Liquid neural networks are among the most important and unique emerging components in the artificial intelligence (AI) landscape.
When a machine or robot needs to react to external stimulus or data, it can be extremely resource-heavy, causing a bottleneck if you are trying to fit intelligence into a very small space.
VentureBeat describes how a classical neural network might need 100,000 artificial neurons to keep the car steady in a task such as driving a vehicle down a road.
However, in an incredible finding, the team at MIT CSAIL developing liquid neural networks was able to perform the same task with just 19 neurons.
The Inspiration Behind Liquid Neural Networks
Liquid neural networks are a type of deep learning architecture developed to solve a challenge for robots performing complex learning and tasks, aiming to cut around the problem of dependency on the cloud or limited internal storage.
Daniela Rus, the director of MIT CSAIL, told VentureBeat: “The inspiration for liquid neural networks was thinking about the existing approaches to machine learning and considering how they fit with the kind of safety-critical systems that robots and edge devices offer.
“On a robot, you cannot really run a large language model because there isn’t really the computation [power] and [storage] space for that.”
The research team found a clue to their problem from the research on biological neurons found in tiny organisms.
What are Liquid Neural Networks?
Think of liquid neural networks as the interconnected cells of a human brain that come together to process information and provide output.
The human brain is a highly complex cell arrangement that performs extremely complex computations.
Liquid neural networks focus on safety-critical applications, such as self-driven vehicles and robots, that need a continuous stream of data being fed into them.
According to Daniela Rus, “In general, liquid networks do well when we have time-series data … you need a sequence in order for liquid networks to work well.
“However, if you try to apply the liquid network solution to some static database like ImageNet, that’s not going to work so well.”
Advantages and Limitations
The research team at the Computer Science and Artificial Intelligence Laboratory at MIT (CSAIL) found the following advantages based on their experience.
- Compactness
Liquid Neural Networks could work with a significantly lesser number of neurons than classic neural networks.
As outlined above, a classic deep-learning neural network would need 100,000 neurons to keep a self-driving car in its lane — a liquid neural network needs just 19 neurons.
- Causality
Liquid neural networks handle causality better than classic deep-learning neural networks. They can spot a clear relationship between the cause and effects, which classic deep-learning neural networks struggle to do.
For example, the classic deep-learning neural networks can consistently identify cause-and-effect relationships between events across diverse settings more efficiently than the classic neural network.
- Interpretability
Understanding an AI system’s interpretation of data is one of the biggest challenges in AI.
Classic deep-learning models often display shallow, unclear, or wrong basis for interpretations of data, but liquid neural networks can explain their basis for interpreting data.
- However…
Liquid neural networks are not a comprehensive solution for everything.
While they handle continuous data streams such as audio streams, temperature data, or video streams well, they struggle with static or fixed data, which are better suited to other AI models.
The Bottom Line
In the AI landscape, liquid neural networks are among the most critical emerging models.
It coexists with the classic deep-learning neural network but appears a better fit for extremely complex tasks such as autonomous vehicles, temperature or climate reading, or stock market assessments, whereas the classic deep-learning neural network does a better job with static or one-time data.
The researchers at the Computer Science and Artificial Intelligence Laboratory at MIT (CSAIL) have been trying to extend the capabilities of liquid neural networks to more use cases, but it will take time.
Both liquid neural networks and classic deep-learning neural networks have their defined roles in the broader AI picture, and it’s definitely a case where two models are better than one.