top of page

Liquid Neural Networks

A New Approach to Neural Nets

Traditional neural networks are trained on a large dataset of data and then used to make predictions on new data. However, once a neural network is trained, its parameters are fixed. This means that the neural network cannot adapt to new information or changes in the environment.

Liquid neural networks are a new type of AI algorithm that can learn on the job, even after their training.

Liquid neural networks are inspired by how brain cells interact in the human brain. In the brain, synapses, which are the connections between neurons, are constantly changing. This allows the brain to adapt to new information and changes in the environment.

Liquid neural networks work in a similar way. The parameters of a liquid neural network are not fixed. Instead, they are constantly changing in response to new information. This allows liquid neural networks to learn and adapt even after they are trained.

Liquid neural networks have several advantages over traditional neural networks. First, they are more robust to changes in the environment. Second, they can learn from new information even after they are trained. Third, they are more efficient in terms of memory and computational resources.

One example of a liquid neural network is the Liquid State Machine (LSM). The LSM is a type of recurrent neural network that is inspired by the hippocampus, a region of the brain that is involved in learning and memory. The LSM is able to learn and adapt to new information by changing the weights of its connections.


The LSM has been used to solve a variety of problems, including:

  • It has been used to train robots to navigate complex environments.

  • It has been used to train language models that can understand and generate natural language.

  • It has been used to train image classifiers that can recognize objects in images.


Liquid neural networks are more robust, efficient, and adaptable than traditional neural networks. As research in liquid neural networks continues, it is likely that they will be used to solve a wider range of problems.

Here is a more detailed example of how liquid neural networks work:

Let's say we have a liquid neural network that is trained to recognize handwritten digits. The neural network has been trained on a dataset of images of handwritten digits, and it has learned to associate certain patterns of pixels with certain digits.

Now, let's say we give the neural network a new image of a handwritten digit. The neural network will start by analyzing the image and trying to find patterns that match the patterns it has learned from the training data. If it finds a match, it will predict the digit that it believes is in the image.

However, if the neural network does not find a perfect match, it will not simply give up. Instead, it will use its liquid parameters to adapt to the new image. The neural network will change its weights and biases in order to better match the new image. This process will continue until the neural network is confident in its prediction.

In this way, liquid neural networks are able to learn and adapt to new information even after they are trained. This makes them a powerful tool for solving a wide variety of problems.

bottom of page