Demis Hassabis - Combining systems neuroscience and machine learning: a new approach to AGI
Neuroscience is rapidly teasing apart the functional roles of the brain's components, and in some cases even the types of algorithms that they use. Machine learning, meanwhile, is producing a growing collection of techniques for specific kinds of problems, but as yet no general purpose algorithm for artificial intelligence. By bringing these two fields together, we can have both a high level architecture for an artificial general intelligence, and working algorithms for implementing many of the required components. In this talk I will outline the case for pursuing this approach, some current work in progress, and some of the challenges we face going forward.
He compares the biological and non-biological approaches.
The non-biological approaches like formal logic, neural networks tend to have several issues from a list of problems like being brittle, not good with uncertainty and ambiguity, time consuming to train... A main example is the CYC project (Lenat). ongoing for 25 years. One million propositions. Adding new assertions can cause a cascade of inconsistencies that can take weeks to resolve if ever.
Bio-inspired- the reverse engineering projects like Blue Brain and Synapse.
Not clear how well this will work and looks like it may take as long as 50 years to achieve good results.
to consider which way to go consider what the searchspace for AGI might look like.
1. Search area possibile - small search space and dense with possibile AGI solutions. then it would not make sense to go to the trouble of reverse engineering the brain
2. Search area possible - large search space and sparse with AGI solutions. Then it makes sense to reverse engineer
Evidence points to second, that it is a large search space with not many AGI solutions.
Evolution which is a search method has produced only one human level intelligence. The large non-biological AGI projects have not been successful and achieved results that are far from human level AGI.
There is also a range of biological AGI approaches.
From Abstract -------------------**--------------->Biologically copying Cognitive science Blue Brain (Markram) SOAR (Laird/Newell) synapse (Modha) ACT-R (Andersen) Open-cog (Goertzel)** The Middle of the range is systems neuroscience approach is to copy the brains algorithms.
Marr's three levels of analysis
Cog Science --> Computational (what are the goals?) algorithmic (How ? algorithms) Biological --> Implementation (the medium)
There are rapid advances in neuroscience
New experimental techniques (imaging, multi-cellular recording...)
Sophisticated analysis tools (MVPA)
Exponential growth in understanding
finding the nuggets of neuroscience that can be applied to AGI from a lot of information
50,000+ neuroscience papers in 2008
How to sift through all of that to find what is relevant for AGI ?
Actively conduct research and 5+ years of intensive study in both disciplines of AGI and neuroscience.
What use is neuroscience?
It provides direction for AGI where there is no good machine learning alternative
It can give a state of art implementation and a possible AGI component.
Poggio vision system
Simple cells and complex cells. Replicating this algorithm for artificial vision system gives a good implementation.
Place cells and grid cells (graph paper for the mind)
Provides a good approach to navigation
Reinforcement learning and dopamine system can be implemented with math functions.
System neuroscience is a hybrid approach to combine the best of machine learning and neuroscience. Use machine learning when there is a good solution if there is no good solution then push machine learning as far as it can go and search neuroscience for inspiration
there is several gaps that neuroscience can fill
Conceptual knowledge acquisition system
Symbolic level - has symbolic logic
Conceptual level - ???
Perceptual level - DBN, HMAX...
Need to move between the different levels.
Demis Hassabis on wikipedia
Demis page at the university college of London
If you liked this article, please give it a quick review on Reddit, or StumbleUpon. Thanks
How to Make Money