Re: Intelligence without representation

The idea that intelligence can emerge bottom up is consistent with theories of evolution where neural architectures will be selected that offer better chances for survival and reproduction. This will include ways to speed up learning through embedding prior knowledge, and effective exploitation of past experience.  This further points to the benefits for having a layered approach to representing and processing data, and how we are able to learn to recognise animate vs inanimate objects, and their identity, constituent parts and behaviours, starting from pixel level representations.

Brooks reacted against assumptions of the need for central control and the idea of an agent sitting in the machine and doing the work on the representations, and central control is indeed not needed for simple organisms.  However, such central control is definitely a feature of the human mind where consciousness is associated with sequential rule execution by the cortico-basal ganglia circuit.

Brooks further asserted that intelligent behaviour can be generated without the need for explicit manipulable internal representations.  However, that just begs the question of what that means in respect to different kinds of internal representations.

For the human brain, we can use multiple levels of description, e.g. biochemical interactions in the synapses, associated chemical and electrical gradients, the transmission of pulses along nerve fibres, statistical correlation of pulse rates across bundles of fibres, and its relation to vectors in spaces with high dimensions, and to concepts with sets of properties, rules that operate upon them, goals, tasks, and emotions.

The cortico-cerebellar circuit is perhaps closer to Brook’s subsumption architecture. This circuit provides a means for actions made at a conscious level to be devolved to a separate system of systems. The cerebellum acts a bit like a traffic controller coordinating the activation of many muscles based upon real-time information from the senses relayed via the cortex. This involves many systems acting in parallel using a layered approach to control.

The above is a long way from the mindset of classical AI that Brooks was reacting against, and is grounded on progress on the scientific study of the human mind and behaviour as conducted in the cognitive sciences, rather than on a narrow conception of AI and KR.  Instead of focusing on manual development of knowledge representations, it would be advantageous to look at how these can be learned through interactions in the real world or simulated virtual worlds, drawing inspiration from the cognitive and linguistic stages of development of young human infants.

This is almost certainly the wrong forum for discussing such ideas, but at least I have given you a sense of the approach I am exploring.


> On 23 Nov 2019, at 02:24, Paola Di Maio <paola.dimaio@gmail.com> wrote:
> 
> I think I found the culprit, at least one of the papers responsible for this madness of doing
> AI without KR
> https://web.stanford.edu/class/cs331b/2016/presentations/paper17.pdf <https://web.stanford.edu/class/cs331b/2016/presentations/paper17.pdf>  
> I find the paper very interesting although I disagree
> 
> Do people know of other papers that purport a similar hypothesis (that KR is not indispensable in AI for whatever reason?) 
> thanks a lot
> PDM
> 

Dave Raggett <dsr@w3.org> http://www.w3.org/People/Raggett
W3C Data Activity Lead & W3C champion for the Web of things