References
— John Hertz et al. Introduction to The Theory Of Neural Computation (Santa Fe Institute Series) 1st Edition.
— Daniel Kahneman*, Pensar rápido, pensar despacio*, Barcelona, Debate, 2012.
— Rubén Moreno Bote*. ¿Cómo tomamos decisiones?* Bonalletra Alcompas, 2018
6 Human vs machine intelligence: an interdisciplinary discussion
Ramón López de Mántaras
IIIA (Artificial Intelligence Research Institute) of the CSIC (Spanish National Research Council)
The Panel "Human versus Machine Intelligence" addressed several points based on the four presentations that we had right before the panel. Namely those of Joan Serrà on "Unintuitive properties of Deep Neural Networks (DNNs)" (see Section 2), Gustavo Deco on "Whole brain modelling and applications", Karina Vold on "Extended minds and machines" (see Section 4) and "Rubén Moreno-Bote on "Slow and fast biases in decision making" (see Section 5).
The panel started with some initial remarks by the moderator briefly relating the four presentations. The main remark was to point out the very big difference between the computational approaches based on deep neural networks and the real brain. Indeed, from the presentations of Gustavo Deco and Joan Serrà it was pretty clear how complex the structure and the functioning of the brain is and how poor and limited are the artificial neural networks (ANNs) are, including DNNs. The practice in DNNs relies too heavily on "trial and error" and this is why we do not really know why sometimes they work so well whereas sometimes they make so dumb errors. The gap between theory and practice is very large! DNNs lack explanatory capabilities and they suffer from what is known as "catastrophic forgetting" which means that they forget the task they have just learned as soon as they are trained to perform a new task. This last fact in itself is an indicator of the big difference that exists nowadays between human and machine intelligence.
The presentation of Gustavo addressed the issue that the brain at "rest" (without stimuli) gives an output. So, the "resting" brain is in fact not resting and displays spatial patterns of correlated activity between different areas of the brain. In summary, from his talk, and the subsequent debate in the panel, we could see that the computational modelling of the whole brain dynamics is extremely difficult and extremely far away from the current computational models used in AI.
From the talk of Karina Vold and the panel debate we could see other clear differences between human and machine intelligence, namely: consciousness, evolutionary history, embodiment, situated cognition (as part of cognitive extension), and very importantly social intelligence.
Indeed, we, humans, are social agents and this fact obviously "shapes" and extends our intelligence. Such fundamental differences raise serious questions regarding whether the goal of building humanlike intelligence is possible. Another interesting aspect that was raised and discussed is that the technology we use can become functionally integrated into our biological cognitive capacities such that the tools become part of our minds on a pair with our brains (in the sense of what Clark and Chalmers call "extended minds") and how this can affect human intelligence and our cognitive capacities. Another related question is how we can prevent humans from being manipulated and if we should relinquish control to machines.
Finally, the presentation of Ruben Moreno-Diaz was about decision making in animals and machines and particularly on the presence of slow and fast biases in decision making due to previously existing preferences and information (slow versus fast biases distinction is related to how long before the previous preference was chosen or how long before we had the piece of information that affected our choice). Given that we almost never start from an indifference state. One important question is whether we can avoid biases. The answer seems to be no in general. However, the biases can vary in strength and can be partially controlled. The strength of the biases seems to be related to how artificial the task -that is the object of decision- is, that is how much it deviates from the
natural setting it was designed to operate. One claim of this work is that thanks to the presence of biases we can better compare human with machine performance in situations where the experimental conditions deviate from natural environments.
As a final wrap up summary we could say that the state of the art of machine intelligence is still extremely far from human intelligence and it is very controversial whether, in spite of the recent results based on deep learning, there has been real scientific progress towards the extremely ambitious goal of achieving humanlike AI. Another relevant question is: Do we really need it?