Rétropropagation du gradient | EPFL Graph Search

En intelligence artificielle, plus précisément en apprentissage automatique, la rétropropagation du gradient est une méthode pour entraîner un réseau de neurones. Elle consiste à mettre à jour les poids de chaque neurone de la dernière couche vers la première. Elle vise à corriger les erreurs selon l'importance de la contribution de chaque élément à celles-ci. Dans le cas des réseaux de neurones, les poids synaptiques qui contribuent plus à une erreur seront modifiés de manière plus importante que les poids qui provoquent une erreur marginale. De façon abusive, on appelle souvent « technique de rétropropagation du gradient » l'algorithme classique de correction des erreurs reposant sur le calcul du gradient grâce à la rétropropagation. C'est cette méthode qui est présentée ici. La correction des erreurs peut se faire selon d'autres méthodes, par exemple le calcul de la dérivée seconde. Ce principe fonde les méthodes de type algorithme du gradient, qui sont utilisées dans des réseau

Advanced Techniques in Optical Diffraction Tomography

Ahmed Bassam Sayed Ayoub Mohamed Emam

In this thesis, we study the 3 challenges described above. First, we study different reconstruction techniques and assess the fidelity of each reconstruction results by means of structured illumination and phase conjugation. By reconstructing the 3D refractive index of the sample using different algorithms (i.e. Born, Rytov, and Radon) and then perform a numerical back-propagation of experimentally measured structured illumination pattern we are able to assess the fidelity of each reconstruction algorithms without prior information about the 3D RI distribution of the sample.The second part of the thesis is concerned with the 3D reconstruction of samples using intensity-only measurements which the need to holographically acquire them. We show that using intensity-only measurements, we could still be able to reconstruct the 3D volume of the sample with edge-enhanced effects which was proven useful for drug delivery applications in which nano-particles were identified on the cell membrane of immune T-cells in a drug delivery studies. Such reconstruction technique would result in more robust imaging system where the commercial imaging microscope systems can be incorporated with LEDs for high-quality speckle noise-free imaging systems. In addition, we show that under certain conditions, we can be able to reconstruct the 3D refractive index distribution of different samples.The third part of the thesis is contributing to high-speed complex wave-front shaping using DMDs. In that part, new modulation technique is demonstrated that can boost the speed of the current time-multiplexing techniques by a factor of 32. The modulation technique is based on amplitude modulation where an amplitude modulator is synchronized withvthe DMD to modulate the intensity of each bit-plane of an 8-bit image and then all the modulated bit-planes are linearly added on the detector. Such modulation technique can be used not only for structured illumination microscopy but also for high-speed 3D printing applications as well as projectors.The last part is concerned with using deep learning approaches to solve the missing cone problem usually accompanied with optical imaging due to the limited numerical aperture of the imaging system. Two techniques are discussed; the first is based on using a physical model to enhance the quality of the 3D RI reconstruction and the second is based on using deep neural network to solve the missing cone problem.

EPFL2022

Advances In Morphological Neural Networks: Training, Pruning And Enforcing Shape Constraints

Nikolaos Dimitriadis

In this paper, we study an emerging class of neural networks, the Morphological Neural networks, from some modern perspectives. Our approach utilizes ideas from tropical geometry and mathematical morphology. First, we state the training of a binary morphological classifier as a Difference-of-Convex optimization problem and extend this method to multiclass tasks. We then focus on general morphological networks trained with gradient descent variants and show, quantitatively via pruning schemes as well as qualitatively, the sparsity of the resulted representations compared to FeedForward networks with ReLU activations as well as the effect the training optimizer has on such compression techniques. Finally, we show how morphological networks can be employed to guarantee monotonicity and present a softened version of a known architecture, based on Maslov Dequantization, which alleviates issues of gradient propagation associated with its "hard" counterparts and moderately improves performance.

IEEE2021

Biologically plausible unsupervised learning in shallow and deep neural networks

Bernd Albert Illing

The way our brain learns to disentangle complex signals into unambiguous concepts is fascinating but remains largely unknown. There is evidence, however, that hierarchical neural representations play a key role in the cortex. This thesis investigates biologically plausible models of unsupervised learning of hierarchical representations as found in the brain and modern computer vision models. We use computational modeling to address three main questions at the intersection of artificial intelligence (AI) and computational neuroscience.The first question is: What are useful neural representations and when are deep hierarchical representations needed? We approach this point with a systematic study of biologically plausible unsupervised feature learning in a shallow 2-layer networks on digit (MNIST) and object (CIFAR10) classification. Surprisingly, random features support high performance, especially for large hidden layers. When combined with localized receptive fields, random feature networks approach the performance of supervised backpropagation on MNIST, but not on CIFAR10. We suggest that future models of biologically plausible learning should outperform such random feature benchmarks on MNIST, or that such models should be evaluated in different ways.The second question is: How can hierarchical representations be learned with mechanisms supported by neuroscientific evidence? We cover this question by proposing a unifying Hebbian model, inspired by common models of V1 simple and complex cells based on unsupervised sparse coding and temporal invariance learning. In shallow 2-layer networks, our model reproduces learning of simple and complex cell receptive fields, as found in V1. In deeper networks, we stack multiple layers of Hebbian learning but find that it does not yield hierarchical representations of increasing usefulness. From this, we hypothesise that standard Hebbian rules are too constrained to build increasingly useful representations, as observed in higher areas of the visual cortex or deep artificial neural networks.The third question is: Can AI inspire learning models that build deep representations and are still biologically plausible? We address this question by proposing a learning rule that takes inspiration from neuroscience and recent advances in self-supervised deep learning. The proposed rule is Hebbian, i.e. only depends on pre- and post-synaptic neuronal activity, but includes additional local factors, namely predictive dendritic input and widely broadcasted modulation factors. Algorithmically, this rule applies self-supervised contrastive predictive learning to a causal, biological setting using saccades. We find that networks trained with this generalised Hebbian rule build deep hierarchical representations of images, speech and video.We see our modeling as a potential starting point for both, new hypotheses, that can be tested experimentally, and novel AI models that could benefit from added biological realism.