On Fri, Jan 29, 2021, 1:51 AM YKY (Yan King Yin, 甄景贤) < [email protected]> wrote:
> > My proposed model has some important properties: > 1. uses deep learning to learn logic formulas Do you have any experiments confirming this? It wasn't clear from your paper how to achieve this. The lack of an efficient learning algorithm has been the one most important obstacle to symbolic AI in spite of decades of research. I realize that humans are able to manually encode knowledge into structured formats like first order logic and it's extensions like CYCL and probabilistic logic. But this ability can't be central to learning in humans because it comes after the knowledge is learned. You don't have to know the difference between a noun and a verb to form grammatically correct sentences. The 3 major obstacles to AGI, in my opinion: 1. Hardware. After decades of research, the best known solutions to vision, language, and robotics use neural networks. A human brain sized network with 6 x 10^14 synapses and 20 ms activation time requires 30 petaflops and a petabyte of memory. Reducing power consumption from a few megawatts to 20 watts (what the brain uses) can't be done by shrinking transistors. Feature sizes are already of the order of the spacing between dopant atoms in silicon, 11 nm at 1 part per million. To reduce power further you need a whole new computing technology based on moving atoms or molecules instead of electrons. 2. Software. The human body has information content of 5 x 10^9 bits, equivalent to 300 million lines of code (based on my compression tests of the human genome and large software projects). That's doable at a cost of USD $30 billion. There isn't a practical alternative to writing the code, as human evolution cost 10^46 DNA copy operations of 10^37 bits over 3 billion years. 3. Knowledge collection. Human long term memory is only 10^9 bits according to Landauer's recall tests of pictures and words. About 99% of this is shared or online (based on the US labor department estimate that it costs 1% of lifetime earnings to replace an employee). The remaining 10^7 bits takes several months to collect through speech and writing, assuming you are willing to make your life memories public so you don't have to answer the same questions over and over. Repeating for 8 billion people (assuming the obvious application of automating labor worldwide) will cost several months global GDP or about $50 trillion assuming unrestricted data sharing. An efficient unsupervised learning algorithm for structured knowledge could greatly reduce at least the hardware cost. But given the long history of failure dating back to the 1950s, I'm skeptical. ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T54594b98b5b98f83-M0ebc23e306b1e66fe98b5a38 Delivery options: https://agi.topicbox.com/groups/agi/subscription
