A long-term objective is to use RCN as a substitute (BACKUP, therefore the name of the project) of damaged neural tissue. In this case, the RCN has to be designed with a proper topology that meets the final task to accomplish. Such type of optimization would boost the use of RCN in therapeutic applications.
High gain/high-risk balance In BACKUP, I am moving at the forefront of the integrated photonic technology and of the interdisciplinary science of the interface between living and condensed matter. Only now, large networks of photonic components are being integrated with electronic driving circuits. Few successful optogenetic experiments have been carried out where the light signal is provided by remote lasers through optical fibres.
RCN with few photonic components have only recently being demonstrated. Therefore, all the activities I am planning have a very high-risk content. However, the success of BACKUP will open definitely new scenarios in computing by allowing the concurrent and simultaneous use of three profoundly different platforms for computation (photonics, electronics and neuronics). I will leap forward in the understanding of our brain by unveiling the relationship between connectivity –topology- of nodes and the functions they perform as a complex single network. I will unfold new possibilities in neurological disorder therapies by supplementing the faulting functions with artificial networks that replace or supplement the nerves or the brain tissues.
Artificial Intelligence (AI) is a Computer Science discipline whose goal is to create artificial systems (e.g., algorithms) which are able to perform intelligent tasks such as image understanding, language translation or problem solving. AI systems are based on many different paradigms. Some of these paradigms are directly inspired to the human brain. For instance, Artificial Neural Networks (ANNs) are a mathematical simulation of a biological neural network. Broadly speaking, an ANN is a graph, where nodes simulate neurons (and are associated with activation functions) and edges simulate synaptic connections. Importantly, the ANN weights, associated with the edges, simulate the synaptic strength of the connections, which may vary during time.
The main characteristic of ANNs is their ability to automatically learn from a set of training samples. For instance, in order to learn to recognize a chair into an input image, a collection of images of chairs is progressively presented to the ANN. Specific training algorithms modify the ANN weights trying to minimize the recognition error on the training set. When training is done, the implicit knowledge, contained in the training set, has been acquired by the ANN and memorized in its weight values. Now the ANN can recognize new, unseen images of chairs.
Recently, many AI fields have been significantly improved due to the introduction of Deep ANNs (DNNs), which basically are ANNs composed of a (deep) cascade of layers, each layer being characterized by a set of homogeneous neurons which embed part of the ANN general knowledge.
In Backup we will use DNNs in order to study and simulate real, biological networks. The idea is sketched in Fig. 1. B is a (small) biological network grown in vitro. Using specific photonic circuits and optogenetics, we can stimulate and read the activation state of each individual neuron in B. We can thus collect a virtually unlimited dataset of pairs stimuli-responses of B. Once this dataset has been collected, it will be used to train the artificial network A. A is trained in order to predict the response of B given a specific stimulus. Basically, A is trained to “think” as B “thinks”. If successful, this experiment will show, for the first time, that the “memories” of a biological network can be artificially reproduced using external hardware and software.
Even more intriguing and challenging is the possibility of building an hybrid computational system, which will be explored in the second part of the project. Specifically, parts of B will be replaced with a DNN, and we will address this question: “Can a computational task be jointly performed by the cooperation of two networks: an artificial and a biological network, connected to each other?”
Figure 1: A schematic representation of an ANN (A) and a biological network (B)
Backup's project logo recap the interaction between a biological brain network and artificial networks we are going to study and to make interact. On the one hand Backup aims to learn from biological systems how to design efficient networks of interacting neurons, on the other hand BACKUP wants to create artificial networks able to substitute damaged tissues.
Building an in vitro memory engram using light
The brain is an intricate network formed by many neurons highly interconnected via synaptic contacts (synapses). Once formed during development, the neuronal circuitry can be modified by neuronal activity and these changes are directly correlated to cognitive processes such as learning and memory. The term memory is referred to the storage of information in the brain, a cognitive function essential for learning and interaction with the external world.
In particular, information are believed to be stored in the brain as physical persistent modifications of ensemble of neurons (engram). These permanent changes involve a reinforcement of synaptic connections among specific neurons that are activated during the encoding of a memory trace.
Only recently, with the availability of optogenetic techniques, it become possible to identify single memory traces in the brain. Optogenetics is a biological method that involves genetic modification of neurons to express light-sensitive membrane channels (i.e. channel rhodopsin) allowing to control neuronal activity by light. Channel rhodopsin can be specifically expressed in neurons activated during learning of a memory trace and scientists demonstrated that is possible to retrieve memory in mice by re-activating these neurons with light.
The aim of BACKUP is to create a memory engram in a small in vitro neuronal network using a photonic chip (Fig. 2). This artificial engram will be created using optogenetic strategies in which patterned-light illumination will correspond to activation of group of interconnected neurons expressing channel rhodopsin along the light path. In this sense patterned light will work as an artificial learning event for generating memory engram.
This reductionist in vitro system will allow comparison of the activity and morphological changes in “engram neurons” (cells activated by light) vs “non engram neurons” (cells not activated by light) to study basic mechanisms of engram cells connectivity and to establish a link between neuronal activity and changes in the connectivity among neurons.
It is well established that some pathological conditions such as memory loss (amnesia) could be associated with a damaged memory engram or with the inability to retrieve it. The hybrid system developed in BACKUP will also give us the opportunity to easily modify or eliminate the artificial engram in order to better understand basic mechanisms underlying memory disfunctions.
Figure 2 In vitro neuronal network grown on a photonic chip.