn would lead to understanding the processes influencing behavioural patterns –it turned out, however, that it is difficult to encode rational thought. On the other hand, mimicking the thought processes going on in our brains, by reverse engineering of the brain, has proven to be infeasible, albeit it led to advances in some other fields of research. Regardless of the approach used for modelling the thought process behind decision making, working on machines that act like people was more or less abandoned because performing well in the imitation game (e.g. the Turing test) did not help in understanding human intelligence. Noticeable advances in artificial intelligence (AI) were reported only when the core focus shifted towards rational acting, disregarding the gist of previous approaches. Nowadays, AI deals with the concept of creating machines thinking and acting like human beings in a rational sense, i.e. agents behaving optimally. Optimal behaviour can be learned (taught) using state-space search algorithms and self-play. Whereas rational acting by an autonomous agent can be considered solved for smaller scale problems in a simulated environment (games and such), the same cannot be stated for large scale problems, those involving uncertainty in a dynamic, ever- changing, environment (i.e., the real world). Here, an agent’s performance largely depends on its ability to learn quickly, and from fewer examples – which can be helped by embedding the agents percepts with machine learning (abstraction of the state space) and enhancing their search strategies using q-learning. Recent advances in machine learning provided an end-to-end modelling framework for learning adequate feature embeddings, via stacked representations, directly from data (i.e. deep learning), which has proven to be rather useful for dealing with highly-nonlinear problems (e.g. those related to sound or vision). This, in turn, resulted in significant improvements in modelling numerous complex problems, previously considered infeasible for real-world applications (because of low fidelity), e.g. image to image translation, translation between written or spoken languages, and image inpainting, just to name a few. The same technology also triggered the development of techniques mimicking human abilities or appearance, e.g. artistic style transfer, speech synthesis, and so on. Although the aforementioned technologies can be utilised for building autonomous agents surpassing human experts in performing numerous highly specialised menial tasks, their intelligence (optimal rationality) is strictly limited to the task at hand, and they can often be easily fooled into suboptimal acting.

artificial intelligence ; rational acting ; game playing

nije evidentirano