Does Teenage Mutant Ninja Turtles ring a bell? If it does, you’ll remember how awkward-looking those mutated creatures were. If not, but you are familiar with mutants – primarily through tv – you most probably link mutation with bad luck. But, have you heard about AI Mutation? I ask because – as you’d later come to see – mutation, when done right, can result in groundbreaking innovation.
Caveat: Despite being about AI, this isn’t one of those difficult-to-read, dull, or scary tech articles you see around. So, relax and prepare to have a smooth ride all through.
AutoML: 1st Gen Machine Learning
If you are a programmer or are close with one, you sure must be aware of the backend struggles that go into creating, and intermittently updating codes. But with the evolution of machine learning (AutoML), this struggle is being overcome.
With machine learning, AI tools are getting equipped with the ability to learn, optimize, and update codes automagically with human contribution limited to the constant provision of new data. That’s what reigns supreme in today’s age of tech development – though we are just barely scratching the surface of faster learning.
See Machine Learning as the human-like characteristics of artificial intelligence systems to learn, get better at tasks through experience and constant processing of data. Let’s say a programmer creates a program that can tell the difference between the picture of a dog and a cat.
Machine learning or faster learning is that which enables the AI system to form statistical patterns amongst other methodologies that help it almost-perfectly always distinguish dogs – no matter the specie – from cats.
With data from different cats and dogs species intermittently fed to the system, the system starts to decipher that cats have shorter noses, and dogs come in a variety of sizes.
Artificial Intelligence Mutation: 2nd Gen Machine Learning
What if there’s a need to create a system that distinguishes species of dogs? The simple answer is that the programmer again goes into seclusion and writes a new set of codes to that effect. Picture the resources that typically go into each manual programming projects.
Thanks to Google Developers, this may gradually be a thing of the past. With their discovery of AutoML-Zero, artificial intelligence systems can finally be encouraged to mutate and create newer algorithms from scratch with much lesser human input.
So, to create a dog specie identification AI system, the programmer only specifies the desired result, and the machine gets the work done by the reworking DNA of existing codes.
Using Darwinian evolution concepts, the mutant program pools together newer algorithms, which are results of random mathematical operation combinations. The program evaluates the performance of each algorithm in identifying dog species and compares it.
The top performers are retained, and others are discarded – sort of like survival of the fittest. The retained top performers have repeatedly mutated to birth much-better algorithms while the top-performing parents are abandoned. On and on the process goes.
Using tricks to speed up its processing, the AI mutant keeps churning out tens of thousands of newer, altered algorithms per seconds, all in a bid to find the perfect set of codes/algorithm that matches the programmer’s needs. All along the way, the system weeds out duplicate algorithm to prevent evolutionary dead-ends.
How Google Experimented AI Mutation
Developers at Google are the one working on this freakingly-smart concept that’s capable of leading to the evolution of AI systems that outperforms anything made by humans. In its released research reports at arXiv, the Google AI mutant experimental system churned out 1000 image-recognition algorithms optimized to identify a specific set of images.
After making 250 computers choose two best algorithms, the best of each chosen two algorithms with the highest accuracy survived, and the poor one got discarded.
The survivors were cloned and mutated to reproduce newer algorithms that interpret and respond to the training data in a slightly different way. The first-gen parents got discarded.
On and on, the process went, with mutated algorithms having slightly better results getting retained. And those that didn’t get discarded until the repeated mutation, birthed an algorithm with 94.6% accuracy.
All this is still, however, in the experimental stage.
How AI Mutation Ties with Darwinian Evolution
Ruminations birthed this interesting Artificial intelligence mutation concept on the Charles Darwinian human evolution. His reasoning, which has since been refined by several scientists, is that calculated alterations in DNA (genetic makeup) leads to either of an advantage or disadvantage to such mutant organism.
If the mutation is of a non-life-threatening nature that lets the organism survive and birth offsprings, such modification is passed along to slightly-optimized offsprings. If it doesn’t survive, the mutation dies with such a parent organism.
In the algorithm space, this process is referred to as neuroevolution – a process that attempts to recreate the “survival-of-the-fittest” human brain-building process in AI.
AI Mutation: Panacea to AI-related Challenges
The AI world is optimistic about the development and market viability of this 2nd-Gen machine learning capability. That’s because it would significantly help in resolving some of the challenges associated with today’s AI models.
Talk of the problem of bias in AI systems, which are unknowingly introduced by programmers when creating AI systems. Or do we talk about the resources – time and money – spent in continually developing and test-running newer sets of codes for more modern algorithms.
Despite the complexities of AI system development, AutoML-Zero could prove to be the much-needed innovation that propels AI systems into highly-efficient, cutting-edge systems they are meant to be.
The future we’ve always anticipated where computers can easily outsmart humans might be almost here. So, if anyone happens to ask if you’ve heard of AI Mutation, don’t say you haven’t.
Ridwan Sharomi is an aspiring tech lawyer and freelance writer (for hire). He works closely with individuals and businesses, providing top-notch blog writing and ghostwriting services.