The world of AI is being dominated by Big Tech – we all know that. Why? They have deep pockets to fund all of the PhD brainpower needed to build the elusive artificial general intelligence that everyone is after. But, on a more granular level, Big Tech has access to vast troves of data. And that data is needed to build the training models that neural networks are built on.
Ok, got it. So, what’s really interesting about this new study from MIT researchers is the development of a neural network that can continuously learn rather than just learn from data once during its ‘training phase.’
Learning on the fly
Called ‘liquid neural networks,’ these machine learning algorithms change their underlying equations to continuously adapt to new data inputs, per an article in SingularityHub.
Translation: In the real world, data is always changing and decisions have to be made based on new data coming in all the time. These new algorithms could be very helpful in many different areas, but the researchers point to medical diagnosis and autonomous driving as two immediate areas where the algos could deliver.
Lessons from biology
The architecture of algorithm behind these liquid neural networks was inspired by the nervous system of C. elegans, a tiny nematode (or worm). This worm only has 302 neurons but is still capable of surprisingly varied behavior.
So here’s the kicker – the worm algorithm was practically as effective as much more advanced, cutting edge machine learning algorithms in tasks like helping to keep a car in its own lane in autonomous driving systems.
“Everyone talks about scaling up their network,” said Ramin Hasani, the study’s lead author. “We want to scale down, to have fewer but richer nodes.”
Another way to build the AI of the future
“Today, deep learning models with many millions of parameters are often used for learning complex tasks such as autonomous driving,” according to another researcher involved in the study, Mathias Lechner. “However, our new approach enables us to reduce the size of the networks by two orders of magnitude. Our systems only use 75,000 trainable parameters.”
For comparison, OpenAI’s GPT-3 algorithm uses a record-setting 175 billion parameters. And a recent Google algorithm trained an NLP model with over a trillion parameters.
Obviously, these models can be viewed as wasteful and bloated if other algorithms can accomplish the same tasks with less. And it’s also one of the first principles of good coding – simpler is better.