AI's Role in Our Urban Future

A Pentagon video obtained by The Intercept envisions a dystopian future for global megacities as populations in the developing world shift from rural to dense urban environments.

"This is the world of our future.  It is one we are not prepared to effectively operate within and it is unavoidable," concludes the video.

Although the problems discussed are politically and socially complex, some of the issues highlighted may be avoidable through the use of AI.

 
 

 

Uncontrolled Development

One key concern raised in the video is the notion of uncontrolled development, "as impoverishment, slums and shantytowns rapidly expand alongside modern high-rises". Dharavi in Maharashtra India, Neza-Chalco-Itza in Mexico City and Kibera in Nairobi are prominent examples of such developments.  

However such massive developments do not happen overnight and Ilievski suggests that machine learning models can "capture the patterns of urban change driven by a diverse set of context factors" like economic growth, demographics changes, proximity, and neighborhood conditions.

Ilievski demonstrates that machine learning can accurately predict how land use evolves over time across categories like undeveloped, main street, peripheral street, industry, house, education, park, etc.

 
Ilievski labeling tool for orthographic images and maps.  Labeled cells were "divided into a training sample of 1,110,786 cell instances and testing set containing 480,576 cells. The period from 1960 to 1996 was used to train the model, while the time period 1997 - 2013 was used to test the performance" (Ilievski)

Ilievski labeling tool for orthographic images and maps.  Labeled cells were "divided into a training sample of 1,110,786 cell instances and testing set containing 480,576 cells. The period from 1960 to 1996 was used to train the model, while the time period 1997 - 2013 was used to test the performance" (Ilievski)

 

Constrained Resources

Another significant challenge raised by the video is shortage of natural resources.  Here AI can play a dual role of more efficiently managing the resources available and accelerating the cultivation of new ones.

DeepMind recently demonstrated that using "machine learning [in] Google data centres [can] reduce the amount of energy we use for cooling by up to 40 percent." (DeepMind)

Another RL example comes from Castelletti.  They propose that reinforcement learning (Q-learning) can be used to model optimal water reservoir operations with an increased number of state variables (while gaining computational efficiency over prevailing Stochastic Dynamic Programming).

Behmann, et al. survey successful applications of supervised and unsupervised methods used for "early detection of weeds, plant diseases and insect pests in crops."

Improvised InfrastructurE

Another area of focus is, what happens to improvised infrastructure like "makeshift power grids" particularly in the face of natural disasters?

An interesting development in this area is the concept of "transiently-powered computing" that are "networked autonomous devices, collections of tens to thousands of nodes organized into cooperative networks." (Balsamo)

Balsamo "presents a new power-neutral paradigm to tackle the power supply challenge for the transient computing system whose operation is solely based on harvested power. Such a paradigm enforces the well match between the instantaneous power consumption and the harvested power through using an innovative control algorithm exploring dynamic frequency scaling" (Hu) 

Doing Our Part

While there are many promising research areas, we should not underestimate the scale and the complexity of the challenges that the Pentagon video highlights.

We at Adversarial.AI (and our parent Startup.ML) are taking some small steps to:

  • Ensure that the world will have more qualified data science and machine learning professionals to work on these problems through our fellowship program.
  • Create adversarial machine learning solutions that enterprises and governments can understand and operate.
  • Grow and inform the machine learning community through our conference and meetup events such as the upcoming Putting Deep Learning into Production event on Jan 21.

 

Reference

Balsamo, Domenico, et al. "Graceful Performance Modulation for Power-Neutral Transient Computing Systems." IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 35.5 (2016): 738-749.

Behmann, Jan, et al. A review of advanced machine learning methods for the detection of biotic stress in precision crop protection. Precision Agriculture 16.3 (2015): 239-260.

Castelletti, A., et al. Tree‐based reinforcement learning for optimal water reservoir operation. Water Resources Research 46.9 (2010).

DeepMind Blog. DeepMind AI Reduces Google Data Centre Cooling Bill by 40% 

Hu, Shiyan, Xiaobo Sharon Hu, and Albert Y. Zomaya. Guest Editorial Leveraging Design Automation Techniques for Cyber-Physical System Design. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 35.5 (2016): 697-698.

Ilievski, Ilija, Sonja Gievska, and Ognen Marina. Discovering Patterns of Urban Development.(2013).

 

acknowledgement

Erin Craig

Juan Carlos Asensio