Schmidhuber paper

Brief overview of Goedel machines; sort of a rebuke of other authors for ignoring the optimality results for them and AIXI etc.

Simultaneously, our non-universal but still rather general fast deep/ recurrent neural networks have already started to outperform traditional pre-programmed methods: they recently collected a string of 1st ranks in many important visual pattern recognition benchmarks, e.g. Graves & Schmidhuber (2009); Ciresan et al. (2011): IJCNN traffic sign competition, NORB, CIFAR10, MNIST, three ICDAR handwriting competitions. Here we greatly profit from recent advances in computing hardware, using GPUs (mini-supercomputers normally used for video games) 100 times faster than today’s CPU cores, and a million times faster than PCs of 20 years ago, complementing the recent above-mentioned progress in the theory of mathematically optimal universal problem solvers.

On falsified predictions of AI progress:

I feel that after 10,000 years of civilization there is no need to justify pessimism through comparatively recent over-optimistic and self-serving predictions (1960s: ‘only 10 instead of 100 years needed to build AIs’) by a few early AI enthusiasts in search of funding.

Pessimism:

All attempts at making sure there will be only provably friendly AIs seem doomed though. Once somebody posts the recipe for practically feasible self-improving Gödel machines or AIs in form of code into which one can plug arbitrary utility functions, many users will equip such AIs with many different goals, often at least partially conflicting with those of humans. The laws of physics and the availability of physical resources will eventually determine which utility functions will help their AIs more than others to multiply and become dominant in competition with AIs driven by different utility functions. The survivors will define in hindsight what’s ‘moral’, since only survivors promote their values...

The Hard Problem dissolved?

But at least we have pretty good ideas where the symbols and self-symbols underlying consciousness and sentience come from (Schmidhuber, 2009a; 2010). They may be viewed as simple by-products of data compression and problem solving. As we interact with the world to achieve goals, we are constructing internal models of the world, predicting and thus partially compressing the data histories we are observing. If the predictor/compressor is an artificial recurrent neural network (RNN) (Werbos, 1988; Williams & Zipser, 1994; Schmidhuber, 1992; Hochreiter & Schmidhuber, 1997; Graves & Schmidhuber, 2009), it will create feature hierarchies, lower level neurons corresponding to simple feature detectors similar to those found in human brains, higher layer neurons typically corresponding to more abstract features, but fine-grained where necessary. Like any good compressor the RNN will learn to identify shared regularities among different already existing internal data structures, and generate prototype encodings (across neuron populations) or symbols for frequently occurring observation sub-sequences, to shrink the storage space needed for the whole. Self-symbols may be viewed as a by-product of this, since there is one thing that is involved in all actions and sensory inputs of the agent, namely, the agent itself. To efficiently encode the entire data history, it will profit from creating some sort of internal prototype symbol or code (e. g. a neural activity pattern) representing itself (Schmidhuber, 2009a; 2010). Whenever this representation becomes activated above a certain threshold, say, by activating the corresponding neurons through new incoming sensory inputs or an internal ‘search light’ or otherwise, the agent could be called self-aware. No need to see this as a mysterious process — it is just a natural by-product of partially compressing the observation history by efficiently encoding frequent observations.