Moin Nabi

Guanglei Yang, Enrico Fini, Dan Xu et al.

A fundamental and challenging problem in deep learning is catastrophic forgetting, i.e. the tendency of neural networks to fail to preserve the knowledge acquired from old tasks when learning new tasks. This problem has been widely investigated in the research community and several Incremental Learning (IL) approaches have been proposed in the past years. While earlier works in computer vision have mostly focused on image classification and object detection, more recently some IL approaches for semantic segmentation have been introduced. These previous works showed that, despite its simplicity, knowledge distillation can be effectively employed to alleviate catastrophic forgetting. In this paper, we follow this research direction and, inspired by recent literature on contrastive learning, we propose a novel distillation framework, Uncertainty-aware Contrastive Distillation (\method). In a nutshell, \method~is operated by introducing a novel distillation loss that takes into account all the images in a mini-batch, enforcing similarity between features associated to all the pixels from the same classes, and pulling apart those corresponding to pixels from different classes. In order to mitigate catastrophic forgetting, we contrast features of the new model with features extracted by a frozen model learned at the previous incremental step. Our experimental results demonstrate the advantage of the proposed distillation technique, which can be used in synergy with previous IL approaches, and leads to state-of-art performance on three commonly adopted benchmarks for incremental semantic segmentation. The code is available at \url{https://github.com/ygjwd12345/UCD}.

Zhiqi Kang, Enrico Fini, Moin Nabi et al.

Despite significant advances, the performance of state-of-the-art continual learning approaches hinges on the unrealistic scenario of fully labeled data. In this paper, we tackle this challenge and propose an approach for continual semi-supervised learning--a setting where not all the data samples are labeled. A primary issue in this scenario is the model forgetting representations of unlabeled data and overfitting the labeled samples. We leverage the power of nearest-neighbor classifiers to nonlinearly partition the feature space and flexibly model the underlying data distribution thanks to its non-parametric nature. This enables the model to learn a strong representation for the current task, and distill relevant information from previous tasks. We perform a thorough experimental evaluation and show that our method outperforms all the existing approaches by large margins, setting a solid state of the art on the continual semi-supervised learning paradigm. For example, on CIFAR-100 we surpass several others even when using at least 30 times less supervision (0.8% vs. 25% of annotations). Finally, our method works well on both low and high resolution images and scales seamlessly to more complex datasets such as ImageNet-100. The code is publicly available on https://github.com/kangzhiq/NNCSL

Mohammad Samragh, Iman Mirzadeh, Keivan Alizadeh Vahid et al. · utoronto

The pre-training phase of language models often begins with randomly initialized parameters. With the current trends in scaling models, training their large number of parameters can be extremely slow and costly. In contrast, small language models are less expensive to train, but they often cannot achieve the accuracy of large models. In this paper, we explore an intriguing idea to connect these two different regimes: Can we develop a method to initialize large language models using smaller pre-trained models? Will such initialization bring any benefits in terms of training time and final accuracy? In this paper, we introduce HyperCloning, a method that can expand the parameters of a pre-trained language model to those of a larger model with increased hidden dimensions. Our method ensures that the larger model retains the functionality of the smaller model. As a result, the larger model already inherits the predictive power and accuracy of the smaller model before the training starts. We demonstrate that training such an initialized model results in significant savings in terms of GPU hours required for pre-training large language models.

Enrico Fini, Pietro Astolfi, Karteek Alahari et al.

Self-supervised learning models have been shown to learn rich visual representations without requiring human annotations. However, in many real-world scenarios, labels are partially available, motivating a recent line of work on semi-supervised methods inspired by self-supervised principles. In this paper, we propose a conceptually simple yet empirically powerful approach to turn clustering-based self-supervised methods such as SwAV or DINO into semi-supervised learners. More precisely, we introduce a multi-task framework merging a supervised objective using ground-truth labels and a self-supervised objective relying on clustering assignments with a single cross-entropy loss. This approach may be interpreted as imposing the cluster centroids to be class prototypes. Despite its simplicity, we provide empirical evidence that our approach is highly effective and achieves state-of-the-art performance on CIFAR100 and ImageNet.

Tassilo Klein, Moin Nabi

In this paper, we propose Self-Contrastive Decorrelation (SCD), a self-supervised approach. Given an input sentence, it optimizes a joint self-contrastive and decorrelation objective. Learning a representation is facilitated by leveraging the contrast arising from the instantiation of standard dropout at different rates. The proposed method is conceptually simple yet empirically powerful. It achieves comparable results with state-of-the-art methods on multiple benchmarks without using contrastive pairs. This study opens up avenues for efficient self-supervised learning methods that are more robust than current contrastive methods.

Tassilo Klein, Moin Nabi

This paper presents miCSE, a mutual information-based contrastive learning framework that significantly advances the state-of-the-art in few-shot sentence embedding. The proposed approach imposes alignment between the attention pattern of different views during contrastive learning. Learning sentence embeddings with miCSE entails enforcing the structural consistency across augmented views for every sentence, making contrastive self-supervised learning more sample efficient. As a result, the proposed approach shows strong performance in the few-shot learning domain. While it achieves superior results compared to state-of-the-art methods on multiple benchmarks in few-shot learning, it is comparable in the full-shot scenario. This study opens up avenues for efficient self-supervised learning methods that are more robust than current contrastive methods for sentence embedding.

Jannik Wolff, Tassilo Klein, Moin Nabi et al.

Machine learning systems are often deployed in domains that entail data from multiple modalities, for example, phenotypic and genotypic characteristics describe patients in healthcare. Previous works have developed multimodal variational autoencoders (VAEs) that generate several modalities. We consider subjective data, where single datapoints from one modality (such as class labels) describe multiple datapoints from another modality (such as images). We theoretically and empirically demonstrate that multimodal VAEs with a mixture of experts posterior can struggle to capture variability in such surjective data.

Lorenzo Noci, Gregor Bachmann, Seyed-Mohsen Moosavi-Dezfooli et al.

Autoregressive language models trained with next-token prediction generate text by sampling one discrete token at a time. Although very scalable, this objective forces the model to commit at every step, preventing it from exploring or reflecting upon multiple plausible continuations. Furthermore, the compute allocation across tokens is uniform; every token is formed based on a single forward-pass, potentially limiting the model's expressiveness in cases where difficult tokens require inherently more compute. Towards addressing these limitations, we introduce latent lookahead, a training strategy that enables models to "think" before generating: at selected positions in the sequence, before committing to the next token, the model performs a multi-step lookahead in latent space. More precisely, instead of sampling future tokens, we leverage the network's latent space by recursively feeding its hidden states back into the context for $τ$ steps, investing more compute on predicting that token. This produces $τ$ latent predictions that are supervised against the next $τ$ ground-truth tokens, encouraging the model to "lookahead" and refine its prediction. We show that latent lookahead substantially outperforms both autoregressive and non-autoregressive baselines on planning tasks such as maze solving, Sudoku, and ProsQA, where foresight is essential.

Victor G. Turrisi da Costa, Enrico Fini, Moin Nabi et al.

This paper presents solo-learn, a library of self-supervised methods for visual representation learning. Implemented in Python, using Pytorch and Pytorch lightning, the library fits both research and industry needs by featuring distributed training pipelines with mixed-precision, faster data loading via Nvidia DALI, online linear evaluation for better prototyping, and many additional training tricks. Our goal is to provide an easy-to-use library comprising a large amount of Self-supervised Learning (SSL) methods, that can be easily extended and fine-tuned by the community. solo-learn opens up avenues for exploiting large-budget SSL solutions on inexpensive smaller infrastructures and seeks to democratize SSL by making it accessible to all. The source code is available at https://github.com/vturrisi/solo-learn.

Gregor Bachmann, Yichen Jiang, Seyed Mohsen Moosavi Dezfooli et al.

Chain-of-thought (CoT) prompting is a de-facto standard technique to elicit reasoning-like responses from large language models (LLMs), allowing them to spell out individual steps before giving a final answer. While the resemblance to human-like reasoning is undeniable, the driving forces underpinning the success of CoT reasoning still remain largely unclear. In this work, we perform an in-depth analysis of CoT traces originating from competition-level mathematics questions, with the aim of better understanding how, and which parts of CoT actually contribute to the final answer. To this end, we introduce the notion of a potential, quantifying how much a given part of CoT increases the likelihood of a correct completion. Upon examination of reasoning traces through the lens of the potential, we identify surprising patterns including (1) its often strong non-monotonicity (due to reasoning tangents), (2) very sharp but sometimes tough to interpret spikes (reasoning insights and jumps) as well as (3) at times lucky guesses, where the model arrives at the correct answer without providing any relevant justifications before. While some of the behaviours of the potential are readily interpretable and align with human intuition (such as insights and tangents), others remain difficult to understand from a human perspective. To further quantify the reliance of LLMs on reasoning insights, we investigate the notion of CoT transferability, where we measure the potential of a weaker model under the partial CoT from another, stronger model. Indeed aligning with our previous results, we find that as little as 20% of partial CoT can ``unlock'' the performance of the weaker model on problems that were previously unsolvable for it, highlighting that a large part of the mechanics underpinning CoT are transferable.

Enrico Fini, Mustafa Shukor, Xiujun Li et al.

We introduce a novel method for pre-training of large-scale vision encoders. Building on recent advancements in autoregressive pre-training of vision models, we extend this framework to a multimodal setting, i.e., images and text. In this paper, we present AIMV2, a family of generalist vision encoders characterized by a straightforward pre-training process, scalability, and remarkable performance across a range of downstream tasks. This is achieved by pairing the vision encoder with a multimodal decoder that autoregressively generates raw image patches and text tokens. Our encoders excel not only in multimodal evaluations but also in vision benchmarks such as localization, grounding, and classification. Notably, our AIMV2-3B encoder achieves 89.5% accuracy on ImageNet-1k with a frozen trunk. Furthermore, AIMV2 consistently outperforms state-of-the-art contrastive models (e.g., CLIP, SigLIP) in multimodal image understanding across diverse settings.

Saleh Ashkboos, Iman Mirzadeh, Keivan Alizadeh et al. · utoronto, uw

While large language models (LLMs) dominate the AI landscape, Small-scale large Language Models (SLMs) are gaining attention due to cost and efficiency demands from consumers. However, there is limited research on the training behavior and computational requirements of SLMs. In this study, we explore the computational bottlenecks of training SLMs (up to 2B parameters) by examining the effects of various hyperparameters and configurations, including GPU type, batch size, model size, communication protocol, attention type, and the number of GPUs. We assess these factors on popular cloud services using metrics such as loss per dollar and tokens per second. Our findings aim to support the broader adoption and optimization of language model training for low-resource AI research institutes.

Kumari Nishu, Sachin Mehta, Samira Abnar et al.

Training large language models (LLMs) for different inference constraints is computationally expensive, limiting control over efficiency-accuracy trade-offs. Moreover, once trained, these models typically process tokens uniformly, regardless of their complexity, leading to static and inflexible behavior. In this paper, we introduce a post-training optimization framework, DynaMoE, that adapts a pre-trained dense LLM to a token-difficulty-driven Mixture-of-Experts model with minimal fine-tuning cost. This adaptation makes the model dynamic, with sensitivity control to customize the balance between efficiency and accuracy. DynaMoE features a token-difficulty-aware router that predicts the difficulty of tokens and directs them to the appropriate sub-networks or experts, enabling larger experts to handle more complex tokens and smaller experts to process simpler ones. Our experiments demonstrate that DynaMoE can generate a range of adaptive model variants of the existing trained LLM with a single fine-tuning step, utilizing only $10B$ tokens, a minimal cost compared to the base model's training. Each variant offers distinct trade-offs between accuracy and performance. Compared to the baseline post-training optimization framework, Flextron, our method achieves similar aggregated accuracy across downstream tasks, despite using only $\frac{1}{9}\text{th}$ of their fine-tuning cost.

Atoosa Chegini, Hamid Kazemi, Iman Mirzadeh et al. · uw

In Large Language Model (LLM) development, Reinforcement Learning from Human Feedback (RLHF) is crucial for aligning models with human values and preferences. RLHF traditionally relies on the Kullback-Leibler (KL) divergence between the current policy and a frozen initial policy as a reference, which is added as a penalty in policy optimization algorithms like Proximal Policy Optimization (PPO). While this constraint prevents models from deviating too far from the initial checkpoint, it limits exploration of the reward landscape, reducing the model's ability to discover higher-quality solutions. As a result, policy optimization is often trapped in a narrow region of the parameter space, leading to suboptimal alignment and performance. This paper presents SALSA (Soup-based Alignment Learning for Stronger Adaptation), a novel approach designed to overcome these limitations by creating a more flexible and better located reference model through weight-space averaging of two independent supervised fine-tuned (SFT) models. This model soup allows for larger deviation in KL divergence and exploring a promising region of the solution space without sacrificing stability. By leveraging this more robust reference model, SALSA fosters better exploration, achieving higher rewards and improving model robustness, out-of-distribution generalization, and performance. We validate the effectiveness of SALSA through extensive experiments on popular open models (Llama2-7B, Mistral-7B, and Gemma-2B) across various benchmarks (MT-Bench, Arena-Hard, UltraFeedback), where it consistently surpasses PPO by fostering deeper exploration and achieving superior alignment in LLMs.

Tassilo Klein, Moin Nabi

How can we learn a representation with high predictive power while preserving user privacy? We present an adversarial representation learning method for sanitizing sensitive content from the learned representation. Specifically, we introduce a variant of entropy - focal entropy, which mitigates the potential information leakage of the existing entropy-based approaches. We showcase feasibility on multiple benchmarks. The results suggest high target utility at moderate privacy leakage.

Tassilo Klein, Moin Nabi

The generation of toxic content by large language models (LLMs) remains a critical challenge for the safe deployment of language technology. We propose a novel framework for implicit knowledge editing and controlled text generation by fine-tuning LLMs with a prototype-based contrastive perplexity objective. Central to our method is the construction of hard negatives - toxic outputs that are generated through adversarial paraphrasing to be semantically similar and model probability to their non-toxic counterparts. By training on these challenging and realistic pairs, our approach ensures robust and stable contrastive optimization. Experimental results in the domain of detoxification demonstrate that our method significantly reduces toxic generation while maintaining strong performance on downstream tasks such as commonsense reasoning and reading comprehension. Our findings highlight the effectiveness of exploiting hard negatives for attribute-aware fine-tuning.

Tassilo Klein, Moin Nabi

Self-supervised learning has recently attracted considerable attention in the NLP community for its ability to learn discriminative features using a contrastive objective. This paper investigates whether contrastive learning can be extended to Transfomer attention to tackling the Winograd Schema Challenge. To this end, we propose a novel self-supervised framework, leveraging a contrastive loss directly at the level of self-attention. Experimental analysis of our attention-based models on multiple datasets demonstrates superior commonsense reasoning capabilities. The proposed approach outperforms all comparable unsupervised approaches while occasionally surpassing supervised ones.

Tassilo Klein, Moin Nabi

Can we get existing language models and refine them for zero-shot commonsense reasoning? This paper presents an initial study exploring the feasibility of zero-shot commonsense reasoning for the Winograd Schema Challenge by formulating the task as self-supervised refinement of a pre-trained language model. In contrast to previous studies that rely on fine-tuning annotated datasets, we seek to boost conceptualization via loss landscape refinement. To this end, we propose a novel self-supervised learning approach that refines the language model utilizing a set of linguistic perturbations of similar concept relationships. Empirical analysis of our conceptually simple framework demonstrates the viability of zero-shot commonsense reasoning on multiple benchmarks.

Enrico Fini, Enver Sangineto, Stéphane Lathuilière et al.

In this paper, we study the problem of Novel Class Discovery (NCD). NCD aims at inferring novel object categories in an unlabeled set by leveraging from prior knowledge of a labeled set containing different, but related classes. Existing approaches tackle this problem by considering multiple objective functions, usually involving specialized loss terms for the labeled and the unlabeled samples respectively, and often requiring auxiliary regularization terms. In this paper, we depart from this traditional scheme and introduce a UNified Objective function (UNO) for discovering novel classes, with the explicit purpose of favoring synergy between supervised and unsupervised learning. Using a multi-view self-labeling strategy, we generate pseudo-labels that can be treated homogeneously with ground truth labels. This leads to a single classification objective operating on both known and unknown classes. Despite its simplicity, UNO outperforms the state of the art by a significant margin on several benchmarks (~+10% on CIFAR-100 and +8% on ImageNet). The project page is available at: https://ncd-uno.github.io.

Frederik Pahde, Mihai Puscas, Tassilo Klein et al.

Although providing exceptional results for many computer vision tasks, state-of-the-art deep learning algorithms catastrophically struggle in low data scenarios. However, if data in additional modalities exist (e.g. text) this can compensate for the lack of data and improve the classification results. To overcome this data scarcity, we design a cross-modal feature generation framework capable of enriching the low populated embedding space in few-shot scenarios, leveraging data from the auxiliary modality. Specifically, we train a generative model that maps text data into the visual feature space to obtain more reliable prototypes. This allows to exploit data from additional modalities (e.g. text) during training while the ultimate task at test time remains classification with exclusively visual data. We show that in such cases nearest neighbor classification is a viable approach and outperform state-of-the-art single-modal and multimodal few-shot learning methods on the CUB-200 and Oxford-102 datasets.

Colin Samplawski, Jannik Wolff, Tassilo Klein et al.

The task of zero-shot learning (ZSL) requires correctly predicting the label of samples from classes which were unseen at training time. This is achieved by leveraging side information about class labels, such as label attributes or word embeddings. Recently, attention has shifted to the more realistic task of generalized ZSL (GZSL) where test sets consist of seen and unseen samples. Recent approaches to GZSL have shown the value of generative models, which are used to generate samples from unseen classes. In this work, we incorporate an additional source of side information in the form of a relation graph over labels. We leverage this graph in order to learn a set of prior distributions, which encourage an aligned variational autoencoder (VAE) model to learn embeddings which respect the graph structure. Using this approach we are able to achieve improved performance on the CUB and SUN benchmarks over a strong baseline.

Enrico Fini, Stéphane Lathuilière, Enver Sangineto et al.

Continual Learning (CL) aims to develop agents emulating the human ability to sequentially learn new tasks while being able to retain knowledge obtained from past experiences. In this paper, we introduce the novel problem of Memory-Constrained Online Continual Learning (MC-OCL) which imposes strict constraints on the memory overhead that a possible algorithm can use to avoid catastrophic forgetting. As most, if not all, previous CL methods violate these constraints, we propose an algorithmic solution to MC-OCL: Batch-level Distillation (BLD), a regularization-based CL approach, which effectively balances stability and plasticity in order to learn from data streams, while preserving the ability to solve old tasks through distillation. Our extensive experimental evaluation, conducted on three publicly available benchmarks, empirically demonstrates that our approach successfully addresses the MC-OCL problem and achieves comparable accuracy to prior distillation methods requiring higher memory overhead.

Tassilo Klein, Moin Nabi

We propose a self-supervised method to solve Pronoun Disambiguation and Winograd Schema Challenge problems. Our approach exploits the characteristic structure of training corpora related to so-called "trigger" words, which are responsible for flipping the answer in pronoun disambiguation. We achieve such commonsense reasoning by constructing pair-wise contrastive auxiliary predictions. To this end, we leverage a mutual exclusive loss regularized by a contrastive margin. Our architecture is based on the recently introduced transformer networks, BERT, that exhibits strong performance on many NLP benchmarks. Empirical results show that our method alleviates the limitation of current supervised approaches for commonsense reasoning. This study opens up avenues for exploiting inexpensive self-supervision to achieve performance gain in commonsense reasoning tasks.

Aiham Taleb, Christoph Lippert, Tassilo Klein et al.

Self-supervised learning approaches leverage unlabeled samples to acquire generic knowledge about different concepts, hence allowing for annotation-efficient downstream task learning. In this paper, we propose a novel self-supervised method that leverages multiple imaging modalities. We introduce the multimodal puzzle task, which facilitates rich representation learning from multiple image modalities. The learned representations allow for subsequent fine-tuning on different downstream tasks. To achieve that, we learn a modality-agnostic feature embedding by confusing image modalities at the data-level. Together with the Sinkhorn operator, with which we formulate the puzzle solving optimization as permutation matrix inference instead of classification, they allow for efficient solving of multimodal puzzles with varying levels of complexity. In addition, we also propose to utilize cross-modal generation techniques for multimodal data augmentation used for training self-supervised tasks. In other words, we exploit synthetic images for self-supervised pretraining, instead of downstream tasks directly, in order to circumvent quality issues associated with synthetic images, while improving data-efficiency and representations quality. Our experimental results, which assess the gains in downstream performance and data-efficiency, show that solving our multimodal puzzles yields better semantic representations, compared to treating each modality independently. Our results also highlight the benefits of exploiting synthetic images for self-supervised pretraining. We showcase our approach on four downstream tasks: Brain tumor segmentation and survival days prediction using four MRI modalities, Prostate segmentation using two MRI modalities, and Liver segmentation using unregistered CT and MRI modalities. We outperform many previous solutions, and achieve results competitive to state-of-the-art.

Abdullah Salama, Oleksiy Ostapenko, Tassilo Klein et al.

Deep Learning models have become the dominant approach in several areas due to their high performance. Unfortunately, the size and hence computational requirements of operating such models can be considerably high. Therefore, this constitutes a limitation for deployment on memory and battery constrained devices such as mobile phones or embedded systems. To address these limitations, we propose a novel and simple pruning method that compresses neural networks by removing entire filters and neurons according to a global threshold across the network without any pre-calculation of layer sensitivity. The resulting model is compact, non-sparse, with the same accuracy as the non-compressed model, and most importantly requires no special infrastructure for deployment. We prove the viability of our method by producing highly compressed models, namely VGG-16, ResNet-56, and ResNet-110 respectively on CIFAR10 without losing any performance compared to the baseline, as well as ResNet-34 and ResNet-50 on ImageNet without a significant loss of accuracy. We also provide a well-retrained 30% compressed ResNet-50 that slightly surpasses the base model accuracy. Additionally, compressing more than 56% and 97% of AlexNet and LeNet-5 respectively. Interestingly, the resulted models' pruning patterns are highly similar to the other methods using layer sensitivity pre-calculation step. Our method does not only exhibit good performance but what is more also easy to implement.

Tassilo Klein, Moin Nabi

Automatic question generation aims at the generation of questions from a context, with the corresponding answers being sub-spans of the given passage. Whereas, most of the methods mostly rely on heuristic rules to generate questions, more recently also neural network approaches have been proposed. In this work, we propose a variant of the self-attention Transformer network architectures model to generate meaningful and diverse questions. To this end, we propose an easy to use model consisting of the conjunction of the Transformer decoder GPT-2 model with Transformer encoder BERT for the downstream task for question answering. The model is trained in an end-to-end fashion, where the language model is trained to produce a question-answer-aware input representation that facilitates to generate an answer focused question. Our result of neural question generation from text on the SQuAD 1.1 dataset suggests that our method can produce semantically correct and diverse questions. Additionally, we assessed the performance of our proposed method for the downstream task of question answering. The analysis shows that our proposed generation & answering collaboration framework relatively improves both tasks and is particularly powerful in the semi-supervised setup. The results further suggest a robust and comparably lean pipeline facilitating question generation in the small-data regime.

Tassilo Klein, Moin Nabi

The recently introduced BERT model exhibits strong performance on several language understanding benchmarks. In this paper, we describe a simple re-implementation of BERT for commonsense reasoning. We show that the attentions produced by BERT can be directly utilized for tasks such as the Pronoun Disambiguation Problem and Winograd Schema Challenge. Our proposed attention-guided commonsense reasoning method is conceptually simple yet empirically powerful. Experimental analysis on multiple datasets demonstrates that our proposed system performs remarkably well on all cases while outperforming the previously reported state of the art by a margin. While results suggest that BERT seems to implicitly learn to establish complex relationships between entities, solving commonsense reasoning tasks might require more than unsupervised models learned from huge text corpora.

Rodrigo Berriel, Stéphane Lathuilière, Moin Nabi et al.

Multi-Domain Learning (MDL) refers to the problem of learning a set of models derived from a common deep architecture, each one specialized to perform a task in a certain domain (e.g., photos, sketches, paintings). This paper tackles MDL with a particular interest in obtaining domain-specific models with an adjustable budget in terms of the number of network parameters and computational complexity. Our intuition is that, as in real applications the number of domains and tasks can be very large, an effective MDL approach should not only focus on accuracy but also on having as few parameters as possible. To implement this idea we derive specialized deep models for each domain by adapting a pre-trained architecture but, differently from other methods, we propose a novel strategy to automatically adjust the computational complexity of the network. To this aim, we introduce Budget-Aware Adapters that select the most relevant feature channels to better handle data from a novel domain. Some constraints on the number of active switches are imposed in order to obtain a network respecting the desired complexity budget. Experimentally, we show that our approach leads to recognition accuracy competitive with state-of-the-art approaches but with much lighter networks both in terms of storage and computation.

Oleksiy Ostapenko, Mihai Puscas, Tassilo Klein et al.

Models trained in the context of continual learning (CL) should be able to learn from a stream of data over an undefined period of time. The main challenges herein are: 1) maintaining old knowledge while simultaneously benefiting from it when learning new tasks, and 2) guaranteeing model scalability with a growing amount of data to learn from. In order to tackle these challenges, we introduce Dynamic Generative Memory (DGM) - a synaptic plasticity driven framework for continual learning. DGM relies on conditional generative adversarial networks with learnable connection plasticity realized with neural masking. Specifically, we evaluate two variants of neural masking: applied to (i) layer activations and (ii) to connection weights directly. Furthermore, we propose a dynamic network expansion mechanism that ensures sufficient model capacity to accommodate for continually incoming tasks. The amount of added capacity is determined dynamically from the learned binary mask. We evaluate DGM in the continual class-incremental setup on visual classification tasks.

Frederik Pahde, Mihai Puscas, Jannik Wolff et al.

Since the advent of deep learning, neural networks have demonstrated remarkable results in many visual recognition tasks, constantly pushing the limits. However, the state-of-the-art approaches are largely unsuitable in scarce data regimes. To address this shortcoming, this paper proposes employing a 3D model, which is derived from training images. Such a model can then be used to hallucinate novel viewpoints and poses for the scarce samples of the few-shot learning scenario. A self-paced learning approach allows for the selection of a diverse set of high-quality images, which facilitates the training of a classifier. The performance of the proposed approach is showcased on the fine-grained CUB-200-2011 dataset in a few-shot setting and significantly improves our baseline accuracy.

Frederik Pahde, Oleksiy Ostapenko, Patrick Jähnichen et al.

State-of-the-art deep learning algorithms yield remarkable results in many visual recognition tasks. However, they still fail to provide satisfactory results in scarce data regimes. To a certain extent this lack of data can be compensated by multimodal information. Missing information in one modality of a single data point (e.g. an image) can be made up for in another modality (e.g. a textual description). Therefore, we design a few-shot learning task that is multimodal during training (i.e. image and text) and single-modal during test time (i.e. image). In this regard, we propose a self-paced class-discriminative generative adversarial network incorporating multimodality in the context of few-shot learning. The proposed approach builds upon the idea of cross-modal data generation in order to alleviate the data sparsity problem. We improve few-shot learning accuracies on the finegrained CUB and Oxford-102 datasets.

Shailza Jolly, Sandro Pezzelle, Tassilo Klein et al.

We introduce MASSES, a simple evaluation metric for the task of Visual Question Answering (VQA). In its standard form, the VQA task is operationalized as follows: Given an image and an open-ended question in natural language, systems are required to provide a suitable answer. Currently, model performance is evaluated by means of a somehow simplistic metric: If the predicted answer is chosen by at least 3 human annotators out of 10, then it is 100% correct. Though intuitively valuable, this metric has some important limitations. First, it ignores whether the predicted answer is the one selected by the Majority (MA) of annotators. Second, it does not account for the quantitative Subjectivity (S) of the answers in the sample (and dataset). Third, information about the Semantic Similarity (SES) of the responses is completely neglected. Based on such limitations, we propose a multi-component metric that accounts for all these issues. We show that our metric is effective in providing a more fine-grained evaluation both on the quantitative and qualitative level.

Frederik Pahde, Patrick Jähnichen, Tassilo Klein et al.

State-of-the-art deep learning algorithms generally require large amounts of data for model training. Lack thereof can severely deteriorate the performance, particularly in scenarios with fine-grained boundaries between categories. To this end, we propose a multimodal approach that facilitates bridging the information gap by means of meaningful joint embeddings. Specifically, we present a benchmark that is multimodal during training (i.e. images and texts) and single-modal in testing time (i.e. images), with the associated task to utilize multimodal data in base classes (with many samples), to learn explicit visual classifiers for novel classes (with few samples). Next, we propose a framework built upon the idea of cross-modal data hallucination. In this regard, we introduce a discriminative text-conditional GAN for sample generation with a simple self-paced strategy for sample selection. We show the results of our proposed discriminative hallucinated method for 1-, 2-, and 5- shot learning on the CUB dataset, where the accuracy is improved by employing multimodal data.

Robin C. Geyer, Tassilo Klein, Moin Nabi

Federated learning is a recent advance in privacy protection. In this context, a trusted curator aggregates parameters optimized in decentralized fashion by multiple clients. The resulting model is then distributed back to all clients, ultimately converging to a joint representative model without explicitly having to share the data. However, the protocol is vulnerable to differential attacks, which could originate from any party contributing during federated optimization. In such an attack, a client's contribution during training and information about their data set is revealed through analyzing the distributed model. We tackle this problem and propose an algorithm for client sided differential privacy preserving federated optimization. The aim is to hide clients' contributions during training, balancing the trade-off between privacy loss and model performance. Empirical studies suggest that given a sufficiently large number of participating clients, our proposed procedure can maintain client-level differential privacy at only a minor cost in model performance.

Mahdyar Ravanbakhsh, Moin Nabi, Enver Sangineto et al.

In this paper we address the abnormality detection problem in crowded scenes. We propose to use Generative Adversarial Nets (GANs), which are trained using normal frames and corresponding optical-flow images in order to learn an internal representation of the scene normality. Since our GANs are trained with only normal data, they are not able to generate abnormal events. At testing time the real data are compared with both the appearance and the motion representations reconstructed by our GANs and abnormal areas are detected by computing local differences. Experimental results on challenging abnormality detection datasets show the superiority of the proposed method compared to the state of the art in both frame-level and pixel-level abnormality detection tasks.

Mahdyar Ravanbakhsh, Enver Sangineto, Moin Nabi et al.

Abnormal crowd behaviour detection attracts a large interest due to its importance in video surveillance scenarios. However, the ambiguity and the lack of sufficient abnormal ground truth data makes end-to-end training of large deep networks hard in this domain. In this paper we propose to use Generative Adversarial Nets (GANs), which are trained to generate only the normal distribution of the data. During the adversarial GAN training, a discriminator (D) is used as a supervisor for the generator network (G) and vice versa. At testing time we use D to solve our discriminative task (abnormality detection), where D has been trained without the need of manually-annotated abnormal data. Moreover, in order to prevent G learn a trivial identity function, we use a cross-channel approach, forcing G to transform raw-pixel data in motion information and vice versa. The quantitative results on standard benchmarks show that our method outperforms previous state-of-the-art methods in both the frame-level and the pixel-level evaluation.

Ravi Shekhar, Sandro Pezzelle, Yauhen Klimovich et al.

In this paper, we aim to understand whether current language and vision (LaVi) models truly grasp the interaction between the two modalities. To this end, we propose an extension of the MSCOCO dataset, FOIL-COCO, which associates images with both correct and "foil" captions, that is, descriptions of the image that are highly similar to the original ones, but contain one single mistake ("foil word"). We show that current LaVi models fall into the traps of this data and perform badly on three tasks: a) caption classification (correct vs. foil); b) foil word detection; c) foil word correction. Humans, in contrast, have near-perfect performance on those tasks. We demonstrate that merely utilising language cues is not enough to model FOIL-COCO and that it challenges the state-of-the-art by requiring a fine-grained understanding of the relation between text and image.

Mahdyar Ravanbakhsh, Hossein Mousavi, Moin Nabi et al.

In this paper we introduce a novel method for segmentation that can benefit from general semantics of Convolutional Neural Network (CNN). Our segmentation proposes visually and semantically coherent image segments. We use binary encoding of CNN features to overcome the difficulty of the clustering on the high-dimensional CNN feature space. These binary encoding can be embedded into the CNN as an extra layer at the end of the network. This results in real-time segmentation. To the best of our knowledge our method is the first attempt on general semantic image segmentation using CNN. All the previous papers were limited to few number of category of the images (e.g. PASCAL VOC). Experiments show that our segmentation algorithm outperform the state-of-the-art non-semantic segmentation methods by a large margin.

Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi et al.

Most of the crowd abnormal event detection methods rely on complex hand-crafted features to represent the crowd motion and appearance. Convolutional Neural Networks (CNN) have shown to be a powerful tool with excellent representational capacities, which can leverage the need for hand-crafted features. In this paper, we show that keeping track of the changes in the CNN feature across time can facilitate capturing the local abnormality. We specifically propose a novel measure-based method which allows measuring the local abnormality in a video by combining semantic information (inherited from existing CNN models) with low-level Optical-Flow. One of the advantage of this method is that it can be used without the fine-tuning costs. The proposed method is validated on challenging abnormality detection datasets and the results show the superiority of our method compared to the state-of-the-art methods.

Mahdyar Ravanbakhsh, Hossein Mousavi, Moin Nabi et al.

In this paper we introduce a novel method for general semantic segmentation that can benefit from general semantics of Convolutional Neural Network (CNN). Our segmentation proposes visually and semantically coherent image segments. We use binary encoding of CNN features to overcome the difficulty of the clustering on the high-dimensional CNN feature space. These binary codes are very robust against noise and non-semantic changes in the image. These binary encoding can be embedded into the CNN as an extra layer at the end of the network. This results in real-time segmentation. To the best of our knowledge our method is the first attempt on general semantic image segmentation using CNN. All the previous papers were limited to few number of category of the images (e.g. PASCAL VOC). Experiments show that our segmentation algorithm outperform the state-of-the-art non-semantic segmentation methods by large margin.

Hamidreza Rabiee, Javad Haddadnia, Hossein Mousavi et al.

In crowd behavior understanding, a model of crowd behavior need to be trained using the information extracted from video sequences. Since there is no ground-truth available in crowd datasets except the crowd behavior labels, most of the methods proposed so far are just based on low-level visual features. However, there is a huge semantic gap between low-level motion/appearance features and high-level concept of crowd behaviors. In this paper we propose an attribute-based strategy to alleviate this problem. While similar strategies have been recently adopted for object and action recognition, as far as we know, we are the first showing that the crowd emotions can be used as attributes for crowd behavior understanding. The main idea is to train a set of emotion-based classifiers, which can subsequently be used to represent the crowd motion. For this purpose, we collect a big dataset of video clips and provide them with both annotations of "crowd behaviors" and "crowd emotions". We show the results of the proposed method on our dataset, which demonstrate that the crowd emotions enable the construction of more descriptive models for crowd behaviors. We aim at publishing the dataset with the article, to be used as a benchmark for the communities.

Enver Sangineto, Moin Nabi, Dubravko Culibrk et al.

In a weakly-supervised scenario object detectors need to be trained using image-level annotation alone. Since bounding-box-level ground truth is not available, most of the solutions proposed so far are based on an iterative, Multiple Instance Learning framework in which the current classifier is used to select the highest-confidence boxes in each image, which are treated as pseudo-ground truth in the next training iteration. However, the errors of an immature classifier can make the process drift, usually introducing many of false positives in the training dataset. To alleviate this problem, we propose in this paper a training protocol based on the self-paced learning paradigm. The main idea is to iteratively select a subset of images and boxes that are the most reliable, and use them for training. While in the past few years similar strategies have been adopted for SVMs and other classifiers, we are the first showing that a self-paced approach can be used with deep-network-based classifiers in an end-to-end training pipeline. The method we propose is built on the fully-supervised Fast-RCNN architecture and can be applied to similar architectures which represent the input image as a bag of boxes. We show state-of-the-art results on Pascal VOC 2007, Pascal VOC 2010 and ILSVRC 2013. On ILSVRC 2013 our results based on a low-capacity AlexNet network outperform even those weakly-supervised approaches which are based on much higher-capacity networks.

Moin Nabi

Visual Recognition is one of the fundamental challenges in AI, where the goal is to understand the semantics of visual data. Employing mid-level representation, in particular, shifted the paradigm in visual recognition. The mid-level image/video representation involves discovering and training a set of mid-level visual patterns (e.g., parts and attributes) and represent a given image/video utilizing them. The mid-level patterns can be extracted from images and videos using the motion and appearance information of visual phenomenas. This thesis targets employing mid-level representations for different high-level visual recognition tasks, namely (i)image understanding and (ii)video understanding. In the case of image understanding, we focus on object detection/recognition task. We investigate on discovering and learning a set of mid-level patches to be used for representing the images of an object category. We specifically employ the discriminative patches in a subcategory-aware webly-supervised fashion. We, additionally, study the outcomes provided by employing the subcategory-based models for undoing dataset bias.