Is Machine Learning The Quantum Physics Of Computer Science ? – Forbes

0 2

Preamble:   Intermittently, I will be introducing some columns which introduce some seemingly outlandish concepts. The purpose is a bit of humor, but also to provoke some thought. Enjoy.   

atom orbit abstract

atom orbit abstract


 “God does not play dice with the universe,” Albert Einstein is reported to have said about the field of Quantum Physics.    He was referring to the great divide at the time in the physics community between general relativity and quantum physics. General relativity was a theory which beautifully explained a great deal of physical phenomena in a deterministic fashion.  Meanwhile, quantum physics grew out of a model which fundamentally had a probabilistic view of the world. Since Einstein made that statement in the mid 1950s, quantum physics has proven to be quite a durable theory, and in fact, it is used in a variety of applications such as semiconductors. 

One might imagine a past leader in computer science such as Donald Knuth exclaiming, “Algorithms should be deterministic.”   That is, given any input, the output should be exact and known. Indeed, since its formation, the field of computer science has focused on building elegant deterministic algorithms which have a clear view of the transformation between inputs and outputs.  Even in the regime of non-determinism such as parallel processing, the objective of the overall algorithm is to be deterministic. That is, despite the fact that operations can run out-of-order, the outputs are still exact and known. Computer scientists work very hard to make that a reality. 

As computer scientists  have engaged with the “real world,”  they frequently face very noisy inputs such as sensors or even worse, human beings.  Computer algorithms continue to focus on faithfully and precisely translating input noise to output noise. This has given rise to the Junk In Junk Out (JIJO) paradigm.  One of the key motivations for pursuing such a structure has been the notion of causality and diagnosability. After all, if the algorithms are noisy, how is one to know the issue is not a bug in the algorithm?  Good point.   

With machine learning, computer science has transitioned to a model where one “trains” a machine to build an algorithm, and this machine can then be used to transform inputs to outputs. Since the process of training is dynamic and often ongoing,  the data and the algorithm are intertwined in a manner which is not easily unwound. Similar to quantum physics, there is a class of applications where this model seems to “work.” Recognizing patterns seems to be a good application. This is a key building block for autonomous vehicles, but the results are probabilistic in nature.  

In quantum physics, there is an implicit understanding that the answers are often “probabilistic”  Perhaps this is the key insight which can allow us to leverage the power of machine learning techniques and avoid the pitfalls.  That is, if the requirements of the algorithm must be exact, perhaps machine learning methods are not appropriate. As an example, if your bank statement was correct with somewhat high probability, this may not be comforting.    However, if machine learning algorithms can provide with high probability the instances of potential fraud, the job of a forensic CPA is made quite a bit more productive. Similar analogies exist in the area of autonomous vehicles. 

Overall, machine learning seems to define the notion of probabilistic algorithms in computer science in a similar manner as quantum physics. The critical challenge for computing is to find the correct mechanisms to design and validate probabilistic results. 

Leave A Reply

Your email address will not be published.