Sunday, June 21, 2015
HTM does not optimize for profit (post to google video)
Jeff mentions there are a whole lot of "motor" connections and other connections going down the cortical column that they do not understand. The HTM is probably about 30% effective in mimicking the cortical column, and that's just on a macro-qualitative basis, not counting the extent and details of what's missing. For example, the HTM using only AND and OR operations whereas dendrites do a lot more. That's is not a qualitative difference because not-AND operations by themselves can be turing complete (a computer can be made exclusively of them...and largely are). But it is there is a mind-boggling difference in capability. A huge qualitative difference is that neurons seek to efficiently use food energy and structural materials at all levels. The HTM is only selecting the number of active synapse connections via permanence values, with only a partial optimization process. The number of segments with AND operations, the number of inputs to the segments, and the threshold of active inputs to trigger a positive AND result are all not being adjusted automatically by the HTM for optimization of resources for optimal pattern recognition. They manually adjust these values based on the application and experimentation. The cortical column does this automatically at many levels. The optimization should cyclically relax conditions and then tighten them in order to discover the best optimization for the data (self-search for Occam's razor). The HTM avoids opportunities to seek occam's razor and in this sense avoids seeking higher intelligence because occam's razor means a compressed model of the data generator (simplist physics laws for example). Optimization of resources for equal or better pattern recognition (aka efficient and high prediction capability) is an implementation of Occam's razor. High intelligence is defined as the greatest profit obtained divided by the storage cost of the algorithm and the computational cost of executing it (these two are physically expressed as mass and energy and have an "exchange rate" conversion factor in getting cost into the same units just as c^2 converts their physical units E=m*c^2 ). These two costs have been formally defined as the length of the algorithm and the execution steps, or some sort of logarithm of them. But basically intelligence is the greatest profit over the widest data set at least cost. "Profit" in the vast majority of AI work is simply pattern recognition, aka prediction, aka classification. This misses something huge: the ability to send a control signal (motor neuron) back into whatever is generating the data to change the data being received in order to increase a desired profit signal (set point) that is separate from mere recognition. You've mention this is a higher-level function of brain, and maybe that's true, but I have to wonder if this is the reason for all the "motor" connections running through the cortical column. The cortical column at all levels may need the control signal as part of its input data, not necessarily needing to know some other part of "itself" was the source, but I doubt the users of the HTM are feeding all available control signals into the HTM . For example, if you make a decision to influence the data generator based on HTM output, then that should be another input stream to the HTM. Getting back to optimization and letting efficient pattern recognition be the only profit desired, the HTM does not seek to reduce storage or computational costs except for letting permanences and segments compete to be active. This is a a big optimization, but it misses the other opportunities I mentioned.