Ghada Sokar

Ghada Sokar, Rishabh Agarwal, Pablo Samuel Castro et al. · deepmind

In this work we identify the dormant neuron phenomenon in deep reinforcement learning, where an agent's network suffers from an increasing number of inactive neurons, thereby affecting network expressivity. We demonstrate the presence of this phenomenon across a variety of algorithms and environments, and highlight its effect on learning. To address this issue, we propose a simple and effective method (ReDo) that Recycles Dormant neurons throughout training. Our experiments demonstrate that ReDo maintains the expressive power of networks by reducing the number of dormant neurons and results in improved performance.

Bram Grooten, Ghada Sokar, Shibhansh Dohare et al.

Tomorrow's robots will need to distinguish useful information from noise when performing different tasks. A household robot for instance may continuously receive a plethora of information about the home, but needs to focus on just a small subset to successfully execute its current chore. Filtering distracting inputs that contain irrelevant data has received little attention in the reinforcement learning literature. To start resolving this, we formulate a problem setting in reinforcement learning called the $\textit{extremely noisy environment}$ (ENE), where up to $99\%$ of the input features are pure noise. Agents need to detect which features provide task-relevant information about the state of the environment. Consequently, we propose a new method termed $\textit{Automatic Noise Filtering}$ (ANF), which uses the principles of dynamic sparse training in synergy with various deep reinforcement learning algorithms. The sparse input layer learns to focus its connectivity on task-relevant features, such that ANF-SAC and ANF-TD3 outperform standard SAC and TD3 by a large margin, while using up to $95\%$ fewer weights. Furthermore, we devise a transfer learning setting for ENEs, by permuting all features of the environment after 1M timesteps to simulate the fact that other information sources can become relevant as the world evolves. Again, ANF surpasses the baselines in final performance and sample complexity. Our code is available at https://github.com/bramgrooten/automatic-noise-filtering

Ghada Sokar, Zahra Atashgahi, Mykola Pechenizkiy et al.

A new line of research for feature selection based on neural networks has recently emerged. Despite its superiority to classical methods, it requires many training iterations to converge and detect informative features. The computational time becomes prohibitively long for datasets with a large number of samples or a very high dimensional feature space. In this paper, we present a new efficient unsupervised method for feature selection based on sparse autoencoders. In particular, we propose a new sparse training algorithm that optimizes a model's sparse topology during training to pay attention to informative features quickly. The attention-based adaptation of the sparse topology enables fast detection of informative features after a few training iterations. We performed extensive experiments on 10 datasets of different types, including image, speech, text, artificial, and biological. They cover a wide range of characteristics, such as low and high-dimensional feature spaces, and few and large training samples. Our proposed approach outperforms the state-of-the-art methods in terms of selecting informative features while reducing training iterations and computational costs substantially. Moreover, the experiments show the robustness of our method in extremely noisy environments.

Murat Onur Yildirim, Elif Ceren Gok Yildirim, Ghada Sokar et al.

Continual learning (CL) refers to the ability of an intelligent system to sequentially acquire and retain knowledge from a stream of data with as little computational overhead as possible. To this end; regularization, replay, architecture, and parameter isolation approaches were introduced to the literature. Parameter isolation using a sparse network which enables to allocate distinct parts of the neural network to different tasks and also allows to share of parameters between tasks if they are similar. Dynamic Sparse Training (DST) is a prominent way to find these sparse networks and isolate them for each task. This paper is the first empirical study investigating the effect of different DST components under the CL paradigm to fill a critical research gap and shed light on the optimal configuration of DST for CL if it exists. Therefore, we perform a comprehensive study in which we investigate various DST components to find the best topology per task on well-known CIFAR100 and miniImageNet benchmarks in a task-incremental CL setup since our primary focus is to evaluate the performance of various DST criteria, rather than the process of mask selection. We found that, at a low sparsity level, Erdos-Rényi Kernel (ERK) initialization utilizes the backbone more efficiently and allows to effectively learn increments of tasks. At a high sparsity level, unless it is extreme, uniform initialization demonstrates a more reliable and robust performance. In terms of growth strategy; performance is dependent on the defined initialization strategy and the extent of sparsity. Finally, adaptivity within DST components is a promising way for better continual learners.

Shiwei Liu, Tianlong Chen, Zahra Atashgahi et al.

The success of deep ensembles on improving predictive performance, uncertainty estimation, and out-of-distribution robustness has been extensively studied in the machine learning literature. Albeit the promising results, naively training multiple deep neural networks and combining their predictions at inference leads to prohibitive computational costs and memory requirements. Recently proposed efficient ensemble approaches reach the performance of the traditional deep ensembles with significantly lower costs. However, the training resources required by these approaches are still at least the same as training a single dense model. In this work, we draw a unique connection between sparse neural network training and deep ensembles, yielding a novel efficient ensemble learning framework called FreeTickets. Instead of training multiple dense networks and averaging them, we directly train sparse subnetworks from scratch and extract diverse yet accurate subnetworks during this efficient, sparse-to-sparse training. Our framework, FreeTickets, is defined as the ensemble of these relatively cheap sparse subnetworks. Despite being an ensemble method, FreeTickets has even fewer parameters and training FLOPs than a single dense model. This seemingly counter-intuitive outcome is due to the ultra training/inference efficiency of dynamic sparse training. FreeTickets surpasses the dense baseline in all the following criteria: prediction accuracy, uncertainty estimation, out-of-distribution (OoD) robustness, as well as efficiency for both training and inference. Impressively, FreeTickets outperforms the naive deep ensemble with ResNet50 on ImageNet using around only 1/5 of the training FLOPs required by the latter. We have released our source code at https://github.com/VITA-Group/FreeTickets.

Johan Obando-Ceron, Ghada Sokar, Timon Willi et al. · mila

The recent rapid progress in (self) supervised learning models is in large part predicted by empirical scaling laws: a model's performance scales proportionally to its size. Analogous scaling laws remain elusive for reinforcement learning domains, however, where increasing the parameter count of a model often hurts its final performance. In this paper, we demonstrate that incorporating Mixture-of-Expert (MoE) modules, and in particular Soft MoEs (Puigcerver et al., 2023), into value-based networks results in more parameter-scalable models, evidenced by substantial performance increases across a variety of training regimes and model sizes. This work thus provides strong empirical evidence towards developing scaling laws for reinforcement learning.

Ghada Sokar, Gintare Karolina Dziugaite, Anurag Arnab et al.

Continual learning is conventionally tackled through sequential fine-tuning, a process that, while enabling adaptation, inherently favors plasticity over the stability needed to retain prior knowledge. While existing approaches attempt to mitigate catastrophic forgetting, a bias towards recent tasks persists as they build upon this sequential nature. In this work we present a new perspective based on model merging to maintain stability while still retaining plasticity. Rather than just sequentially updating the model weights, we propose merging newly trained task parameters with previously learned ones, promoting a better balance. To maximize the effectiveness of the merging process, we propose a simple mechanism that promotes learning aligned weights with previous ones, thereby avoiding interference when merging. We evaluate this approach on large Vision-Language Models (VLMs), and demonstrate its effectiveness in reducing forgetting, increasing robustness to various task orders and similarities, and improving generalization.

Ghada Sokar, Pablo Samuel Castro

Scaling deep reinforcement learning in pixel-based environments presents a significant challenge, often resulting in diminished performance. While recent works have proposed algorithmic and architectural approaches to address this, the underlying cause of the performance drop remains unclear. In this paper, we identify the connection between the output of the encoder (a stack of convolutional layers) and the ensuing dense layers as the main underlying factor limiting scaling capabilities; we denote this connection as the bottleneck, and we demonstrate that previous approaches implicitly target this bottleneck. As a result of our analyses, we present global average pooling as a simple yet effective way of targeting the bottleneck, thereby avoiding the complexity of earlier approaches.

Bram Grooten, Farid Hasanov, Chenxiang Zhang et al.

Model ensembles have long been a cornerstone for improving generalization and robustness in deep learning. However, their effectiveness often comes at the cost of substantial computational overhead. To address this issue, state-of-the-art methods aim to replicate ensemble-class performance without requiring multiple independently trained networks. Unfortunately, these algorithms often still demand considerable compute at inference. In response to these limitations, we introduce $\textbf{NeuroTrails}$, a sparse multi-head architecture with dynamically evolving topology. This unexplored model-agnostic training paradigm improves ensemble performance while reducing the required resources. We analyze the underlying reason for its effectiveness and observe that the various neural trails induced by dynamic sparsity attain a $\textit{Goldilocks zone}$ of prediction diversity. NeuroTrails displays efficacy with convolutional and transformer-based architectures on computer vision and language tasks. Experiments on ResNet-50/ImageNet, LLaMA-350M/C4, among many others, demonstrate increased accuracy and stronger robustness in zero-shot generalization, while requiring significantly fewer parameters.

Ghada Sokar, Decebal Constantin Mocanu, Mykola Pechenizkiy

Using task-specific components within a neural network in continual learning (CL) is a compelling strategy to address the stability-plasticity dilemma in fixed-capacity models without access to past data. Current methods focus only on selecting a sub-network for a new task that reduces forgetting of past tasks. However, this selection could limit the forward transfer of relevant past knowledge that helps in future learning. Our study reveals that satisfying both objectives jointly is more challenging when a unified classifier is used for all classes of seen tasks-class-Incremental Learning (class-IL)-as it is prone to ambiguities between classes across tasks. Moreover, the challenge increases when the semantic similarity of classes across tasks increases. To address this challenge, we propose a new CL method, named AFAF, that aims to Avoid Forgetting and Allow Forward transfer in class-IL using fix-capacity models. AFAF allocates a sub-network that enables selective transfer of relevant knowledge to a new task while preserving past knowledge, reusing some of the previously allocated components to utilize the fixed-capacity, and addressing class-ambiguities when similarities exist. The experiments show the effectiveness of AFAF in providing models with multiple CL desirable properties, while outperforming state-of-the-art methods on various challenging benchmarks with different semantic similarities.

Ghada Sokar, Elena Mocanu, Decebal Constantin Mocanu et al.

Deep reinforcement learning (DRL) agents are trained through trial-and-error interactions with the environment. This leads to a long training time for dense neural networks to achieve good performance. Hence, prohibitive computation and memory resources are consumed. Recently, learning efficient DRL agents has received increasing attention. Yet, current methods focus on accelerating inference time. In this paper, we introduce for the first time a dynamic sparse training approach for deep reinforcement learning to accelerate the training process. The proposed approach trains a sparse neural network from scratch and dynamically adapts its topology to the changing data distribution during training. Experiments on continuous control tasks show that our dynamic sparse agents achieve higher performance than the equivalent dense methods, reduce the parameter count and floating-point operations (FLOPs) by 50%, and have a faster learning speed that enables reaching the performance of dense agents with 40-50% reduction in the training steps.

Ghada Sokar, Decebal Constantin Mocanu, Mykola Pechenizkiy

Continual learning aims to provide intelligent agents capable of learning multiple tasks sequentially with neural networks. One of its main challenging, catastrophic forgetting, is caused by the neural networks non-optimal ability to learn in non-stationary distributions. In most settings of the current approaches, the agent starts from randomly initialized parameters and is optimized to master the current task regardless of the usefulness of the learned representation for future tasks. Moreover, each of the future tasks uses all the previously learned knowledge although parts of this knowledge might not be helpful for its learning. These cause interference among tasks, especially when the data of previous tasks is not accessible. In this paper, we propose a new method, named Self-Attention Meta-Learner (SAM), which learns a prior knowledge for continual learning that permits learning a sequence of tasks, while avoiding catastrophic forgetting. SAM incorporates an attention mechanism that learns to select the particular relevant representation for each future task. Each task builds a specific representation branch on top of the selected knowledge, avoiding the interference between tasks. We evaluate the proposed method on the Split CIFAR-10/100 and Split MNIST benchmarks in the task agnostic inference. We empirically show that we can achieve a better performance than several state-of-the-art methods for continual learning by building on the top of selected representation learned by SAM. We also show the role of the meta-attention mechanism in boosting informative features corresponding to the input data and identifying the correct target in the task agnostic inference. Finally, we demonstrate that popular existing continual learning methods gain a performance boost when they adopt SAM as a starting point.

Ghada Sokar, Decebal Constantin Mocanu, Mykola Pechenizkiy

Continual learning aims to provide intelligent agents that are capable of learning continually a sequence of tasks, building on previously learned knowledge. A key challenge in this learning paradigm is catastrophically forgetting previously learned tasks when the agent faces a new one. Current rehearsal-based methods show their success in mitigating the catastrophic forgetting problem by replaying samples from previous tasks during learning a new one. However, these methods are infeasible when the data of previous tasks is not accessible. In this work, we propose a new pseudo-rehearsal-based method, named learning Invariant Representation for Continual Learning (IRCL), in which class-invariant representation is disentangled from a conditional generative model and jointly used with class-specific representation to learn the sequential tasks. Disentangling the shared invariant representation helps to learn continually a sequence of tasks, while being more robust to forgetting and having better knowledge transfer. We focus on class incremental learning where there is no knowledge about task identity during inference. We empirically evaluate our proposed method on two well-known benchmarks for continual learning: split MNIST and split Fashion MNIST. The experimental results show that our proposed method outperforms regularization-based methods by a big margin and is better than the state-of-the-art pseudo-rehearsal-based method. Finally, we analyze the role of the shared invariant representation in mitigating the forgetting problem especially when the number of replayed samples for each previous task is small.

Zahra Atashgahi, Ghada Sokar, Tim van der Lee et al.

Major complications arise from the recent increase in the amount of high-dimensional data, including high computational costs and memory requirements. Feature selection, which identifies the most relevant and informative attributes of a dataset, has been introduced as a solution to this problem. Most of the existing feature selection methods are computationally inefficient; inefficient algorithms lead to high energy consumption, which is not desirable for devices with limited computational and energy resources. In this paper, a novel and flexible method for unsupervised feature selection is proposed. This method, named QuickSelection, introduces the strength of the neuron in sparse neural networks as a criterion to measure the feature importance. This criterion, blended with sparsely connected denoising autoencoders trained with the sparse evolutionary training procedure, derives the importance of all input features simultaneously. We implement QuickSelection in a purely sparse manner as opposed to the typical approach of using a binary mask over connections to simulate sparsity. It results in a considerable speed increase and memory reduction. When tested on several benchmark datasets, including five low-dimensional and three high-dimensional datasets, the proposed method is able to achieve the best trade-off of classification and clustering accuracy, running time, and maximum memory usage, among widely used approaches for feature selection. Besides, our proposed method requires the least amount of energy among the state-of-the-art autoencoder-based feature selection methods.

Ghada Sokar, Decebal Constantin Mocanu, Mykola Pechenizkiy

The continual learning (CL) paradigm aims to enable neural networks to learn tasks continually in a sequential fashion. The fundamental challenge in this learning paradigm is catastrophic forgetting previously learned tasks when the model is optimized for a new task, especially when their data is not accessible. Current architectural-based methods aim at alleviating the catastrophic forgetting problem but at the expense of expanding the capacity of the model. Regularization-based methods maintain a fixed model capacity; however, previous studies showed the huge performance degradation of these methods when the task identity is not available during inference (e.g. class incremental learning scenario). In this work, we propose a novel architectural-based method referred as SpaceNet for class incremental learning scenario where we utilize the available fixed capacity of the model intelligently. SpaceNet trains sparse deep neural networks from scratch in an adaptive way that compresses the sparse connections of each task in a compact number of neurons. The adaptive training of the sparse connections results in sparse representations that reduce the interference between the tasks. Experimental results show the robustness of our proposed method against catastrophic forgetting old tasks and the efficiency of SpaceNet in utilizing the available capacity of the model, leaving space for more tasks to be learned. In particular, when SpaceNet is tested on the well-known benchmarks for CL: split MNIST, split Fashion-MNIST, and CIFAR-10/100, it outperforms regularization-based methods by a big performance gap. Moreover, it achieves better performance than architectural-based methods without model expansion and achieved comparable results with rehearsal-based methods, while offering a huge memory reduction.

Shiwei Liu, Tim Van der Lee, Anil Yaman et al.

Sparse neural networks are effective approaches to reduce the resource requirements for the deployment of deep neural networks. Recently, the concept of adaptive sparse connectivity, has emerged to allow training sparse neural networks from scratch by optimizing the sparse structure during training. However, comparing different sparse topologies and determining how sparse topologies evolve during training, especially for the situation in which the sparse structure optimization is involved, remain as challenging open questions. This comparison becomes increasingly complex as the number of possible topological comparisons increases exponentially with the size of networks. In this work, we introduce an approach to understand and compare sparse neural network topologies from the perspective of graph theory. We first propose Neural Network Sparse Topology Distance (NNSTD) to measure the distance between different sparse neural networks. Further, we demonstrate that sparse neural networks can outperform over-parameterized models in terms of performance, even without any further structure optimization. To the end, we also show that adaptive sparse connectivity can always unveil a plenitude of sparse sub-networks with very different topologies which outperform the dense model, by quantifying and comparing their topological evolutionary processes. The latter findings complement the Lottery Ticket Hypothesis by showing that there is a much more efficient and robust way to find "winning tickets". Altogether, our results start enabling a better theoretical understanding of sparse neural networks, and demonstrate the utility of using graph theory to analyze them.