What is a Liquid Neural Network?
A liquid neural network (LNN) is a time-continuous recurrent neural network built with a dynamic architecture of neurons. These neurons are able to process time-series data while making predictions based on observations and continuously adapting to new inputs.
Their adaptable nature gives them the ability to continually learn and adapt and, ultimately, process time-series data more effectively than traditional neural networks.
LNNs were originally developed by the Computer Science and Artificial Intelligence Laboratory at MIT (CSAIL), which attempted to make a machine learning (ML) solution capable of learning on the job and adapting to new inputs.
The concept was inspired by the microscopic nematode C.elegans, a worm that only has 302 neurons in its nervous system but still manages to respond dynamically to its environment.
Liquid Neural Networks vs. Neural Networks
One of the key differences between LNNs and neural networks is that the former uses dynamic connections between neurons, whereas traditional neural networks have fixed connections and weights between each neuron.
These flexible connections mean that liquid neural networks can continuously adapt to and learn from new data inputs in a way that traditional neural networks can’t, as they are dependent on their training data. This makes LNNs better at processing time-series data but is also less effective at processing static or fixed data than other neural networks.
It’s important to note that the dynamic architecture of liquid neural networks also requires fewer overall neurons than a neural network and consumes less overall computing power. Their low computational needs mean they can be used to run on lightweight computers and hardware such as microcontrollers.
LNNs are more interpretable than more complex black-box neural networks because it’s easier to see how data inputs are influencing outputs.
What are Liquid Neural Networks Used For?
As mentioned above, LNNs are generally used for time series data processing and prediction on smaller computers. The lower computational needs of these solutions mean they can run on devices with minimal computing power, from robots to devices at the network’s edge.
This makes them ideal for a wide range of use cases running from natural language processing (NPL) and video processing to autonomous robotics, vehicles, drones, and medical diagnosis.
LNNs and Automated Drones
In April 2023, MIT researchers unveiled research demonstrating how liquid neural networks could be used to help teach aerial drones to navigate to a given object and to respond correctly in complex environments like forests and urban landscapes.
As Daniela Rus, CSAIL director and the Andrew and Erna Viterbi Professor of Electrical Engineering and Computer Science at MIT, explained:
“Our experiments demonstrate that we can effectively teach a drone to locate an object in a forest during summer and then deploy the model in winter, with vastly different surroundings, or even in urban settings, with varied tasks such as seeking and following.”
Traditional deep learning solutions would have been poorly suited to this use due to their lack of ability to adapt to changing conditions, particularly when considering that a drone needed to traverse a real-world environment while avoiding obstacles.
“This adaptability is made possible by the causal underpinnings of our solutions. These flexible algorithms could one day aid in decision-making based on data streams that change over time, such as medical diagnosis and autonomous driving applications.”
Liquid Neural Networks and Autonomous Vehicles
Another test conducted by MIT examined how liquid neural networks could be used to help autonomous vehicles navigate. In this test, researchers equipped a car with a camera and computing units and then got human participants to drive the car.
The onboard cameras recorded the angle the humans held the steering wheel and passed them to a training platform, which taught the liquid neural network to map the steering wheel to the angle shown in the image. The network then used these camera inputs to autonomously steer the vehicle.
At a high level, this exercise demonstrated how liquid neural networks can be used to design neural controllers to help power an autonomous vehicle control system.
Benefits of LNNs
Liquid neural networks offer a number of core benefits. Some of these are:
- Real-time decision-making capabilities;
- The ability to process time series data;
- Respond quickly to a wide range of data distributions;
- Resilient and able to filter out anomalous or noisy data;
- More interpretability than a black-box machine learning algorithm;
- Reduced computational costs.
Challenges of LNNs
While liquid neural networks are very useful, they aren’t without their own set of unique challenges. These include:
- Struggle with processing static or fixed data
- Training difficulties due to elevated or vanishing gradients;
- Limitations in learning long-term dependencies due to gradient problems;
- Lack of extensive research into the functioning of liquid neural networks;
- The time-consuming parameter-tuning process;
- Challenges in processing static or fixed data.
Liquid neural networks are an important innovation due to their ability to help process time-series data and open the door to some exciting use cases in piloting drones and autonomous vehicles.