|Questions about HAL|
Why is HAL two-dimensional?While our existing model is two-dimensional, we hope it is obvious how the model scales to higher dimensions. We have even developed three-dimensional, computation-universal reversible cellular automata specifically designed for this application.
To simply address the question of why we are using a two-dimensional model, we are targetting hardware that is currently two-dimensional and we need locality considerations in our model to directly match the locality-considerations of the target hardware.
Developing a three-dimensional model and implementing it on essentially two-dimensional hardware would force components in one of the dimensions to become arbitrarily distant from their "neighbours" as the size of the automata increases. This would defeat the purpose of using a local model in the first place.
If a three-dimensional model were to be used it would make sense to strongly restrict the size of one of the dimensions. As one road-map for the development of silicon-based computational substrates envisages the third dimension becoming employed gradually through a build-up of conducting layers, such a model may even eventually make reasonable sense.
However for the moment, FPGAs are thoroughly two-dimensional. While we look forward to the development of three-dimensional computing substrates as much as anyone, and believe that the use of reversible logic of the type we are advocating will assist the heat dissipation issues raised by such an architecture, until it exists we will continue to work primarily with two-dimensional models.
Why doesn't HAL employ neural nets?Neural nets appear to be in fashion.
While it would be possible to use cellular neural nets as a target substrate for HAL, and it is possible to use HAL as a genetic substrate for breeding neural nets in a separate medium, neural nets do not appear to be appropriate for most of the target problems we're considering. In particular, although the best players of the Game of Go in the world use neural nets, a pure neural net approach seems to have very severe weaknesses when it comes to reading out tactical sequences. We doubt an intelligent Go player design would be purely neural-net based.
Neural nets do not appear to be a particularly good match for the target hardware either - which typically resembles a network of boolean logic gates more than a network of neurons.
HAL constitutes an attempt to develop computational devices by mimicing the fundamental processes of living organisms. It concentrates firmly on the genetics and development of living organisms. From the perspective of the creation of living forms, neural nets don't appear to be relevant in the slightest.
HAL is fundamentally a low-level approach. If HAL finds itself faced with problems for which a connectionist approach is most suited, then we expect it to evolve the relevant apparatus - and expect that the resulting apparatus will prove more capable than if it had been designed by hand.
Evolving a connectionist representation will undoubtedly be a time-consuming process, but it will only need to be performed once - and allowing this aspect of the design to be fine-tuned by evolutionary techniques should hopefully allow HAL to address a wider range of target problems, than if a neural representation was hard-wired into the design.