From Machine Learning to Machine Creativity

20 Jan,2022

 

 

By Ashoke Agarrwal

 

Ashoke Agarrwal

Ashoke Agarrwal

Understanding and encouraging creativity has been a significant part of my professional endeavours as an advertising strategist.

 

I have understood the difference between strategic planning and creative development as a “P versus NP” type issue.

 

A P-type problem is a problem that yields a solution in a finite amount of time.

 

Simply put, a P-type problem is humanly soluble provided one puts the right type and amount of effort!.

 

The NP-type problem is a problem whose solution is checkable for correctness in a finite amount of time. For example, Sudoku is an NP-type problem. Whatever the size of the Sudoku grid, one can check the correctness of a given solution in a finite amount of time.

 

A significant issue at the frontier of maths and computer science is whether an NP-type problem is also a P-type problem.

 

As a practising advertising professional, I have my own take on the P-type and N_-type problems dichotomy.

 

As I think of it, creating a marketing and advertising strategy addresses a P-type problem where a given strategy meets a given objective through a series of rationale and finite steps.

 

On the other hand, creative solutions are the result of addressing an NP-type problem.

 

Most seasoned professionals can judge the effectiveness of a given creative solution. However, the debate has always been whether creative solutions can be arrived at through a series of logical steps. The consensus, as of now, is that in this case, NP-type is not equal to P-type. In all walks of life, creative output results from a creative leap well beyond the restrictions of logical and rational steps.

 

However, there is now a clear challenge to the above notion coming from the cutting edge of Machine Intelligence (MI). Over the past few years, the second generation of MI has emerged through a set of neural network techniques based on Deep Learning (DL) principles. DL took a giant leap forward when Jürgen Schmidhuber of Lugano University and his student Sepp Hochreitter proposed a Recurrent Neural Network (RNN) architecture called the Long Short-Term Memory (LSTM).

 

With this, DL evolved and gained a specific ability of human intelligence – learning how to learn.

 

The result of this new generation of DL has been magical.

 

For example, in the late nineties, before DL has evolved, Deep Blue – the IBM AI engine that beat at Chess, the reigning world champion – Gar Kasparov – using a recursive analysis of millions of chess games that the machine had on record before every time it had to make a move. The essence of Deep Blue lay in the super-fast speed that enabled it to analyse millions of alternatives before every move.

 

Cut to 2016. Using LSTM type RNN techniques, DeepMind, a Deep Learning engine, sets out to learn Go (a strategy game considerably more complex than Chess) and Chess. DeepMind was fed the game’s rules, and the program began to learn by playing against itself.

 

The results have been astonishing. AlphaGo, the DeepMind engine for Go, has consistently beat world champions. AlphaZero. DeepMind’s chess engine obliterated Stockfish, the highest-rated “old-world AI” chess engine.

 

Every passing month DeepMind is taking DL deeper, crossing new frontiers holding out the prospect of AI finally leaping to General Intelligence (GI).

 

However, to my mind, for computer science to finally prove that NP-type problems are also P-type problems, one more revolutionary step is required.

 

The next generation of DL will have to move from supervised learning to active learning. The first step to active learning would be to give the AI engine agency. Then, to let the AI decide in an autonomous way on what to observe and study – an attribute called curiosity. In other words, an AI system capable of generating and acting upon Machine Curiosity.

 

The curiosity will be directed by a higher-order goal setting than the tight purposive framing the current generation of MI operates on. The next step would be for the MI system to act on this curiosity. Schmidhuber, a founder of Deep Mind, believes that a critical capability for a MI system to act on its curiosity is not just to observe and act on available data but also to create data by poking, prodding and experimenting with the real world. This next generation of curiosity-driven, experiment-making MI systems will unleash creativity of an order higher than the world has yet to see. According to the experts, we can see this next generation of MI within the next decade or two.

 

That will be a big step towards computer science, finally proving that most NP-type problems can be P-type problems.

 

And for me to add further ballast to my assertion that I work at the creative end of strategy and the strategic end of creativity.

 

Further, in an earlier post on MxMIndia, I had written about Concierge Intelligence. The creation and deployment of Concierge Intelligence will be possible only by Machine Learning empowered by Machine Curiosity leading to Machine Creativity in the service and control of the individual instead of business or government entities.

 

You can read more about the P-NP type problems in my post on Medium titled “Advertising and the P-NP Problem”, dated July 4, 2019.

 

A post on Medium titled “Machine Intelligence to Machine Curiosity – The Road to Machine Creativity”, dated June 26, 2019, delves a little deeper into Machine Creativity.

 

 

Ashoke Agarrwal is a veteran advertising professional with around four decades in advertising and marketing services. Agarrwal, a chemical engineer from IIT Mumbai and a postgraduate from IIM Bangalore, is a pro-entrepreneur with past and current ventures in market research, advertising, CGI, e-learning and brand consultancy. He will write on MxMIndia every other Thursday. His views here are personal.

 

Post a Comment 

Comments are closed.