Search

Activation functions are one of the key components of a neural network. The most commonly used activation functions can be classed into the category of continuously differentiable (e.g. tanh) and linear-unit functions (e.g. ReLU), both having their own strengths and drawbacks with respect to downstream performance and representation capacity through learning (e.g. measured by the number of dead neurons and the effective rank). In reinforcement learning, the performance of continuously differentiable activations often falls short as compared to linear-unit functions. From the perspective of the activations in the last hidden layer, this paper provides insights regarding this sub-optimality and explores how activation functions influence the occurrence of dead neurons and the magnitude of the effective rank. Additionally, a novel neural architecture is proposed that leverages the product of independent activation values. In the Atari domain, we show faster learning, a reduction in dead neurons and increased effective rank., Comment: 22 pages, 17 figures, 4 tables

As telecommunication systems evolve to meet increasing demands, integrating deep neural networks (DNNs) has shown promise in enhancing performance. However, the trade-off between accuracy and flexibility remains challenging when replacing traditional receivers with DNNs. This paper introduces a novel probabilistic framework that allows a single DNN demapper to demap multiple QAM and APSK constellations simultaneously. We also demonstrate that our framework allows exploiting hierarchical relationships in families of constellations. The consequence is that we need fewer neural network outputs to encode the same function without an increase in Bit Error Rate (BER). Our simulation results confirm that our approach approaches the optimal demodulation error bound under an Additive White Gaussian Noise (AWGN) channel for multiple constellations. Thereby, we address multiple important issues in making DNNs flexible enough for practical use as receivers., Comment: To appear in the ICMLCN 2024 proceedings

Encryption on the internet with the shift to HTTPS has been an important step to improve the privacy of internet users. However, there is an increasing body of work about extracting information from encrypted internet traffic without having to decrypt it. Such attacks bypass security guarantees assumed to be given by HTTPS and thus need to be understood. Prior works showed that the variable bitrates of video streams are sufficient to identify which video someone is watching. These works generally have to make trade-offs in aspects such as accuracy, scalability, robustness, etc. These trade-offs complicate the practical use of these attacks. To that end, we propose a deep metric learning framework based on the triplet loss method. Through this framework, we achieve robust, generalisable, scalable and transferable encrypted video stream detection. First, the triplet loss is better able to deal with video streams not seen during training. Second, our approach can accurately classify videos not seen during training. Third, we show that our method scales well to a dataset of over 1000 videos. Finally, we show that a model trained on video streams over Chrome can also classify streams over Firefox. Our results suggest that this side-channel attack is more broadly applicable than originally thought. We provide our code alongside a diverse and up-to-date dataset for future research., Comment: Published in the WI-IAT 2023 proceedings

Coping with distributional shifts is an important part of transfer learning methods in order to perform well in real-life tasks. However, most of the existing approaches in this area either focus on an ideal scenario in which the data does not contain noises or employ a complicated training paradigm or model design to deal with distributional shifts. In this paper, we revisit the robustness of the minimum error entropy (MEE) criterion, a widely used objective in statistical signal processing to deal with non-Gaussian noises, and investigate its feasibility and usefulness in real-life transfer learning regression tasks, where distributional shifts are common. Specifically, we put forward a new theoretical result showing the robustness of MEE against covariate shift. We also show that by simply replacing the mean squared error (MSE) loss with the MEE on basic transfer learning algorithms such as fine-tuning and linear probing, we can achieve competitive performance with respect to state-of-the-art transfer learning algorithms. We justify our arguments on both synthetic data and 5 real-world time-series data., Comment: Manuscript accepted at ECAI-23. Code available at https://github.com/lpsilvestrin/mee-finetune

Nowadays, the deployment of deep learning models on edge devices for addressing real-world classification problems is becoming more prevalent. Moreover, there is a growing popularity in the approach of early classification, a technique that involves classifying the input data after observing only an early portion of it, aiming to achieve reduced communication and computation requirements, which are crucial parameters in edge intelligence environments. While early classification in the field of time series analysis has been broadly researched, existing solutions for multivariate time series problems primarily focus on early classification along the temporal dimension, treating the multiple input channels in a collective manner. In this study, we propose a more flexible early classification pipeline that offers a more granular consideration of input channels and extends the early classification paradigm to the channel dimension. To implement this method, we utilize reinforcement learning techniques and introduce constraints to ensure the feasibility and practicality of our objective. To validate its effectiveness, we conduct experiments using synthetic data and we also evaluate its performance on real datasets. The comprehensive results from our experiments demonstrate that, for multiple datasets, our method can enhance the early classification paradigm by achieving improved accuracy for equal input utilization.

Performant Convolutional Neural Network (CNN) architectures must be tailored to specific tasks in order to consider the length, resolution, and dimensionality of the input data. In this work, we tackle the need for problem-specific CNN architectures. We present the Continuous Convolutional Neural Network (CCNN): a single CNN able to process data of arbitrary resolution, dimensionality and length without any structural changes. Its key component are its continuous convolutional kernels which model long-range dependencies at every layer, and thus remove the need of current CNN architectures for task-dependent downsampling and depths. We showcase the generality of our method by using the same architecture for tasks on sequential ($1{\rm D}$), visual ($2{\rm D}$) and point-cloud ($3{\rm D}$) data. Our CCNN matches and often outperforms the current state-of-the-art across all tasks considered.

An eco-system of agents each having their own policy with some, but limited, generalizability has proven to be a reliable approach to increase generalization across procedurally generated environments. In such an approach, new agents are regularly added to the eco-system when encountering a new environment that is outside of the scope of the eco-system. The speed of adaptation and general effectiveness of the eco-system approach highly depends on the initialization of new agents. In this paper we propose different initialization techniques, inspired from Deep Neural Network initialization and transfer learning, and study their impact., Comment: 12 pages

In the context of MDPs with high-dimensional states, downstream tasks are predominantly applied on a compressed, low-dimensional representation of the original input space. A variety of learning objectives have therefore been used to attain useful representations. However, these representations usually lack interpretability of the different features. We present a novel approach that is able to disentangle latent features into a controllable and an uncontrollable partition. We illustrate that the resulting partitioned representations are easily interpretable on three types of environments and show that, in a distribution of procedurally generated maze environments, it is feasible to interpretably employ a planning algorithm in the isolated controllable latent partition., Comment: 14 pages (8 main paper pages), 15 figures

In current research, machine and deep learning solutions for the classification of temporal data are shifting from single-channel datasets (univariate) to problems with multiple channels of information (multivariate). The majority of these works are focused on the method novelty and architecture, and the format of the input data is often treated implicitly. Particularly, multivariate datasets are often treated as a stack of univariate time series in terms of input preprocessing, with scaling methods applied across each channel separately. In this evaluation, we aim to demonstrate that the additional channel dimension is far from trivial and different approaches to scaling can lead to significantly different results in the accuracy of a solution. To that end, we test seven different data transformation methods on four different temporal dimensions and study their effect on the classification accuracy of five recent methods. We show that, for the large majority of tested datasets, the best transformation-dimension configuration leads to an increase in the accuracy compared to the result of each model with the same hyperparameters and no scaling, ranging from 0.16 to 76.79 percentage points. We also show that if we keep the transformation method constant, there is a statistically significant difference in accuracy results when applying it across different dimensions, with accuracy differences ranging from 0.23 to 47.79 percentage points. Finally, we explore the relation of the transformation methods and dimensions to the classifiers, and we conclude that there is no prominent general trend, and the optimal configuration is dataset- and classifier-specific., Comment: IEEE ICMLA 2022

The use of Convolutional Neural Networks (CNNs) is widespread in Deep Learning due to a range of desirable model properties which result in an efficient and effective machine learning framework. However, performant CNN architectures must be tailored to specific tasks in order to incorporate considerations such as the input length, resolution, and dimentionality. In this work, we overcome the need for problem-specific CNN architectures with our Continuous Convolutional Neural Network (CCNN): a single CNN architecture equipped with continuous convolutional kernels that can be used for tasks on data of arbitrary resolution, dimensionality and length without structural changes. Continuous convolutional kernels model long range dependencies at every layer, and remove the need for downsampling layers and task-dependent depths needed in current CNN architectures. We show the generality of our approach by applying the same CCNN to a wide set of tasks on sequential (1$\mathrm{D}$) and visual data (2$\mathrm{D}$). Our CCNN performs competitively and often outperforms the current state-of-the-art across all tasks considered., Comment: First two authors contributed equally to this work

Adapting a Reinforcement Learning (RL) agent to an unseen environment is a difficult task due to typical over-fitting on the training environment. RL agents are often capable of solving environments very close to the trained environment, but when environments become substantially different, their performance quickly drops. When agents are retrained on new environments, a second issue arises: there is a risk of catastrophic forgetting, where the performance on previously seen environments is seriously hampered. This paper proposes a novel approach that exploits an eco-system of agents to address both concerns. Hereby, the (limited) adaptive power of individual agents is harvested to build a highly adaptive eco-system., Comment: 9 pages, 12 figures

Nowadays, with the rising number of sensors in sectors such as healthcare and industry, the problem of multivariate time series classification (MTSC) is getting increasingly relevant and is a prime target for machine and deep learning approaches. Their expanding adoption in real-world environments is causing a shift in focus from the pursuit of ever-higher prediction accuracy with complex models towards practical, deployable solutions that balance accuracy and parameters such as prediction speed. An MTSC model that has attracted attention recently is ROCKET, based on random convolutional kernels, both because of its very fast training process and its state-of-the-art accuracy. However, the large number of features it utilizes may be detrimental to inference time. Examining its theoretical background and limitations enables us to address potential drawbacks and present LightWaveS: a framework for accurate MTSC, which is fast both during training and inference. Specifically, utilizing wavelet scattering transformation and distributed feature selection, we manage to create a solution that employs just 2.5% of the ROCKET features, while achieving accuracy comparable to recent MTSC models. LightWaveS also scales well across multiple compute nodes and with the number of input channels during training. In addition, it can significantly reduce the input size and provide insight to an MTSC problem by keeping only the most useful channels. We present three versions of our algorithm and their results on distributed training time and scalability, accuracy, and inference speedup. We show that we achieve speedup ranging from 9x to 53x compared to ROCKET during inference on an edge device, on datasets with comparable accuracy., Comment: This work has been accepted as a short paper at DCOSS 2022

Novel prediction methods should always be compared to a baseline to know how well they perform. Without this frame of reference, the performance score of a model is basically meaningless. What does it mean when a model achieves an $F_1$ of 0.8 on a test set? A proper baseline is needed to evaluate the `goodness' of a performance score. Comparing with the latest state-of-the-art model is usually insightful. However, being state-of-the-art can change rapidly when newer models are developed. Contrary to an advanced model, a simple dummy classifier could be used. However, the latter could be beaten too easily, making the comparison less valuable. This paper presents a universal baseline method for all binary classification models, named the Dutch Draw (DD). This approach weighs simple classifiers and determines the best classifier to use as a baseline. We theoretically derive the DD baseline for many commonly used evaluation measures and show that in most situations it reduces to (almost) always predicting either zero or one. Summarizing, the DD baseline is: (1) general, as it is applicable to all binary classification problems; (2) simple, as it is quickly determined without training or parameter-tuning; (3) informative, as insightful conclusions can be drawn from the results. The DD baseline serves two purposes. First, to enable comparisons across research papers by this robust and universal baseline. Secondly, to provide a sanity check during the development process of a prediction model. It is a major warning sign when a model is outperformed by the DD baseline.

With the development of new sensors and monitoring devices, more sources of data become available to be used as inputs for machine learning models. These can on the one hand help to improve the accuracy of a model. On the other hand, combining these new inputs with historical data remains a challenge that has not yet been studied in enough detail. In this work, we propose a transfer learning algorithm that combines new and historical data with different input dimensions. This approach is easy to implement, efficient, with computational complexity equivalent to the ordinary least-squares method, and requires no hyperparameter tuning, making it straightforward to apply when the new data is limited. Different from other approaches, we provide a rigorous theoretical study of its robustness, showing that it cannot be outperformed by a baseline that utilizes only the new data. Our approach achieves state-of-the-art performance on 9 real-life datasets, outperforming the linear DSFT, another linear transfer learning algorithm, and performing comparably to non-linear DSFT., Comment: Manuscript accepted for publication at the Journal of Computational Mathematics and Data Science. Code available at https://github.com/lpsilvestrin/incremental_input_tl

When designing Convolutional Neural Networks (CNNs), one must select the size\break of the convolutional kernels before training. Recent works show CNNs benefit from different kernel sizes at different layers, but exploring all possible combinations is unfeasible in practice. A more efficient approach is to learn the kernel size during training. However, existing works that learn the kernel size have a limited bandwidth. These approaches scale kernels by dilation, and thus the detail they can describe is limited. In this work, we propose FlexConv, a novel convolutional operation with which high bandwidth convolutional kernels of learnable kernel size can be learned at a fixed parameter cost. FlexNets model long-term dependencies without the use of pooling, achieve state-of-the-art performance on several sequential datasets, outperform recent works with learned kernel sizes, and are competitive with much deeper ResNets on image benchmark datasets. Additionally, FlexNets can be deployed at higher resolutions than those seen during training. To avoid aliasing, we propose a novel kernel parameterization with which the frequency of the kernels can be analytically controlled. Our novel kernel parameterization shows higher descriptive power and faster convergence speed than existing parameterizations. This leads to important improvements in classification accuracy., Comment: First two authors contributed equally to this work

Over the past decade, the advent of cybercrime has accelarated the research on cybersecurity. However, the deployment of intrusion detection methods falls short. One of the reasons for this is the lack of realistic evaluation datasets, which makes it a challenge to develop techniques and compare them. This is caused by the large amounts of effort it takes for a cyber analyst to classify network connections. This has raised the need for methods (i) that can learn from small sets of labeled data, (ii) that can make predictions on large sets of unlabeled data, and (iii) that request the label of only specially selected unlabeled data instances. Hence, Active Learning (AL) methods are of interest. These approaches choose specific unlabeled instances by a query function that are expected to improve overall classification performance. The resulting query observations are labeled by a human expert and added to the labeled set. In this paper, we propose a new hybrid AL method called Jasmine. Firstly, it determines how suitable each observation is for querying, i.e., how likely it is to enhance classification. These properties are the uncertainty score and anomaly score. Secondly, Jasmine introduces dynamic updating. This allows the model to adjust the balance between querying uncertain, anomalous and randomly selected observations. To this end, Jasmine is able to learn the best query strategy during the labeling process. This is in contrast to the other AL methods in cybersecurity that all have static, predetermined query functions. We show that dynamic updating, and therefore Jasmine, is able to consistently obtain good and more robust results than querying only uncertainties, only anomalies or a fixed combination of the two., Comment: 35 pages, 15 figures

Time-series forecasting plays an important role in many domains. Boosted by the advances in Deep Learning algorithms, it has for instance been used to predict wind power for eolic energy production, stock market fluctuations, or motor overheating. In some of these tasks, we are interested in predicting accurately some particular moments which often are underrepresented in the dataset, resulting in a problem known as imbalanced regression. In the literature, while recognized as a challenging problem, limited attention has been devoted on how to handle the problem in a practical setting. In this paper, we put forward a general approach to analyze time-series forecasting problems focusing on those underrepresented moments to reduce imbalances. Our approach has been developed based on a case study in a large industrial company, which we use to exemplify the approach.

While reinforcement learning (RL) has proven to be the approach of choice for tackling many complex problems, it remains challenging to develop and deploy RL agents in real-life scenarios successfully. This paper presents pH-RL (personalization in e-Health with RL) a general RL architecture for personalization to bring RL to health practice. pH-RL allows for various levels of personalization in health applications and allows for online and batch learning. Furthermore, we provide a general-purpose implementation framework that can be integrated with various healthcare applications. We describe a step-by-step guideline for the successful deployment of RL policies in a mobile application. We implemented our open-source RL architecture and integrated it with the MoodBuster mobile application for mental health to provide messages to increase daily adherence to the online therapeutic modules. We then performed a comprehensive study with human participants over a sustained period. Our experimental results show that the developed policies learn to select appropriate actions consistently using only a few days' worth of data. Furthermore, we empirically demonstrate the stability of the learned policies during the study.

Conventional neural architectures for sequential data present important limitations. Recurrent networks suffer from exploding and vanishing gradients, small effective memory horizons, and must be trained sequentially. Convolutional networks are unable to handle sequences of unknown size and their memory horizon must be defined a priori. In this work, we show that all these problems can be solved by formulating convolutional kernels in CNNs as continuous functions. The resulting Continuous Kernel Convolution (CKConv) allows us to model arbitrarily long sequences in a parallel manner, within a single operation, and without relying on any form of recurrence. We show that Continuous Kernel Convolutional Networks (CKCNNs) obtain state-of-the-art results in multiple datasets, e.g., permuted MNIST, and, thanks to their continuous nature, are able to handle non-uniformly sampled datasets and irregularly-sampled data natively. CKCNNs match or perform better than neural ODEs designed for these purposes in a faster and simpler manner.

Finding well-defined clusters in data represents a fundamental challenge for many data-driven applications, and largely depends on good data representation. Drawing on literature regarding representation learning, studies suggest that one key characteristic of good latent representations is the ability to produce semantically mixed outputs when decoding linear interpolations of two latent representations. We propose the Mixing Consistent Deep Clustering method which encourages interpolations to appear realistic while adding the constraint that interpolations of two data points must look like one of the two inputs. By applying this training method to various clustering (non-)specific autoencoder models we found that using the proposed training method systematically changed the structure of learned representations of a model and it improved clustering performance for the tested ACAI, IDEC, and VAE models on the MNIST, SVHN, and CIFAR-10 datasets. These outcomes have practical implications for numerous real-world clustering tasks, as it shows that the proposed method can be added to existing autoencoders to further improve clustering performance.

Leveraging the symmetries inherent to specific data domains for the construction of equivariant neural networks has lead to remarkable improvements in terms of data efficiency and generalization. However, most existing research focuses on symmetries arising from planar and volumetric data, leaving a crucial data source largely underexplored: time-series. In this work, we fill this gap by leveraging the symmetries inherent to time-series for the construction of equivariant neural network. We identify two core symmetries: *scale and translation*, and construct scale-translation equivariant neural networks for time-series learning. Intriguingly, we find that scale-translation equivariant mappings share strong resemblance with the wavelet transform. Inspired by this resemblance, we term our networks Wavelet Networks, and show that they perform nested non-linear wavelet-like time-frequency transforms. Empirical results show that Wavelet Networks outperform conventional CNNs on raw waveforms, and match strongly engineered spectrogram techniques across several tasks and time-series types, including audio, environmental sounds, and electrical signals. Our code is publicly available at https://github.com/dwromero/wavelet_networks.

Although group convolutional networks are able to learn powerful representations based on symmetry patterns, they lack explicit means to learn meaningful relationships among them (e.g., relative positions and poses). In this paper, we present attentive group equivariant convolutions, a generalization of the group convolution, in which attention is applied during the course of convolution to accentuate meaningful symmetry combinations and suppress non-plausible, misleading ones. We indicate that prior work on visual attention can be described as special cases of our proposed framework and show empirically that our attentive group equivariant convolutional networks consistently outperform conventional group convolutional networks on benchmark image datasets. Simultaneously, we provide interpretability to the learned concepts through the visualization of equivariant attention maps., Comment: Proceedings of the 37th International Conference on Machine Learning (ICML), 2020

Trace clustering has increasingly been applied to find homogenous process executions. However, current techniques have difficulties in finding a meaningful and insightful clustering of patients on the basis of healthcare data. The resulting clusters are often not in line with those of medical experts, nor do the clusters guarantee to help return meaningful process maps of patients' clinical pathways. After all, a single hospital may conduct thousands of distinct activities and generate millions of events per year. In this paper, we propose a novel trace clustering approach by using sample sets of patients provided by medical experts. More specifically, we learn frequent sequence patterns on a sample set, rank each patient based on the patterns, and use an automated approach to determine the corresponding cluster. We find each cluster separately, while the frequent sequence patterns are used to discover a process map. The approach is implemented in ProM and evaluated using a large data set obtained from a university medical center. The evaluation shows F1-scores of 0.7 for grouping kidney injury, 0.9 for diabetes, and 0.64 for head/neck tumor, while the process maps show meaningful behavioral patterns of the clinical pathways of these groups, according to the domain experts.

Equivariance is a nice property to have as it produces much more parameter efficient neural architectures and preserves the structure of the input through the feature mapping. Even though some combinations of transformations might never appear (e.g. an upright face with a horizontal nose), current equivariant architectures consider the set of all possible transformations in a transformation group when learning feature representations. Contrarily, the human visual system is able to attend to the set of relevant transformations occurring in the environment and utilizes this information to assist and improve object recognition. Based on this observation, we modify conventional equivariant feature mappings such that they are able to attend to the set of co-occurring transformations in data and generalize this notion to act on groups consisting of multiple symmetries. We show that our proposed co-attentive equivariant neural networks consistently outperform conventional rotation equivariant and rotation & reflection equivariant neural networks on rotated MNIST and CIFAR-10., Comment: Proceedings of the 8th International Conference on Learning Representations (ICLR), 2020

Language systems have been of great interest to the research community and have recently reached the mass market through various assistant platforms on the web. Reinforcement Learning methods that optimize dialogue policies have seen successes in past years and have recently been extended into methods that personalize the dialogue, e.g. take the personal context of users into account. These works, however, are limited to personalization to a single user with whom they require multiple interactions and do not generalize the usage of context across users. This work introduces a problem where a generalized usage of context is relevant and proposes two Reinforcement Learning (RL)-based approaches to this problem. The first approach uses a single learner and extends the traditional POMDP formulation of dialogue state with features that describe the user context. The second approach segments users by context and then employs a learner per context. We compare these approaches in a benchmark of existing non-RL and RL-based methods in three established and one novel application domain of financial product recommendation. We compare the influence of context and training experiences on performance and find that learning approaches generally outperform a handcrafted gold standard.

The huge wealth of data in the health domain can be exploited to create models that predict development of health states over time. Temporal learning algorithms are well suited to learn relationships between health states and make predictions about their future developments. However, these algorithms: (1) either focus on learning one generic model for all patients, providing general insights but often with limited predictive performance, or (2) learn individualized models from which it is hard to derive generic concepts. In this paper, we present a middle ground, namely parameterized dynamical systems models that are generated from data using a Genetic Programming (GP) framework. A fitness function suitable for the health domain is exploited. An evaluation of the approach in the mental health domain shows that performance of the model generated by the GP is on par with a dynamical systems model developed based on domain knowledge, significantly outperforms a generic Long Term Short Term Memory (LSTM) model and in some cases also outperforms an individualized LSTM model.

Grouping patients meaningfully can give insights about the different types of patients, their needs, and the priorities. Finding groups that are meaningful is however very challenging as background knowledge is often required to determine what a useful grouping is. In this paper we propose an approach that is able to find groups of patients based on a small sample of positive examples given by a domain expert. Because of that, the approach relies on very limited efforts by the domain experts. The approach groups based on the activities and diagnostic/billing codes within health pathways of patients. To define such a grouping based on the sample of patients efficiently, frequent patterns of activities are discovered and used to measure the similarity between the care pathways of other patients to the patients in the sample group. This approach results in an insightful definition of the group. The proposed approach is evaluated using several datasets obtained from a large university medical center. The evaluation shows F1-scores of around 0.7 for grouping kidney injury and around 0.6 for diabetes.

Personalization is very powerful in improving the effectiveness of health interventions. Reinforcement learning (RL) algorithms are suitable for learning these tailored interventions from sequential data collected about individuals. However, learning can be very fragile. The time to learn intervention policies is limited as disengagement from the user can occur quickly. Also, in e-Health intervention timing can be crucial before the optimal window passes. We present an approach that learns tailored personalization policies for groups of users by combining RL and clustering. The benefits are two-fold: speeding up the learning to prevent disengagement while maintaining a high level of personalization. Our clustering approach utilizes dynamic time warping to compare user trajectories consisting of states and rewards. We apply online and batch RL to learn policies over clusters of individuals and introduce our self-developed and publicly available simulator for e-Health interventions to evaluate our approach. We compare our methods with an e-Health intervention benchmark. We demonstrate that batch learning outperforms online learning for our setting. Furthermore, our proposed clustering approach for RL finds near-optimal clusterings which lead to significantly better policies in terms of cumulative reward compared to learning a policy per individual or learning one non-personalized policy across all individuals. Our findings also indicate that the learned policies accurately learn to send interventions at the right moments and that the users workout more and at the right times of the day.