Algorithmic Information Dynamics (or Algorithmic Dynamics in short) is thus a new type of discrete calculus based on computer programming to study causation by generating mechanistic models to help find first principles for physical phenomena building up the next generation of data analytics in what we call Algorithmic Machine Learning. Teaming up with immunologists, bioinformaticians, toxicologists, oncologists, cognitive scientists and molecular biologists, our lab applies all these mathematical ideas and fundamental findings to, among other areas: behavioural, evolutionary and molecular biology.
Contributions
We aim at connecting computability theory, algorithmic information and dynamical systems to infer causation and generating mechanisms of synthetic and biological systems helping to produce models and reprogramming living systems.
- We have introduced the field of Algorithmic Information Dynamics [see J31, J32, J30, J29 and online course] and an algorithmic causal calculus which contributes to:
- the study of dynamical systems from a data-model-driven approach instead of data-statistics-approach
- the reconstruction of space-time dynamics of systems
- the manipulation and reprogramming of dynamical systems at the level of generative models applied both to synthetic and biological relevant examples ([see J17, P11, J4, online and online].)
- the identifying breaking points in static and dynamical systems such as graphs and networks
- a fundamental new Maximum Randomness/Entropy Principle MaxEnt replacing entropy by a hybrid method able to take into consideration algorithmic probability accounting for recursive non-recursiveness hence truly for random model methods.
- qualitatively reconstruct attractor landscapes e.g. the reconstruction of cell differentiation order.
- We also introduced a significant methodological and numerical improvement and a precise calculation pipeline to infer and reverse engineer networks based on adaptive differentiation [see paper J27].
- We have introduced novel and more powerful measures to what we call algorithmic machine learning in two areas of specific application (and more are coming), such as:
- methods for minimization of dimensions both for network (sparsification) and general data dimensionality reduction (see J33)
- generative mechanism deconvolution (see J34).
- ways to combine symbolic computation, statistical analysis and differentiable programming (e.g. deep learning)
- We have shown that algorithmic probability may explain phenomenology associated to biological evolution, in particular accelerating convergence rates, promoting memory and producing modularity reminiscent of genetic code, and making massive extinctions a mechanism of evolution and thus potentially providing an intrinsic explanation for them (see J30).
- We have shown that indexes based on Shannon Entropy cannot be considered robust measures of randomness especially in the context of maximal entropy models [see J25, P12, J29 and J34].
- We have shown that computability, causality and dynamical systems are deeply connected by way of algorithmic probability [J, 22, P7, J4].
- We have introduced a minimal mathematical model for Multiple Sclerosis that displays the most important stages and dynamic features of the disease [J24].
- We have shown how natural selection may harness computation and how algorithmic dynamics can explain aspects of biological evolution and , such as speed of convergence, natural mass extinctions and diversity explosions, among others. By taking organisms moving in software space we have always defined a concept of back mutations, this is mutations that are algorithmically more likely to occur because they are in the computationally evolutionary pathway of the system. [online]
- We have shown how our indexes can help in challenges of structural biology such as nucleosome occupancy/positioning. [online]
For more information, visit Publications.
We have produced these videos to help explain our research interests and contributions:
|
|
These videos do not, of course, contain all the details and information that only reading the full papers can provide.
And a teaser of an online course (MOOC) in preparation to come
out in the Sta Fe Institute's Complexity Explorer platform:
out in the Sta Fe Institute's Complexity Explorer platform: