Implantable devices for the stimulation of the central and peripheral nervous systems (NS) may lead to a revolution in the treatment of a variety of diseases. Ideally, brain dysfunctions and disorders in organs can be treated by a direct interaction with central neurons and peripheral nerves. Within an adaptive closed-loop, the intelligent controller recognizes pathological activity patterns and adapts stimulation accordingly to restore function. Importantly, while stimulation is always operated at the level of the NS, sensing may occur also in other organs (e.g. monitoring glucose concentration in the blood).
This is the vision, but to make it real, novel implantable devices must be conceived. Precise control of neuronal discharge and 'intelligent' operation, i.e. by smartly adapting stimulation to the varying neuronal responses must be realized. In essence, these devices must emulate and support the NS in supervising smart controls.
The goal of CIDES is to capture the neural dynamics of predictive decision-making and to reproduce this dynamic in a computational model. This information will be ready to be used in advanced neuromorphic devices able to generate efficient decision-making strategies in closed loop protocols. To this end, we will i) carry out neurophysiology recordings from multiple brain areas of animal models performing predictive decision-making tasks; ii) include modulation in the network neural activity after stimulation of the subthalamic nucleus; iii) develop an adaptive multi-area spiking neural network model that reproduces the experimental results.
The project results will advance our understanding of the brain's control processes and have a great potential impact on the electronics and biomedical European industry, paving the way to a new generation of brain-inspired processors (smart decision makers) for an efficient autonomous control of brain computer interfaces including closed-loop brain stimulators.
The concept of adaptive smart control developed within CIDES can be easily extended to the whole domain of brain machine interfaces (BMI), in fact posing a crucial technological challenge: that is, how to develop a platform supporting the development of 'Intelligent' neurostimulation implants that are adaptive, precise and that can be customized for various therapy applications. We propose to contribute to a ground-breaking leap forwards in this direction, taking advantage of miniaturized electrodes and multisite recordings we will provide fundamental information for building up brain-inspired - i.e., based on spiking neurons - neuromorphic processing units (NPUs), that are 'plastic' and reconfigurable, and that are capable to 'intelligently' adapt stimulation responses by reinforcement learning to changes of biological signals in input. In the ideal situation, we propose to pair such NPUs to high-resolution and large-scale interfaces and develop suitable information processing algorithms to extract the maximum of relevant information from recorded signals and to precisely tune neurostimulation and its effect on neurons and nerves.
Among the most ambitious goals of CIDES is the gathering of sufficient experimental data to instantiate models and computer simulations on the dynamic of functional states in neural networks during a decision-making process. Integrate and fire models of cortical decision-making have been extensively studied in the past decade [17] to become one of the most advanced fields in computational neuroscience [18]. Still, the domain of application of these models has been mainly bounded to reproduce and explain decision-making for discriminating ambiguous external stimuli [19], and applications to the case of unpredictable conditions have been rare and limited to simple models [15]. A second limitation is that the complex architecture of the pathway leading to decision-making has usually been summarized into a single bi-stable network with inputs and outputs, while recent advancements show that inter-network interaction can significantly alter the whole process [20]. For the first time we will integrate two learning timescales: we will modify the classic accumulation of the evidence model [17] we will implement spike time dependent plasticity in order to achieve optimal synaptic weights for decision-making [21]. Our model will be the first to encompass the whole architecture responsible of decision-making in the cortex in the case of unpredictable conditions useful for whatever adaptive controller. We will analyse, characterize and model the specific learning processes taking place in each of the areas involved and in the interactions among them to achieve an accurate description of multiscale hierarchical learning in decision-making.
Understanding how to reproduce robust and reliable decision-making behavior in neuromorphic systems will be crucial for developing a new class of non von Neumann information processing architectures that will exploit ultra-low power sub-threshold analog circuits and be able to produce robust computation with inhomogeneous and unreliable components. Indeed, much effort is currently being invested in the quest for developing new computing paradigms for information and communication technologies (ICT).
The very recent interest for BMI and smart machines by Mark Zuckerberg, founder of Facebook, and by Elon Musk, founder of Tesla, Inc., suggest that extremely high levels of electronic integration and substantial (cloud) computing resources might become soon available for biomedical devices and BMIs, as they are already for our mobile phones. Designing novel algorithms and developing the know-how on the combination of brain signals represent right now one of the most important priority to enable innovation in the medical (and possibly the consumer) domains.
1. Fox D (2017) Nature 545:20-22
2. Lebedev MA and Nicolelis MAL (2017) Physiol Rev. 97:767-837
3. Giulioni M, et al. (2015) Sci Rep. 14;5:14730
4. Benabid AL et al. (1991) Lancet 337, 403-40610
5. Penry JK & Dean JC (1990) Epilepsia. 31 Suppl 2:S40-3
6. Sternson SM & Roth BL (2014) Annu Rev Neurosci. 37:387-407
7. Wichmann T & DeLong MR (2016) Neurotherapeutics. 13:264-83
8. Wichmann T, et al.(1994) J Neurophysiol. 72:521-30
9. Sun FT & Morell MJ (2014) Expert Rev. Med. Devices 11:563572
10. Little S et al (2013) Ann Neurol. 74:449-57
11. Rosin B et al (2011) Neuron 72:370-84
12. Wichmann T & Soares J (2006) J. Neurophysiol. 95:2120-33
13. Bevan MD et al.(2002) TINS 25:525-31
14. Neumann W-J et al.(2016) Neuromodulation 19:20-24
15. Marcos E et al (2013). Neuron 78:249-255
16. Mattia M et al (2013) J Neurosci 33:11155-68
17. Wang, X-J (2002). Neuron 36:955-68
18. Gerstner W et al (2012). Science 338: 60-65
19. Shadlen MN et al (2013). Neuron 80: 791-806
20. Wimmer K et al (2015). Nat. Commun. 6:6177.7
21. Lo C-C, & Wang X-J (2006). Nat. Neur. 9:956963