- Researchers use machine learning algorithm (originally developed for spam filtering) to analyze Hydra’s behavior.
- It can analyze behavior in real time and show how Hydra’s neurons respond to varying environments.
- To do this, they applied bag-of-words classification model to the video containing all Hydra’s activities.
We have been studying animals’ behavior for centuries. It involves a lot of detailed observations and painstaking attention. But there are some efficient techniques that can automate the identification and classification process.
Recently, scientists at Columbia University demonstrated how a spam-filtering algorithm could be used to analyze animals’ behavior. They built an automatic behavior analysis pipeline that learns to pick out (from video) the complete behavioral repertoire of Hydra.
Hydra is a freshwater organism with regenerative ability – they don’t die of old age. They don’t have any brain, but hundreds of neurons run along their translucent body coordinating multiple behaviors.
They behave in a predictable manner and by comparing their behaviors to their neuron-firing, scientists could understand how nervous system of complex animals function.
Artificial Intelligence has been used to partly examine how worms crawl, how some flies fly, but this is the first time researchers are using machine learning to describe the behavior of an animal.
The Machine Learning Algorithm
The algorithm can analyze the behavior in realtime – it helps researchers to observe if Hydra can learn anything, and if they do, how their neurons respond.
In 2017, researchers found 4 types of neural circuits responsible for controlling 4 different bending and elongation behaviors. This helped them to understand how nervous system of Hydra controls its behavior.
Now they have gone a step further: they have cataloged a complete set of Hydra’s behaviors.
The team applied well-known bag-of-words model to the video that contains all activities of Hydra. The model is simplifying representation used in information retrieval and natural language processing. Moreover, researchers identified unsupervised methods and unannotated behaviors.
Neurons of Hydra shown as a green fluorescence indicator | Credit: Columbia University
The bag-of-words model considers videos and images as “bags” of visual words, like small patches in the pictures, or shape and video features extracted from such small patches. Compared to other methods, it’s more robust against challenges like orientation, occlusion and change in viewing angle.
In order to make it more efficient, researchers integrated this model with other computational methods, including dense trajectory (encodes shape and motion stats), body part segmentation (describes spatial information) and Fisher vectors (represents visual words in statistical way).
The algorithm cycled through the hours of video and detected repetitive moves, just like it examines how many times words appear in a text-body to pick out topics/subject and flag the email.
Credit:Yuste Lab / Columbia University
10 of the previously reported behaviors were successfully identified by the algorithm. In fact, it evaluated 6 of those behaviors that responded to varying surrounding scenarios. The results were quite interesting, the behavior of Hydra hardly altered. Whether you kept the lights off or on, fed it or not, it repeated the same thing several times.
Researchers plan to experiment with stimuli to observe any changes in Hydra’s behavior. The ultimate goal is to uncover the neural code that reveals how Hydra’s neural networks produce behavior. In future, the technique can also be infused with other organisms that have evolved over thousands of centuries.
Things learned from this study could be helpful to other engineering branches concerned with maintaining accurate control and stability in machines like planes and ships, navigating in harsh environments.