Home

LOGO_FUTURE

BAMBI ended at the end of year 2016

You may read the Public BAMBI Final Report

The BAMBIĀ (Bottom-up Approaches to Machines dedicated to Bayesian Inference) project is a EU collaborative FET Project ( FP7-ICT-2013-C, project number 618024)

We proposed a theory and a hardware implementation of probabilistic computation inspired by biochemical cell signalling.

 

Scheme of BAMBI organization. The three main axes (theory of probabilistic computation, probabilistic inference in biology, and hardware implementation) are oriented according to our bottom-up approach. The concentric circles illustrate cross-fertilization through inter-disciplinary tasks and transverse milestones.

Scheme of BAMBI organization. The three main axes (theory of probabilistic computation, probabilistic inference in biology, and hardware implementation) are oriented according to our bottom-up approach. The concentric circles illustrate cross-fertilization through inter-disciplinary tasks and transverse milestones.

In the BAMBI project, our ambition was to lay the foundations of new computing machines, largely inspired by biology, and specially designed to efficiently perform probabilistic inferences. For that, our main effort focused on three research axes: to develop a new theory of probabilistic computation, to investigate how simple living organisms process information and perform basic inferences, to build completely new hardware based on these principles.

In the first axis, we first developed a Bayesian algebra and we proposed the Bayesian gates that can be seen as the extensions of the Boolean algebra and logical gates, the essential components of current computers. Then, we detailed and simulated new architectures capable to solve probabilistic inference problems of various complexity. A key feature of these architectures is to rely on stochastic computation and on components that behave randomly: an essential characteristic of bio-molecules participating to cell signaling, as well as of new magnetic nanocomponents. The numerous simulations that we performed during the project showed that these new architectures could solve very efficiently several inference problems, including sensor fusion, Bayesian filters, parameter learning and approximation of highly intractable inference problems.

In the second axis, we aimed to establish the link between some general principal governing biological signal processing at various space-time scales and the basic principles of probabilistic computation. The experimental model we used for that is a micro-alga, Chlamydomonas reinhardtii, on which we recorded numerous genomic, metabolic and behavioral data. We showed how a relatively simple stochastic model can account for the capacity of these micro-organisms to swim toward or away from a light source in order to adapt to their current energy needs. The surprisal analysis was used to identify in various experimental dataset the balance state and the additional constraints required to quantify the observed deviations and dynamical evolutions with respect to the balance state. Then, we developed the analytical theory for characterizing how a time evolving system adjusts according to Le Chatelier’s Principle specifically when it is not in equilibrium. Finally, we successfully developed probabilistic inference model for biochemical kinetics cast in the form of a nonlinear finite state machine which is massively parallel.

In the third axis, we followed two tracks. The first track is for short-term, it is based on the use of existing fully integrated circuits, namely the reconfigurable array of logical elements (FPGA). The long-term track is based on existing components, the super paramagnetic tunnel junctions (SMTJs) which behave stochastically at room temperature, but are not yet massively integrated in existing circuits, except for their promising applications in non-volatile memory chips. The short-term track was very important for us, since it allowed today to produce real circuits implementing the new architectures simulated in the first axis. The long-term track is very promising since, as we have shown, the energy efficiency will be several orders of magnitude better than one could expect with existing hardware. Considering the possible future hardware developments, we have identified two critical issues: the first one is the capacity to integrate a large number of entropy sources at low energy cost, the second one is the capacity to integrate local memories within the basic processing units in order to avoid the well-known von Neumann bottleneck which results from the physical separation of memory and processing units.

Our project was highly interdisciplinary. As such, it challenges mainstream ideas in computer science, biology and hardware industry. We successfully adopted complementary approaches in all our research axes, and all partners in our consortium benefit from intensive scientific interactions. The project as a whole has produced more than 50 publications and 3 patents. We are now considering future academic research projects and industrial developments, based on the numerous promising results of BAMBI.