Develop biological evolutionary genetic algorithms to identify objects of interest in large digital images using Cloud Computing.

Image Recognition in the Cloud – EvoDevo

The Problem

  • No existing automated processes come close to the ability of a human being looking at images to identify objects of interest.
  • Years and billions of R&D’s money to develop software algorithms to find objects of interest in images have had limited success.
  • The volume of images to analyze has continued to grow.

The Solution

  • Using recent advances in our understanding of how the brain processes images and the biological evolution process, develop and evolve algorithms that approach the efficiency and accuracy of the human brain for image understanding.
  • Use cloud computing to evolve a solution to visual object recognition using an EvoDevo approach in combination with Genetic Algorithms to create a neural network.

The Impact

  • Achieving 97.5% recognition using EvoDevo evolved in the cloud.

Develop biological evolutionary genetic algorithms to identify objects of interest in large digital images using Cloud Computing.

The Problem

Under contract to the U.S. Government, Next Century is researching the use of biological evolution and development to evolve neural networks that recognize objects in digital images. The results of the first phase of this project were very promising, warranting further research and experimentation.

A severe constraint during the first phase of EvoDevo was the lack of computer power. For example, during the final run of the initial phase, it took eight processing cores two calendar weeks to “evolve” the neural networks, resulting in 80 generations of neural network evolution. Even after monopolizing the expensive hardware for this length of time, it was clear that more runs and more generations per run were needed to further the research. To accomplish this goal and overcome the limits of the original computing resources, Next Century evaluated EvoDevo using cloud computing.

The Technical Approach

Use Amazon’s EC2 cloud to dramatically increase the compute power and reduce run time.

For this project Next Century used Amazon Elastic Compute Cloud (EC2) as the underlying infrastructure, leveraged Amazon Web Services (AWS) security processes, and used Appistry CloudIQ as the fabric to hold infrastructure together and coordinate processing. Genetic algorithms are highly parallel. We used the genetic algorithm to spread individual development and evaluation to many machines.  After all evaluations, we created a new generation from the best individuals and then re-evaluated.

Current Performance Comparisons (500 Individuals, 92 Generations)

On in-house server “ls4”

  • Dell Xeon, 8 cores
  • Runtime (est.):   50 hr 45 min
  • Purchase price: $19,936

Note: To get a comparable “ls4” runtime, to equal the EC2, would require four dedicated servers at a total cost of $79,744.

On Amazon EC2

  • 100 small instances, 1 core each
  • Runtime:   12 hrs. 5 min.
  • EC2 usage cost: $130

Cost Comparison

If EC2 was used in Phase I of the project, there would have been a significant calendar time savings and reduced labor costs!

Original Project Usage Cost

  • Estimated CPU hours used: 2,420 hours
  • Cost of ls4: $19,936 (+ electricity, cooling, support)
  • Cost per hour used: $19,936/2,420 = $8.24/hour
  • Project length: 8 months (5,760 hours)
  • Effective ls4 usage rate:  2,420/5,760 = 42%

Equivalent EC2 Usage Cost

  • Cost per hour used (extra large, high-CPU): $0.80/hour
  • EC2 for 2,420 hours: $1,936

Additional Technical Information

ls4 EC2 Small Instance EC2 Large Instance
2GHz Xeon ~ ¼ ls4 ~ equal to ls4
2 quad-core CPUs 1 core 8 cores
4GB memory 1.7 GB memory 7GB of memory

Note: The resulting algorithm is currently achieving 97% positive (3% false negatives) and 90% (10% false positives) ship recognition!

Categories: Cloud Computing, Image Processing