arXiv subjects

Daniel Bennequin

Publications and source records attributed to Daniel Bennequin.

13 recordsLinked to original sources

arXiv2022

Topos and Stacks of Deep Neural Networks

Every known artificial deep neural network (DNN) corresponds to an object in a canonical Grothendieck's topos; its learning dynamic corresponds to a flow of morphisms in this topos. Invariance structures in the layers (like CNNs or LSTMs) correspond to Giraud's stacks. This invariance is supposed to be responsible of the generalization property, that is extrapolation from learning data under constraints. The fibers represent pre-semantic categories (Culioli, Thom), over which artificial languages are defined, with internal logics, intuitionist, classical or linear (Girard). Semantic functioning of a network is its ability to express theories in such a language for answering questions in output about input data. Quantities and spaces of semantic information are defined by analogy with the homological interpretation of Shannon's entropy of P.Baudot and D.Bennequin in 2015). They generalize the measures found by Carnap and Bar-Hillel (1952). Amazingly, the above semantical structures are classified by geometric fibrant objects in a closed model category of Quillen, then they give rise to homotopical invariants of DNNs and of their semantic functioning. Intentional type theories (Martin-Loef) organize these objects and fibrations between them. Information contents and exchanges are analyzed by Grothendieck's derivators.

math.AT↗

arXiv2021

Logical Information Cells I

In this study we explore the spontaneous apparition of visible intelligible reasoning in simple artificial networks, and we connect this experimental observation with a notion of semantic information. We start with the reproduction of a DNN model of natural neurons in monkeys, studied by Neromyliotis and Moschovakis in 2017 and 2018, to explain how "motor equivalent neurons", coding only for the action of pointing, are supplemented by other neurons for specifying the actor of the action, the eye E, the hand H, or the eye and the hand together EH. There appear inner neurons performing a logical work, making intermediary proposition, for instance E V EH. Then, we remarked that adding a second hidden layer and choosing a symmetric metric for learning, the activities of the neurons become almost quantized and more informative. Using the work of Carnap and Bar-Hillel 1952, we define a measure of the logical value for collections of such cells. The logical score growths with the depth of the layer, i.e. the information on the output decision increases, which confirms a kind of bottleneck principle. Then we study a bit more complex tasks, a priori involving predicate logic. We compare the logic and the measured weights. This shows, for groups of neurons, a neat correlation between the logical score and the size of the weights. It exhibits a form of sparsity between the layers. The most spectacular result concerns the triples which can conclude for all conditions: when applying their weight matrices to their logical matrix, we recover the classification. This shows that weights precisely perform the proofs.

cs.AI↗

arXiv2021

Superfield equations in the Berezin-Kostant-Leites category

Using the functor of points, we prove that the Wess-Zumino equations for massive chiral superfields in dimension 4|4 can be represented by supersymmetric equations in terms of superfunctions in the Berezin-Kostant-Leites sense (involving ordinary fields, with real and complex valued components). Then, after introducing an appropriate supersymmetric extension of the Fourier transform, we prove explicitly that these supersymmetric equations provide a realization of the irreducible unitary representations with positive mass and zero superspin of the super Poincaré group in dimension 4|4.

math-ph↗

arXiv2020

Extra-fine sheaves and interaction decompositions

We introduce an original notion of extra-fine sheaf on a topological space, and a variant (hyper-extra-fine) for which Čech cohomology in strictly positive degree vanishes. We provide a characterization of such sheaves when the topological space is a partially ordered set (poset) equipped with the Alexandrov topology. Then we further specialize our results to some sheaves of vector spaces and injective maps, where extra-fineness is (essentially) equivalent to the decomposition of the sheaf into a direct sum of subfunctors, known as interaction decomposition, and can be expressed by a sum-intersection condition. We use these results to compute the dimension of the space of global sections when the presheaves are freely generated over a functor of sets, generalizing classical counting formulae for the number of solutions of the linearized marginal problem (Kellerer and Matúš). We finish with a comparison theorem between the Čech cohomology associated to a covering and the topos cohomology of the poset with coefficients in the presheaf, which is also the cohomology of a cosimplicial local system over the nerve of the poset. For that, we give a detailed treatment of cosimplicial local systems on simplicial sets. The appendixes present presheaves, sheaves and Čech cohomology, and their application to the marginal problem.

math.AT↗

arXiv2020

A functional equation related to generalized entropies and the modular group

We solve a functional equation connected to the algebraic characterization of generalized information functions. To prove the symmetry of the solution, we study a related system of functional equations, which involves two homographies. These transformations generate the modular group, and this fact plays a crucial role in solving the system. The method suggests a more general relation between conditional probabilities and arithmetic.

math.CA↗

arXiv2019

Topological Information Data Analysis

This paper presents methods that quantify the structure of statistical interactions within a given data set, and was first used in \cite{Tapia2018}. It establishes new results on the k-multivariate mutual-informations (I_k) inspired by the topological formulation of Information introduced in. In particular we show that the vanishing of all I_k for 2\leq k \leq n of n random variables is equivalent to their statistical independence. Pursuing the work of Hu Kuo Ting and Te Sun Han, we show that information functions provide co-ordinates for binary variables, and that they are analytically independent on the probability simplex for any set of finite variables. The maximal positive I_k identifies the variables that co-vary the most in the population, whereas the minimal negative I_k identifies synergistic clusters and the variables that differentiate-segregate the most the population. Finite data size effects and estimation biases severely constrain the effective computation of the information topology on data, and we provide simple statistical tests for the undersampling bias and the k-dependences following. We give an example of application of these methods to genetic expression and unsupervised cell-type classification. The methods unravel biologically relevant subtypes, with a sample size of 41 genes and with few errors. It establishes generic basic methods to quantify the epigenetic information storage and a unified epigenetic unsupervised learning formalism. We propose that higher-order statistical interactions and non identically distributed variables are constitutive characteristics of biological systems that should be estimated in order to unravel their significant statistical structure and diversity. The topological information data analysis presented here allows to precisely estimate this higher-order structure characteristic of biological systems.

stat.OT↗

arXiv2018

The Moon Illusion explained by the Projective Consciousness Model

The Moon often appears larger near the perceptual horizon and smaller high in the sky though the visual angle subtended is invariant. We show how this illusion results from the optimization of a projective geometrical frame for conscious perception through free energy minimization, as articulated in the Projective Consciousness Model. The model accounts for all documented modulations of the illusion without anomalies (e.g., the size-distance paradox), surpasses other theories in explanatory power, makes sense of inter- and intra-subjective variability vis-a-vis the illusion, and yields new quantitative and qualitative predictions.

q-bio.NC↗

arXiv2016

The unfolded universe of elementary particles. A geometric explanation of the standard model structure

We propose a geometric explanation of the standard model of Glashow, Weinberg and Salam for the known elementary particles. Our model is a generic Quantum Field Theory in dimension four, obtained by developing along a Lorentz sub-manifold the lagrangian of Einstein and Dirac in dimension twelve. The main mechanism which gives birth to the standard model is a certain gauge fixing of triality, which permits to identify the multiplicity of fermions, as seen from the four dimensional world, with the eight unseen dimensions of the generating universe. In this way we get the known tables of particles, explaining the series of fermions and the gauge bosons. We suggest that the Higgs field dynamics could appear through a bosonization of the right handed neutrino and correspond to a displacement in the unseen dimensions. We also propose hypotheses for dark matter, and perhaps dark energy. Then we suggest predictions to go beyond the standard model.

physics.gen-ph↗

arXiv2014

Parsimony, exhaustivity and balanced detection in neocortex

One fascinating aspect of the brain is its ability to process information in a fast and reliable manner. The functional architecture is thought to play a central role in this task, by encoding efficiently complex stimuli and facilitating higher level processing. In the early visual cortex of higher mammals, information is processed within functional maps whose layout is thought to underlie visual perception. The possible principles underlying the topology of the different maps, as well as the role of a specific functional architecture on information processing, is however poorly understood. We demonstrate mathematically here that two natural principles, local exhaustivity of representation and parsimony, would constrain the orientation and spatial frequency maps to display co-located singularities around which the orientation is organized as a pinwheel and spatial frequency as a dipole. This observation is perfectly in line with new optical imaging data on the cat visual cortex we analyze in a companion paper. Here we further focus on the theoretical implications of this structure. Using a computational model, we show that this architecture allows a trade-off in the local perception of orientation and spatial frequency, but this would occur for sharper selectivity than the tuning width reported in the literature. We therefore re-examined physiological data and show that indeed the spatial frequency selectivity substantially sharpens near maps singularities, bringing to the prediction that the system tends to optimize balanced detection between different attributes. These results shed new light on the principles at play in the emergence of functional architecture of cortical maps, as well as their potential role in processing information.

q-bio.NC↗

arXiv2014

Supersymmetric field equations from momentum space

It is known that every irreducible unitary representation of positive energy of the Poincaré group can be realized as a subspace of tensor fields on Minkowski spacetime subjected to suitable partial differential equations. We first describe geometrically the general mechanism that produces, via Fourier transform, the invariant differential operators corresponding to those representations. Then, using a super-version of the Fourier transform, we show explicitly how a massive irreducible unitary representation of the super Poincaré group in dimension $(4|4)$ can be realized as a linear sub-supermanifold of suitably constrained superfunctions. In this way, we obtain supersymmetric equations in terms of ordinary (non-Grassmannian) fields. Finally, using the functor of points, we show how our equations can be related in a natural way to the Wess-Zumino equations for massive chiral superfields.

math-ph↗

arXiv2014

Optimized Schwarz Waveform Relaxation for Advection Reaction Diffusion Equations in Two Dimensions

Optimized Schwarz Waveform Relaxation methods have been developed over the last decade for the parallel solution of evolution problems. They are based on a decomposition in space and an iteration, where only subproblems in space-time need to be solved. Each subproblem can be simulated using an adapted numerical method, for example with local time stepping, or one can even use a different model in different subdomains, which makes these methods very suitable also from a modeling point of view. For rapid convergence however, it is important to use effective transmission conditions between the space-time subdomains, and for best performance, these transmission conditions need to take the physics of the underlying evolution problem into account. The optimization of these transmission conditions leads to a mathematically hard best approximation problem of homographic type. We study in this paper in detail this problem for the case of linear advection reaction diffusion equations in two spatial dimensions. We prove comprehensively best approximation results for transmission conditions of Robin and Ventcel type. We give for each case closed form asymptotic values for the parameters, which guarantee asymptotically best performance of the iterative methods. We finally show extensive numerical experiments, and we measure performance corresponding to our analysis

math.NA↗

arXiv2012

Affine invariance of human hand movements: a direct test

Geometrical invariance, in particular affine invariance, has been recently proposed as an important principle underlying the production of hand movements. However, tests of affine invariance have traditionally been applied to the consequences of this principle rather to the principle itself. Here, we designed and performed an original, direct, test of affine invariance in a scribbling experiment. In each of the 10800 pairs of randomly-selected scribbling segments, we compared the time parameterizations obtained by transforming the first segment using four different transportation rules - affine, equi-affine, Euclidian and constant - with the experimentally-observed parameterization of the second segment. We observed that, when the two paths are affinely-similar, the affine transportation of the first segment yields the time parameterization that best matches the experimental parameterization of the second segment, which directly demonstrates the existence of affine invariance in the production of hand movements.

q-bio.OT↗

arXiv2009

Representants lagrangiens de l'homologie des surfaces projectives complexes

Using results by Donaldson and Auroux on pseudo-holomorphic curves as well as Duval's rational convexity construction, the paper investigates the existence of smooth Lagrangian surfaces representing 2-dimensional homology classes in complex projective surfaces. We prove that if the projective surface X is minimal, of general type, with uneven geometric genus, and has an effective, smooth and connected canonical divisor K, then there exists a non-empty convex open cone in the real 2-dimensional homology group H2(X) such that a multiple of every integral homology class in this cone can be represented by an embedded Lagrangian surface in X\K. A corollary asserts that such a surface X is of simple type in the sense of Kronheimer and Mrowka.

math.DG↗