Viterbi Faculty of Electrical Engineering, Technion
Spatial source separation in random networks of cortical neurons using manifold learning
Groups of neurons produce sequences of action-potentials (spikes) that propagate across extended parts of neuronal networks. These waves of spikes are believed to form neural representations (codes) that give rise to network functionality and behavior. Directed by sensory input or generated through interactions with a responding environment, these sequences are affected by processes acting at lower biophysical and biological levels. Cardinal processes in this regard are synaptic plasticity and cellular intrinsic excitability which exhibit substantial spontaneous changes over wide range of time scales. These processes are characterized by high degrees of variance and non-stationarity, in contrast to the stability and invariance needed to maintain coherent behaviors. Common techniques used to explore ‘neural codes’ and propagation patterns tend to be supervised with strict constraints and are rather task oriented. Therefore, they are limited in fully capturing the spatio-temporal network complexity. This study developed and applied a generic semi-supervised approach, named Rated-Time-Hidden-Manifold (RTHM), to explore and analyze neuronal activity by combining both rate-based and time-based representations paradigms to a single encoding scheme (Rated-Time scheme). Using this encoding, we find that we can expose low-dimensional, non-linear embeddings through manifold-learning dimensionality reduction techniques (e.g. Diffusion-Maps). Here, we use this approach to analyze the problem of source separation (also referred to as source identification), that is, identifying the source of input delivered to a neuronal network from the output (the network activity evoked by the input). To that end, we used cultured large-scale random networks of cortical neurons developing in-vitro on substrate-integrated Multi-Electrode-Arrays (MEA’s) as our experimental setting. We performed experiments in which periodic stimulations were delivered from different spatial sources while continuously recording the network activity. We then used RTHM to describe network activity through low-dimensional embeddings, extracted from the Rated-Time encoding. The resulted embeddings were successful in the source-separation task. Embeddings were then used to analyze intrinsic characteristics of networks, such as sensitivity of delay-in-response and similarity of responses originating from adjacent sources. Our method is comparable with results produced by supervised methods and is successful when additional temporal and spatial constraints are introduced. RTHM can be applied online as well as in closed-loop frameworks and can be utilized in additional tasks and various settings.
* MSc seminar under supervision of Prof. Shimon Marom and prof. Noam Ziv.