From Liquid State Machine to First Steps with Cortical Neurons
Development in recurrent neuronal networks led to the discovery that randomly connected nonlinear function units can compute a large variety of nonlinear functions on the input. With a sufficiently large variety of such units, it is theoretically possible to perform mathematical operations like classification or function mapping for applications such as speech recognition or computer vision.
One particular kind of a recurrent network is known as a Liquid State Machine (LSM). Each unit in this network computes one specific nonlinear function with feedback, like a Leaky Integrate-and-Fire Neuron Model.
Recent studies have implemented LSMs using various “features” that originated from biological networks. This biology inspired modifications enabled the LSM to be both robust to noise and hardware failure as well as to compute complex continuous real-valued patterns without the need for discretization and digital encoding. This model effectively simulates neuronal network activity and is, thus, used to study the emergence of representational capacities in random and modular networks and to explore the underlying biophysical and biological processes that give rise to these capacities. However, one major drawback of the LSMs is their inability to learn and adapt in real-time.
Inspired by the complex behavior of the LSM and its ability to borrow properties directly from biology, I have developed an experimental setup that combines generic ex-vivo networks of neurons, neuron-computer interfaces and electrophysiology, instead of the simulated neurons in the liquid in the LSM. The aim of this system is to be adaptive to neuronal activity in real time. It is based on established experimental techniques that will allow for long-term closed-loop experiments. This real time closed-loop system serves as a proxy of the environment in the sense that it will provide both electrical and “neuromodulatory” input and, more importantly, will allow for “feedback” from and interactions with the simulated “environment”.
Preliminary results indeed show that when ex-vivo cultures are stimulated with various frequencies of stimulations with combination with neuro modulation, their activity can be differentiated in real-time using standard machine learning techniques and, thus, provide the first step in closing the loop with the environment.