Reversibility and machine intelligenceThis essay is about the intersection of computational reversibility and machine intelligence. It argues that reversibility is important for intelligent systems, and that this association impacts on the probability of the hypothesis that we exist inside a computer simulation.
The benefits of reversibilityComputational reversibility has well-known benefits relating to minimizing power consumption and heat generation. These are a consequence of the thermodynamic cost of erasing bits - an idea known as Landauer's principle However there are also some times when reversibility helps with actually running a system in reverse - for example when providing 'undo' functionality or when backtracking. The need for backtracking is common when performing a tree search - and tree searches are used ubiquitously in the forecasting component of machine intelligence systems.When performing a tree search it is common to want to go back and reconsider a previous node. For example, in a game of go if you play A, your opponent plays B, you play C and your opponent plays D - and then your position evaluation tells you that you are going to lose the game after this, you may want to back track and reconsider your move at C. If no move at C seems any good you may want to back track further to reconsider your move at A. This process involves running your simulation of the game both forwards and backwards. This example involves an adversary and a series of branching points corresponding to agent actions, but the same situation applies to practically any forecasting problem involving uncertainty. What makes the tree of future possibilities branch is not necessarily a choice between possible actions, but any form of uncertainty which could potentially lead to multiple possible outcomes. This makes the need for reversibility very widespread in forecasting systems. Another thing to mention is that forecasting systems are a central and ubiquitous component of machine intelligence systems. Forecasting the possible consequences of actions to allow an agent to choose between the resulting outcomes is a the primary function of animal brains. Machine intelligence works on the same principle. For more on this topic, see my essay titled: The broad scope of inductive inference. The simplest way to run a system backwards is to store a complete history of previously-considered nodes in the tree. However, that is often very inefficient. More often, you want to do something like store a diff. That's where reversible logic comes in. Reversible logic lets you efficiently run any computation in reverse. It uses Toffoli gates, the Margolous neighbourhood, or other techniques to run the system backwards. Computational universality allows a reversible computer to simulate any irreversible computation. A classic text about egnineering reversibile systems is the book Cellular Automata Machines by Tommaso Toffoli and Norman Margolus.
Reversibility and simulismIt has been widely speculated that we may be living inside a computer simulation. The possibility has long been a staple of science fiction with early examples including:The Tunnel under the World (1955), Time Out of Joint (1959) and (1964). Modern treatments include The_Matrix (1999), Total Recall (1990), The Thirteenth Floor (1999) The Lawnmower Man (1992) Dark City (1998) Strange Days (1995) and eXistenZ (1999). The topic has become of interest to scientists and philosophers. Frank Tipler's (1997) book, The Physics of Immortality: Modern Cosmology, God and the Resurrection of the Dead speculated that modern humans might be resurrected n future computer simulations. Hans Moravec's (1998) paper Simulation, Consciousness, Existence was another early contribution to the topic. Nick Bostrom's 2003 paper Are You Living In a Computer Simulation? also went on to address the subject. The laws of physics appear to exhibit microscopic reversibility. They are time symmetric invariant. This observation was contested in the 20th century by quantum physics pioneers who claimed that a hypothetical irreversible process known as "wave function collapse" might play a significant role in physics. However, there is no evidence that wavefunction collapse is a real phenomenon. As it turns out, quantum physics does not depend on the idea in any way. The best evidence we have is consistent with exact microscopic reversibility. Not all possible laws of physics exhibit reversibility. For example, irreversible rules massively outnumber reversible ones in cellular automata. Of the 256 Wolfram rules, only 6 are reversible. The proportion of reversible rules goes down rapidly as the complexity of the rule increases. In one of the next simplest one dimensional systems, 62 rules out of 4 billion are reversible. One possible explanation of reversibility invokes observation selection effects. Maybe mature living systems can only exist in universes with reversible physical laws - since in irreversible universe the information in them is gradually erased and before intelligent life has time to evolve, it is all gone. However, this possibility doesn't seem sufficient to explain the observations. Life can certainly exist in irreversible universes - and it may also be able to persist for billions of years - provided that the degree of irreversibility is not so great that the universe rapidly leaks away all the information inside it. Exact physical reversibility appears to be not fully explained by this idea. One possible explanation for physical reversibility is that the universe is the product of intelligent design. Reversibility helps to minimize power requirements and heat generation. The link between machine intelligence and reversibility described in this essay provides some more possibilities. Reversibility could be important to function in other ways. Reversibility allows the implementation of a "rewind" function - allowing the simulation to be examined at any point in time. Also, reversibility is a likely property of world simulations that exist inside advanced minds. These are likely to be engineered using reversibility to allow for efficient backtracking. Some critics have suggested that advanced machine intelligences would not squander resources on detailed ancestor simulations. However, we are quite interested in our history and origins. Understanding the past is key to predicting the future. Major evolutionary transitions are likely to be of particular interest to our descendants. In particular, in the alien race, it could be important to understand how the transition to an engineered future turns out, because that might be our best clue about the possible forms of aliens we might meet. Understanding the forms aliens could take could be key strategic information - should we ever meet them. Physical reversibility thus appears to be a piece of evidence that favors simulism over other possibilities. If our universe is the product of intelligent design that makes sense of our reversible laws of physics. Otherwise, we seem faced with a massive coincidence - physical law is reversible - although the odds of it being reversible would seem to be miniscule.
|
