Can machine learning bring about an intelligence explosion? (Extended abstract)

Itamar Arel, Department of Electrical Engineering and Computer Science, The University of Tennessee

Reward-Driven Learning and the Threat of an Adversarial Artificial General Intelligence Singularity 

A myriad of evidence exists in support of the notion that mammalian learning is driven by rewards. Recent findings from cognitive psychology and neuroscience strongly suggest that much of human behavior is propelled by both positive and negative feedback received from the environments with which we interact. The notion of reward is not limited to indicators originating from a physical environment. It also embraces signaling generated internally in the brain, based on intrinsic cognitive processes. Artificial General Intelligence (AGI), coarsely viewed as human- level intelligence manifested over non-biological platforms, is commonly perceived as one of the paths that may lead to the singularity. Such a path has the potential of being either beneficially transformative or devastating to the human race, to a great extent depending on the very nature of the AGI. 


Reinforcement learning (RL) is a fairly mature field within artificial intelligence, delivering a rigorous mathematical framework for machine learning through interaction with an environment. Consequently, it serves as one of the promising foundations for advancing AGI research. An RL agent attempts to map situations to actions with the goal of maximizing its expected future rewards. Strategic thinking properties emerge from the inherent consideration of both short and long term rewards. The key challenge in RL, as a mechanism for decision making under uncertainty, remains that of scalability. The latter refers to dealing with high-dimensional observations spanning large state and action spaces, which characterize real-world setting.

In late the 1950s, Richard Bellman who introduced dynamic programming theory and pioneered the field of optimal control, predicted that high-dimensionality data will remain a fundamental obstacle for many science and engineering systems over decades to come. The main difficulty he highlighted was that learning complexity grows exponentially with a linear increase in the dimensionality of the data. He coined this phenomenon the curse of dimensionality and his prediction proved amazingly true.

Recent neuroscience findings have provided insight into the principles governing information representation in the mammal brain, leading to new paradigms for designing systems that represent information. Deep machine learning has recently been coined as a biologically-inspired framework for dealing with the curse of dimensionality, by exploiting a hierarchical architecture for information representation. Such architectures attempt to mimic the manner by which the cortex learns to represent regularities in real world observations. In addition to the spatial aspects of real-life data, the temporal component often plays a key role. To that end, robust spatiotemporal modeling of observations serves as a primary goal for deep learning systems.

It has recently been hypothesized that the fusion between deep learning, as a scalable situation inference engine, and reinforcement learning as a decision-making system my hold the key to place us on the path to AGI. Assuming that this hypothesis is correct, many critical questions arise, the first of which is how do we avoid a potentially devastating conflict between a reward-driven AGI system and the human race? One can argue that such a scenario is inescapable, given the assumption that an RL-based AGI will be allowed to evolve. In that case, does evolution have to continue over biochemical substrates, or will the next step in evolution involve semiconductor-based realizations of life? Consequently, will AGI bring the human era to an inevitable end? Transhumanism may very well emerge as a transitional period at the end of which post-humanism will commence where life will seize to exist in a biochemical form.

History suggests that pragmatic concerns pertaining to the potential dangers and threats of novel technologies have never impeded such technologies from being widely embraced. Nuclear technology is an obvious example, particularly in that debate over its benefits verses its threats has persistently accompanied its deployment. Although technological progress is needed to make AGI a reality, there is some likelihood that the pieces of the puzzle needed to make AGI a reality are in fact readily available, in which case now is the time to consider the colossal implications of an AGI-driven singularity.