Continual learning on nano drones

Miniaturized flying robotic platforms, called nano-drones, have the potential to revolutionize the autonomous robots industry sector thanks to their very small form factor. The nano-drones’ limited payload only allows for a sub-100mW microcontroller unit for the on-board computations. Therefore, traditional computer vision and control algorithms are too computationally expensive to be executed on board these palm-sized robots, and we are forced to rely on artificial intelligence to trade off accuracy in favor of lightweight pipelines for autonomous tasks. However, relying on deep learning exposes us to the problem of generalization since the deployment scenario of a convolutional neural network (CNN) is often composed by different visual cues and different features from those learned during training, leading to poor inference performances. Our objective is to develop and deploy and adaptation algorithm, based on the concept of latent replays, that would allow us to fine-tune a CNN to work in new and diverse deployment scenarios. To do so we start from an existing model for visual human pose estimation, called PULPFrontnet, which is used to identify the pose of a human subject in space through its 4 output variables, and we present the design of our novel adaptation algorithm, which features automatic data gathering and labeling and on-device deployment. We therefore showcase the ability of our algorithm to adapt PULP-Frontnet to new deployment scenarios, improving the R2 scores of the four network outputs, with respect to an unknown environment, from approximately [−0.2, 0.4, 0.0,−0.7] to [0.25, 0.45, 0.2, 0.1]. Finally we demonstrate how it is possible to fine-tune our neural network in real time (i.e., under 76 seconds), using the target parallel ultra-low power GAP 8 System-on-Chip on board the nano-drone, and we show how all adaptation operations can take place using less than 2mWh of energy, a small fraction of the available battery power.

On Building a Memory Evolutive System for Application to Learning and Cognition Modeling

We address here aspects of the implementation of a memory evolutive system (MES), based on the model proposed by A. Ehresmann and J. Vanbremeersch (2007), by means of a simulated network of spiking neurons with time dependent plasticity. We point out the advantages and challenges of applying category theory for the representation of cognition, by using the MES architecture. Then we discuss the issues concerning the minimum requirements that an artificial neural network (ANN) should fulfill in order that it would be capable of expressing the categories and mappings between them, underlying the MES. We conclude that a pulsed ANN based on Izhikevich`s formal neuron with STDP (spike time-dependent plasticity) has sufficient dynamical properties to achieve these requirements, provided it can cope with the topological requirements. Finally, we present some perspectives of future research concerning the proposed ANN topology.

Learning the dynamics of embedded clauses

Recent work by Siegelmann has shown that the computational power of recurrent neural networks matches that of Turing Machines. One important implication is that complex language classes (infinite languages with embedded clauses) can be represented in neural networks. Proofs are based on a fractal encoding of states to simulate the memory and operations of stacks. In the present work, it is shown that similar stack-like dynamics can be learned in recurrent neural networks from simple sequence prediction tasks. Two main types of network solutions are found and described qualitatively as dynamical systems: damped oscillation and entangled spiraling around fixed points. The potential and limitations of each solution type are established in terms of generalization on two different context-free languages. Both solution types constitute novel stack implementations - generally in line with Siegelmann's theoretical work - which supply insights into how embedded structures of languages can be handled in analog hardware.

Contextual intelligent load management with ANN adaptive learning module

With the current increase of energy resources prices and environmental concerns intelligent load management systems are gaining more and more importance. This paper concerns a SCADA House Intelligent Management (SHIM) system that includes an optimization module using deterministic and genetic algorithm approaches. SHIM undertakes contextual load management based on the characterization of each situation. SHIM considers available generation resources, load demand, supplier/market electricity price, and consumers’ constraints and preferences. The paper focus on the recently developed learning module which is based on artificial neural networks (ANN). The learning module allows the adjustment of users’ profiles along SHIM lifetime. A case study considering a system with fourteen discrete and four variable loads managed by a SHIM system during five consecutive similar weekends is presented.

Hand gesture recognition for human computer interaction : a comparative study of different image features

Hand gesture recognition for human computer interaction, being a natural way of human computer interaction, is an area of active research in computer vision and machine learning. This is an area with many different possible applications, giving users a simpler and more natural way to communicate with robots/systems interfaces, without the need for extra devices. So, the primary goal of gesture recognition research is to create systems, which can identify specific human gestures and use them to convey information or for device control. For that, vision-based hand gesture interfaces require fast and extremely robust hand detection, and gesture recognition in real time. In this study we try to identify hand features that, isolated, respond better in various situations in human-computer interaction. The extracted features are used to train a set of classifiers with the help of RapidMiner in order to find the best learner. A dataset with our own gesture vocabulary consisted of 10 gestures, recorded from 20 users was created for later processing. Experimental results show that the radial signature and the centroid distance are the features that when used separately obtain better results, with an accuracy of 91% and 90,1% respectively obtained with a Neural Network classifier. These to methods have also the advantage of being simple in terms of computational complexity, which make them good candidates for real-time hand gesture recognition.

A behavior-based scheme using reinforcement learning for autonomous underwater vehicles

This paper presents a hybrid behavior-based scheme using reinforcement learning for high-level control of autonomous underwater vehicles (AUVs). Two main features of the presented approach are hybrid behavior coordination and semi on-line neural-Q_learning (SONQL). Hybrid behavior coordination takes advantages of robustness and modularity in the competitive approach as well as efficient trajectories in the cooperative approach. SONQL, a new continuous approach of the Q_learning algorithm with a multilayer neural network is used to learn behavior state/action mapping online. Experimental results show the feasibility of the presented approach for AUVs

Time Series Input Selection using Multiple Kernel Learning

In this paper we study the relevance of multiple kernel learning (MKL) for the automatic selection of time series inputs. Recently, MKL has gained great attention in the machine learning community due to its flexibility in modelling complex patterns and performing feature selection. In general, MKL constructs the kernel as a weighted linear combination of basis kernels, exploiting different sources of information. An efficient algorithm wrapping a Support Vector Regression model for optimizing the MKL weights, named SimpleMKL, is used for the analysis. In this sense, MKL performs feature selection by discarding inputs/kernels with low or null weights. The approach proposed is tested with simulated linear and nonlinear time series (AutoRegressive, Henon and Lorenz series).

Strong minimax lower bounds for learning

Minimax lower bounds for concept learning state, for example, thatfor each sample size $n$ and learning rule $g_n$, there exists a distributionof the observation $X$ and a concept $C$ to be learnt such that the expectederror of $g_n$ is at least a constant times $V/n$, where $V$ is the VC dimensionof the concept class. However, these bounds do not tell anything about therate of decrease of the error for a {\sl fixed} distribution--concept pair.\\In this paper we investigate minimax lower bounds in such a--stronger--sense.We show that for several natural $k$--parameter concept classes, includingthe class of linear halfspaces, the class of balls, the class of polyhedrawith a certain number of faces, and a class of neural networks, for any{\sl sequence} of learning rules $\{g_n\}$, there exists a fixed distributionof $X$ and a fixed concept $C$ such that the expected error is larger thana constant times $k/n$ for {\sl infinitely many n}. We also obtain suchstrong minimax lower bounds for the tail distribution of the probabilityof error, which extend the corresponding minimax lower bounds.

Statistical learning theory for geospatial data. Case study: Aral sea

In recent years there has been an explosive growth in the development of adaptive and data driven methods. One of the efficient and data-driven approaches is based on statistical learning theory (Vapnik 1998). The theory is based on Structural Risk Minimisation (SRM) principle and has a solid statistical background. When applying SRM we are trying not only to reduce training error ? to fit the available data with a model, but also to reduce the complexity of the model and to reduce generalisation error. Many nonlinear learning procedures recently developed in neural networks and statistics can be understood and interpreted in terms of the structural risk minimisation inductive principle. A recent methodology based on SRM is called Support Vector Machines (SVM). At present SLT is still under intensive development and SVM find new areas of application (www.kernel-machines.org). SVM develop robust and non linear data models with excellent generalisation abilities that is very important both for monitoring and forecasting. SVM are extremely good when input space is high dimensional and training data set i not big enough to develop corresponding nonlinear model. Moreover, SVM use only support vectors to derive decision boundaries. It opens a way to sampling optimization, estimation of noise in data, quantification of data redundancy etc. Presentation of SVM for spatially distributed data is given in (Kanevski and Maignan 2004).

Inter organizational learning within and by innovation networks

The thesis deals with the phenomenon of learning between organizations in innovation networks that develop new products, services or processes. Inter organizational learning is studied especially at the level of the network. The role of the network can be seen as twofold: either the network is a context for inter organizational learning, if the learner is something else than the network (organization, group, individual), or the network itself is the learner. Innovations are regarded as a primary source of competitiveness and renewal in organizations. Networking has become increasingly common particularly because of the possibility to extend the resource base of the organization through partnerships and to concentrate on core competencies. Especially in innovation activities, networks provide the possibility to answer the complex needs of the customers faster and to share the costs and risks of the development work. Networked innovation activities are often organized in practice as distributed virtual teams, either within one organization or as cross organizational co operation. The role of technology is considered in the research mainly as an enabling tool for collaboration and learning. Learning has been recognized as one important collaborative process in networks or as a motivation for networking. It is even more important in the innovation context as an enabler of renewal, since the essence of the innovation process is creating new knowledge, processes, products and services. The thesis aims at providing enhanced understanding about the inter organizational learning phenomenon in and by innovation networks, especially concentrating on the network level. The perspectives used in the research are the theoretical viewpoints and concepts, challenges, and solutions for learning. The methods used in the study are literature reviews and empirical research carried out with semi structured interviews analyzed with qualitative content analysis. The empirical research concentrates on two different areas, firstly on the theoretical approaches to learning that are relevant to innovation networks, secondly on learning in virtual innovation teams. As a result, the research identifies insights and implications for learning in innovation networks from several viewpoints on organizational learning. Using multiple perspectives allows drawing a many sided picture of the learning phenomenon that is valuable because of the versatility and complexity of situations and challenges of learning in the context of innovation and networks. The research results also show some of the challenges of learning and possible solutions for supporting especially network level learning.

Deep generative models for facial keypoints detection

A new area of machine learning research called deep learning, has moved machine learning closer to one of its original goals: artificial intelligence and general learning algorithm. The key idea is to pretrain models in completely unsupervised way and finally they can be fine-tuned for the task at hand using supervised learning. In this thesis, a general introduction to deep learning models and algorithms are given and these methods are applied to facial keypoints detection. The task is to predict the positions of 15 keypoints on grayscale face images. Each predicted keypoint is specified by an (x,y) real-valued pair in the space of pixel indices. In experiments, we pretrained deep belief networks (DBN) and finally performed a discriminative fine-tuning. We varied the depth and size of an architecture. We tested both deterministic and sampled hidden activations and the effect of additional unlabeled data on pretraining. The experimental results show that our model provides better results than publicly available benchmarks for the dataset.

Learning about brain physiology and complexity from the study of the epilepsies

The brain is a complex system, which produces emergent properties such as those associated with activity-dependent plasticity in processes of learning and memory. Therefore, understanding the integrated structures and functions of the brain is well beyond the scope of either superficial or extremely reductionistic approaches. Although a combination of zoom-in and zoom-out strategies is desirable when the brain is studied, constructing the appropriate interfaces to connect all levels of analysis is one of the most difficult challenges of contemporary neuroscience. Is it possible to build appropriate models of brain function and dysfunctions with computational tools? Among the best-known brain dysfunctions, epilepsies are neurological syndromes that reach a variety of networks, from widespread anatomical brain circuits to local molecular environments. One logical question would be: are those complex brain networks always producing maladaptive emergent properties compatible with epileptogenic substrates? The present review will deal with this question and will try to answer it by illustrating several points from the literature and from our laboratory data, with examples at the behavioral, electrophysiological, cellular and molecular levels. We conclude that, because the brain is a complex system compatible with the production of emergent properties, including plasticity, its functions should be approached using an integrated view. Concepts such as brain networks, graphics theory, neuroinformatics, and e-neuroscience are discussed as new transdisciplinary approaches dealing with the continuous growth of information about brain physiology and its dysfunctions. The epilepsies are discussed as neurobiological models of complex systems displaying maladaptive plasticity.

Comparison of classification ability of hyperball algorithms to neural network and k-nearest neighbour algorithms

The main focus of this thesis is to evaluate and compare Hyperbalilearning algorithm (HBL) to other learning algorithms. In this work HBL is compared to feed forward artificial neural networks using back propagation learning, K-nearest neighbor and 103 algorithms. In order to evaluate the similarity of these algorithms, we carried out three experiments using nine benchmark data sets from UCI machine learning repository. The first experiment compares HBL to other algorithms when sample size of dataset is changing. The second experiment compares HBL to other algorithms when dimensionality of data changes. The last experiment compares HBL to other algorithms according to the level of agreement to data target values. Our observations in general showed, considering classification accuracy as a measure, HBL is performing as good as most ANn variants. Additionally, we also deduced that HBL.:s classification accuracy outperforms 103's and K-nearest neighbour's for the selected data sets.

Apprentissage de représentations sur-complètes par entraînement d’auto-encodeurs

Les avancés dans le domaine de l’intelligence artificielle, permettent à des systèmes informatiques de résoudre des tâches de plus en plus complexes liées par exemple à la vision, à la compréhension de signaux sonores ou au traitement de la langue. Parmi les modèles existants, on retrouve les Réseaux de Neurones Artificiels (RNA), dont la popularité a fait un grand bond en avant avec la découverte de Hinton et al. [22], soit l’utilisation de Machines de Boltzmann Restreintes (RBM) pour un pré-entraînement non-supervisé couche après couche, facilitant grandement l’entraînement supervisé du réseau à plusieurs couches cachées (DBN), entraînement qui s’avérait jusqu’alors très difficile à réussir. Depuis cette découverte, des chercheurs ont étudié l’efficacité de nouvelles stratégies de pré-entraînement, telles que l’empilement d’auto-encodeurs traditionnels(SAE) [5, 38], et l’empilement d’auto-encodeur débruiteur (SDAE) [44]. C’est dans ce contexte qu’a débuté la présente étude. Après un bref passage en revue des notions de base du domaine de l’apprentissage machine et des méthodes de pré-entraînement employées jusqu’à présent avec les modules RBM, AE et DAE, nous avons approfondi notre compréhension du pré-entraînement de type SDAE, exploré ses différentes propriétés et étudié des variantes de SDAE comme stratégie d’initialisation d’architecture profonde. Nous avons ainsi pu, entre autres choses, mettre en lumière l’influence du niveau de bruit, du nombre de couches et du nombre d’unités cachées sur l’erreur de généralisation du SDAE. Nous avons constaté une amélioration de la performance sur la tâche supervisée avec l’utilisation des bruits poivre et sel (PS) et gaussien (GS), bruits s’avérant mieux justifiés que celui utilisé jusqu’à présent, soit le masque à zéro (MN). De plus, nous avons démontré que la performance profitait d’une emphase imposée sur la reconstruction des données corrompues durant l’entraînement des différents DAE. Nos travaux ont aussi permis de révéler que le DAE était en mesure d’apprendre, sur des images naturelles, des filtres semblables à ceux retrouvés dans les cellules V1 du cortex visuel, soit des filtres détecteurs de bordures. Nous aurons par ailleurs pu montrer que les représentations apprises du SDAE, composées des caractéristiques ainsi extraites, s’avéraient fort utiles à l’apprentissage d’une machine à vecteurs de support (SVM) linéaire ou à noyau gaussien, améliorant grandement sa performance de généralisation. Aussi, nous aurons observé que similairement au DBN, et contrairement au SAE, le SDAE possédait une bonne capacité en tant que modèle générateur. Nous avons également ouvert la porte à de nouvelles stratégies de pré-entraînement et découvert le potentiel de l’une d’entre elles, soit l’empilement d’auto-encodeurs rebruiteurs (SRAE).

Algorithmes d'apprentissage pour la recommandation

L'ère numérique dans laquelle nous sommes entrés apporte une quantité importante de nouveaux défis à relever dans une multitude de domaines. Le traitement automatique de l'abondante information à notre disposition est l'un de ces défis, et nous allons ici nous pencher sur des méthodes et techniques adaptées au filtrage et à la recommandation à l'utilisateur d'articles adaptés à ses goûts, dans le contexte particulier et sans précédent notable du jeu vidéo multi-joueurs en ligne. Notre objectif est de prédire l'appréciation des niveaux par les joueurs. Au moyen d'algorithmes d'apprentissage machine modernes tels que les réseaux de neurones profonds avec pré-entrainement non-supervisé, que nous décrivons après une introduction aux concepts nécessaires à leur bonne compréhension, nous proposons deux architectures aux caractéristiques différentes bien que basées sur ce même concept d'apprentissage profond. La première est un réseau de neurones multi-couches pour lequel nous tentons d'expliquer les performances variables que nous rapportons sur les expériences menées pour diverses variations de profondeur, d'heuristique d'entraînement, et des méthodes de pré-entraînement non-supervisé simple, débruitant et contractant. Pour la seconde architecture, nous nous inspirons des modèles à énergie et proposons de même une explication des résultats obtenus, variables eux aussi. Enfin, nous décrivons une première tentative fructueuse d'amélioration de cette seconde architecture au moyen d'un fine-tuning supervisé succédant le pré-entrainement, puis une seconde tentative où ce fine-tuning est fait au moyen d'un critère d'entraînement semi-supervisé multi-tâches. Nos expériences montrent des performances prometteuses, notament avec l'architecture inspirée des modèles à énergie, justifiant du moins l'utilisation d'algorithmes d'apprentissage profonds pour résoudre le problème de la recommandation.