1,721,016 research outputs found
Partial Hypernetworks for Continual Learning
No full textHypernetworks mitigate forgetting in continual learning (CL) by generating task-dependent weights and penalizing weight changes at a meta-model level. Unfortunately, generating all weights is not only computationally expensive for larger architectures, but also, it is not well understood whether generating all model weights is necessary. Inspired by latent replay methods in CL, we propose partial weight generation for the final layers of a model using hypernetworks while freezing the initial layers. With this objective, we first answer the question of how many layers can be frozen without compromising the final performance. Through several experiments, we empirically show that the number of layers that can be frozen is proportional to the distributional similarity in the CL stream. Then, to demonstrate the effectiveness of hypernetworks, we show that noisy streams can significantly impact the performance of latent replay methods, leading to increased forgetting when features from noisy experiences are replayed with old samples. In contrast, partial hypernetworks are more robust to noise by maintaining accuracy on previous experiences. Finally, we conduct experiments on the split CIFAR-100 and TinyImagenet benchmarks and compare different versions of partial hypernetworks to latent replay methods. We conclude that partial weight generation using hypernetworks is a promising solution to the problem of forgetting in neural networks. It can provide an effective balance between computation and final test accuracy in CL streams
RaSP: Relation-aware Semantic Prior for Weakly Supervised Incremental Segmentation
Get PDFClass-incremental semantic image segmentation assumes multiple model updates, each enriching the model to segment new categories. This is typically carried out by providing expensive pixel-level annotations to the training algorithm for all new objects, limiting the adoption of such methods in practical applications. Approaches that solely require image-level labels offer an attractive alternative, yet, such coarse annotations lack precise information about the location and boundary of the new objects. In this paper we argue that, since classes represent not just indices but semantic entities, the conceptual relationships between them can provide valuable information that should be leveraged. We propose a weakly supervised approach that exploits such semantic relations to transfer objectness prior from the previously learned classes into the new ones, complementing the supervisory signal from image-level labels. We validate our approach on a number of continual learning tasks, and show how even a simple pairwise interaction between classes can significantly improve the segmentation mask quality of both old and new classes. We show these conclusions still hold for longer and, hence, more realistic sequences of tasks and for a challenging few-shot scenari
Understanding learning from non-stationary data
No full textDeep neural networks are ubiquitous, powering necessary predictions for numerous AI applications. These models are trained on evergrowing data collections, making the old paradigm of optimising neural networks from scratch on a fixed data set inefficient and obsolete. As advocated by the proponents of continual learning, these networks need to efficiently absorb knowledge in stages. We need, therefore, to assume that the training data changes during learning, and models need to adapt fast to the corresponding shifts in the output of the optimisation functions. In this monograph, I discuss the challenges that come with this expectation.
This thesis puts together three contributions to this topic. First, we study a surprising fact in supervised learning: pre-training negatively affects the generalisation of fine-tuned models compared to those trained from scratch. We look at this ``generalisation gap'', discuss the conditions for its manifestation and put forward a hypothesis for why it happens. An antidote to this generalisation gap is partially resetting parameters during training. Therefore, the second contribution of the thesis focuses on this regularisation technique and asks how it interacts with more commonly used regularisers and why state-of-the-art setups do not use re-initialisations. For the third contribution, we turn our attention to reinforcement learning, a scenario where the training data distribution shifts continuously because of the changes in the policy interacting with the environment. We propose a technique based on spectral normalisation and demonstrate its efficacy for value-based deep reinforcement learning agents on the Atari benchmark.
Hopefully, this thesis brings a couple of useful observations in the era of designing perpetually learning agents and fine-tuning large foundational models.Open Acces
Graphframex: Towards systematic evaluation of explainability methods for graph neural networks
No full textISSN:2640-349
On Recurrent and Deep Neural Networks
Get PDFL'apprentissage profond est un domaine de recherche en forte croissance en apprentissage automatique qui est parvenu à des résultats impressionnants dans différentes tâches allant de la classification d'images à la parole, en passant par la modélisation du langage. Les réseaux de neurones récurrents, une sous-classe d'architecture profonde, s'avèrent particulièrement prometteurs. Les réseaux récurrents peuvent capter la structure temporelle dans les données. Ils ont potentiellement la capacité d'apprendre des corrélations entre des événements éloignés dans le temps et d'emmagasiner indéfiniment des informations dans leur mémoire interne. Dans ce travail, nous tentons d'abord de comprendre pourquoi la profondeur est utile. Similairement à d'autres travaux de la littérature, nos résultats démontrent que les modèles profonds peuvent être plus efficaces pour représenter certaines familles de fonctions comparativement aux modèles peu profonds. Contrairement à ces travaux, nous effectuons notre analyse théorique sur des réseaux profonds acycliques munis de fonctions d'activation linéaires par parties, puisque ce type de modèle est actuellement l'état de l'art dans différentes tâches de classification. La deuxième partie de cette thèse porte sur le processus d'apprentissage. Nous analysons quelques techniques d'optimisation proposées récemment, telles l'optimisation Hessian free, la descente de gradient naturel et la descente des sous-espaces de Krylov. Nous proposons le cadre théorique des méthodes à région de confiance généralisées et nous montrons que plusieurs de ces algorithmes développés récemment peuvent être vus dans cette perspective. Nous argumentons que certains membres de cette famille d'approches peuvent être mieux adaptés que d'autres à l'optimisation non convexe. La dernière partie de ce document se concentre sur les réseaux de neurones récurrents. Nous étudions d'abord le concept de mémoire et tentons de répondre aux questions suivantes: Les réseaux récurrents peuvent-ils démontrer une mémoire sans limite? Ce comportement peut-il être appris? Nous montrons que cela est possible si des indices sont fournis durant l'apprentissage. Ensuite, nous explorons deux problèmes spécifiques à l'entraînement des réseaux récurrents, à savoir la dissipation et l'explosion du gradient. Notre analyse se termine par une solution au problème d'explosion du gradient qui implique de borner la norme du gradient. Nous proposons également un terme de régularisation conçu spécifiquement pour réduire le problème de dissipation du gradient. Sur un ensemble de données synthétique, nous montrons empiriquement que ces mécanismes peuvent permettre aux réseaux récurrents d'apprendre de façon autonome à mémoriser des informations pour une période de temps indéfinie. Finalement, nous explorons la notion de profondeur dans les réseaux de neurones récurrents. Comparativement aux réseaux acycliques, la définition de profondeur dans les réseaux récurrents est souvent ambiguë. Nous proposons différentes façons d'ajouter de la profondeur dans les réseaux récurrents et nous évaluons empiriquement ces propositions.Deep Learning is a quickly growing area of research in machine learning, providing impressive results on different tasks ranging from image classification to speech and language modelling. In particular, a subclass of deep models, recurrent neural networks, promise even more. Recurrent models can capture the temporal structure in the data. They can learn correlations between events that might be far apart in time and, potentially, store information for unbounded amounts of time in their innate memory. In this work we first focus on understanding why depth is useful. Similar to other published work, our results prove that deep models can be more efficient at expressing certain families of functions compared to shallow models. Different from other work, we carry out our theoretical analysis on deep feedforward networks with piecewise linear activation functions, the kind of models that have obtained state of the art results on different classification tasks. The second part of the thesis looks at the learning process. We analyse a few recently proposed optimization techniques, including Hessian Free Optimization, natural gradient descent and Krylov Subspace Descent. We propose the framework of generalized trust region methods and show that many of these recently proposed algorithms can be viewed from this perspective. We argue that certain members of this family of approaches might be better suited for non-convex optimization than others. The last part of the document focuses on recurrent neural networks. We start by looking at the concept of memory. The questions we attempt to answer are: Can recurrent models exhibit unbounded memory? Can this behaviour be learnt? We show this to be true if hints are provided during learning. We explore, afterwards, two specific difficulties of training recurrent models, namely the vanishing gradients and exploding gradients problem. Our analysis concludes with a heuristic solution for the exploding gradients that involves clipping the norm of the gradients. We also propose a specific regularization term meant to address the vanishing gradients problem. On a toy dataset, employing these mechanisms, we provide anecdotal evidence that the recurrent model might be able to learn, with out hints, to exhibit some sort of unbounded memory. Finally we explore the concept of depth for recurrent neural networks. Compared to feedforward models, for recurrent models the meaning of depth can be ambiguous. We provide several ways in which a recurrent model can be made deep and empirically evaluate these proposals
On Recurrent and Deep Neural Networks
Get PDFL'apprentissage profond est un domaine de recherche en forte croissance en apprentissage automatique qui est parvenu à des résultats impressionnants dans différentes tâches allant de la classification d'images à la parole, en passant par la modélisation du langage. Les réseaux de neurones récurrents, une sous-classe d'architecture profonde, s'avèrent particulièrement prometteurs. Les réseaux récurrents peuvent capter la structure temporelle dans les données. Ils ont potentiellement la capacité d'apprendre des corrélations entre des événements éloignés dans le temps et d'emmagasiner indéfiniment des informations dans leur mémoire interne. Dans ce travail, nous tentons d'abord de comprendre pourquoi la profondeur est utile. Similairement à d'autres travaux de la littérature, nos résultats démontrent que les modèles profonds peuvent être plus efficaces pour représenter certaines familles de fonctions comparativement aux modèles peu profonds. Contrairement à ces travaux, nous effectuons notre analyse théorique sur des réseaux profonds acycliques munis de fonctions d'activation linéaires par parties, puisque ce type de modèle est actuellement l'état de l'art dans différentes tâches de classification. La deuxième partie de cette thèse porte sur le processus d'apprentissage. Nous analysons quelques techniques d'optimisation proposées récemment, telles l'optimisation Hessian free, la descente de gradient naturel et la descente des sous-espaces de Krylov. Nous proposons le cadre théorique des méthodes à région de confiance généralisées et nous montrons que plusieurs de ces algorithmes développés récemment peuvent être vus dans cette perspective. Nous argumentons que certains membres de cette famille d'approches peuvent être mieux adaptés que d'autres à l'optimisation non convexe. La dernière partie de ce document se concentre sur les réseaux de neurones récurrents. Nous étudions d'abord le concept de mémoire et tentons de répondre aux questions suivantes: Les réseaux récurrents peuvent-ils démontrer une mémoire sans limite? Ce comportement peut-il être appris? Nous montrons que cela est possible si des indices sont fournis durant l'apprentissage. Ensuite, nous explorons deux problèmes spécifiques à l'entraînement des réseaux récurrents, à savoir la dissipation et l'explosion du gradient. Notre analyse se termine par une solution au problème d'explosion du gradient qui implique de borner la norme du gradient. Nous proposons également un terme de régularisation conçu spécifiquement pour réduire le problème de dissipation du gradient. Sur un ensemble de données synthétique, nous montrons empiriquement que ces mécanismes peuvent permettre aux réseaux récurrents d'apprendre de façon autonome à mémoriser des informations pour une période de temps indéfinie. Finalement, nous explorons la notion de profondeur dans les réseaux de neurones récurrents. Comparativement aux réseaux acycliques, la définition de profondeur dans les réseaux récurrents est souvent ambiguë. Nous proposons différentes façons d'ajouter de la profondeur dans les réseaux récurrents et nous évaluons empiriquement ces propositions.Deep Learning is a quickly growing area of research in machine learning, providing impressive results on different tasks ranging from image classification to speech and language modelling. In particular, a subclass of deep models, recurrent neural networks, promise even more. Recurrent models can capture the temporal structure in the data. They can learn correlations between events that might be far apart in time and, potentially, store information for unbounded amounts of time in their innate memory. In this work we first focus on understanding why depth is useful. Similar to other published work, our results prove that deep models can be more efficient at expressing certain families of functions compared to shallow models. Different from other work, we carry out our theoretical analysis on deep feedforward networks with piecewise linear activation functions, the kind of models that have obtained state of the art results on different classification tasks. The second part of the thesis looks at the learning process. We analyse a few recently proposed optimization techniques, including Hessian Free Optimization, natural gradient descent and Krylov Subspace Descent. We propose the framework of generalized trust region methods and show that many of these recently proposed algorithms can be viewed from this perspective. We argue that certain members of this family of approaches might be better suited for non-convex optimization than others. The last part of the document focuses on recurrent neural networks. We start by looking at the concept of memory. The questions we attempt to answer are: Can recurrent models exhibit unbounded memory? Can this behaviour be learnt? We show this to be true if hints are provided during learning. We explore, afterwards, two specific difficulties of training recurrent models, namely the vanishing gradients and exploding gradients problem. Our analysis concludes with a heuristic solution for the exploding gradients that involves clipping the norm of the gradients. We also propose a specific regularization term meant to address the vanishing gradients problem. On a toy dataset, employing these mechanisms, we provide anecdotal evidence that the recurrent model might be able to learn, with out hints, to exhibit some sort of unbounded memory. Finally we explore the concept of depth for recurrent neural networks. Compared to feedforward models, for recurrent models the meaning of depth can be ambiguous. We provide several ways in which a recurrent model can be made deep and empirically evaluate these proposals
Hadamard domain training with integers for class incremental quantized learning
No full textContinual learning (CL) implemented directly on-device is crucial for practical deployment of applications to battery powered devices, where privacy needs must be balanced with personalization, and agility in adapting to new data. Existing CL techniques can be cost-prohibitive on such devices requiring quantized operations to enable practical deployment. However, as we show commonly used fully quantized training (FQT) solutions do not converge when applied to CL with low-precision hardware. We propose Hadamard Domain Quantized Training (HDQT), that uses the Hadamard transform to facilitate FQT with 4-bit integer operands. HDQT enables low-precision, on-device training where other FQT solutions fail. An examination of gradient alignment reveals that for early feature detection layers, HDQT gradients are better aligned to the unquantized baselines than those generated by other FQT methods. This improved alignment translates to consistently better performance over the course of learning, reflected in the training trajectories through the model loss-landscape. Numerical experiments conducted on Human Activity Recognition (HAR) datasets reveal a ≪ 1% average accuracy reduction across various competitive CL methods, even under aggressive 4-bit quantization with 8-bit accumulators. Additionally, on CIFAR100 there is no loss of accuracy observed when the accumulator precision is relaxed to 12 bits for competitive non-dynamic architecture CL methods
- …
