general overhaul, better images, better texts
This commit is contained in:
@@ -6,8 +6,9 @@ excerpt: "We propose an approach to annotate trajectories using sequences of spa
|
||||
header:
|
||||
teaser: assets/figures/0_trajectory_reconstruction_teaser.png
|
||||
---
|
||||
<figure class="half">
|
||||
<img src="/assets/figures/0_trajectory_isovist.jpg" alt="" style="width:48%">
|
||||
<img src="/assets/figures/0_trajectory_reconstruction.jpg" alt="" style="width:48%">
|
||||
</figure>
|
||||
This work establishes a foundation for enhancing interaction between robots and humans in shared spaces by developing reliable systems for verbal communication. It introduces an unsupervised learning method using neural autoencoding to learn continuous spatial representations from trajectory data, enabling clustering of movements based on spatial context. The approach yields semantically meaningful encodings of spatio-temporal data for creating prototypical representations, setting a promising direction for future applications in robotic-human interaction. {% cite feld2018trajectory %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto; width:350px"}
|
||||
In the near future, more and more machines will perform tasks in the vicinity of human spaces or support them directly in their spatially bound activities. In order to simplify the verbal communication and the interaction between robotic units and/or humans, reliable and robust systems w.r.t. noise and processing results are needed. This work builds a foundation to address this task. By using a continuous representation of spatial perception in interiors learned from trajectory data, our approach clusters movement in dependency to its spatial context. We propose an unsupervised learning approach based on a neural autoencoding that learns semantically meaningful continuous encodings of spatio-temporal trajectory data. This learned encoding can be used to form prototypical representations. We present promising results that clear the path for future applications. {% cite feld2018trajectory %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto; width:350px"}
|
||||
|
||||
@@ -7,9 +7,8 @@ header:
|
||||
teaser: assets/figures/1_self_replication_pca_space.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto; width:350px"}
|
||||
{:style="display:block; width:40%" .align-right}
|
||||
|
||||
The foundation of biological structures is self-replication. Neural networks are the prime structure used for the emergent construction of complex behavior in computers. We analyze how various network types lend themselves to self-replication. We argue that backpropagation is the natural way to navigate the space of network weights and show how it allows non-trivial self-replicators to arise naturally. We then extend the setting to construct an artificial chemistry environment of several neural networks.
|
||||
{% cite gabor2019self %}
|
||||
This text discusses the fundamental role of self-replication in biological structures and its application to neural networks for developing complex behaviors in computing. It explores different network types for self-replication, highlighting the effectiveness of backpropagation in navigating network weights and fostering the emergence of non-trivial self-replicators. The study further delves into creating an artificial chemistry environment comprising several neural networks, offering a novel approach to understanding and implementing self-replication in computational models. For in-depth insights, refer to the work by {% cite gabor2019self %}.
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
@@ -7,6 +7,5 @@ header:
|
||||
teaser: assets/figures/3_deep_neural_baselines_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto; width:250px"}
|
||||
Detecting sleepiness from spoken language is an ambitious task, which is addressed by the Interspeech 2019 Computational Paralinguistics Challenge (ComParE). We propose an end-to-end deep learning approach to detect and classify patterns reflecting sleepiness in the human voice. Our approach is based solely on a moderately complex deep neural network architecture. It may be applied directly on the audio data without requiring any specific feature engineering, thus remaining transferable to other audio classification tasks. Nevertheless, our approach performs similar to state-of-the-art machine learning models.
|
||||
{% cite elsner2019deep %}
|
||||
{:style="display:block; width:30%" .align-right}
|
||||
The study presents an innovative end-to-end deep learning method to identify sleepiness in spoken language, as part of the Interspeech 2019 ComParE challenge. This method utilizes a deep neural network architecture to analyze audio data directly, eliminating the need for specific feature engineering. This approach not only achieves performance comparable to state-of-the-art models but is also adaptable to various audio classification tasks. For more details, refer to the work by {% cite elsner2019deep %}.
|
||||
|
||||
@@ -6,8 +6,6 @@ excerpt: "Team market value estimation, similarity search and rankings."
|
||||
header:
|
||||
teaser: assets/figures/2_steve_algo.jpg
|
||||
---
|
||||
{:style="display:block; width:60%" .align-center}
|
||||
|
||||
In this work we present STEVE - Soccer TEam VEctors, a principled approach for learning real valued vectors for soccer teams where similar teams are close to each other in the resulting vector space. STEVE only relies on freely available information about the matches teams played in the past. These vectors can serve as input to various machine learning tasks. Evaluating on the task of team market value estimation, STEVE outperforms all its competitors. Moreover, we use STEVE for similarity search and to rank soccer teams.
|
||||
{% cite muller2020soccer %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto;"}
|
||||
This study introduces STEVE (Soccer Team Vectors), a novel method for generating real-valued vectors representing soccer teams, organized so that similar teams are proximate in vector space. Utilizing publicly available match data, these vectors facilitate various machine learning applications, notably excelling in team market value estimation and enabling effective similarity search and team ranking. STEVE demonstrates superior performance over competing models in these domains. For further details, please consult the work by {% cite muller2020soccer %}.
|
||||
|
||||
@@ -7,9 +7,8 @@ header:
|
||||
teaser: assets/figures/4_point_cloud_segmentation_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto;"}
|
||||
{:style="display:block; width:100%" .align-center}
|
||||
|
||||
The segmentation and fitting of solid primitives to 3D point clouds is a complex task. Existing systems are restricted either in the number of input points or the supported primitive types. This paper proposes a hybrid pipeline that is able to reconstruct spheres, bounded cylinders and rectangular cuboids on large point sets. It uses a combination of deep learning and classical RANSAC for primitive fitting, a DBSCAN-based clustering scheme for increased stability and a specialized Genetic Algorithm for robust cuboid extraction. In a detailed evaluation, its performance metrics are discussed and resulting solid primitive sets are visualized. The paper concludes with a discussion of the approach’s limitations.
|
||||
{% cite friedrich2020hybrid %}
|
||||
This paper introduces a hybrid approach for segmenting and fitting solid primitives to 3D point clouds, overcoming limitations in handling large datasets and diverse primitive shapes. By integrating deep learning with RANSAC for primitive fitting, employing DBSCAN for clustering, and utilizing a specialized Genetic Algorithm for cuboid extraction, this method achieves enhanced stability and robustness. It excels in reconstructing spheres, cylinders, and cuboids from large point sets, with performance metrics and visualizations provided to demonstrate its effectiveness, alongside a discussion on its limitations. For more detailed insights, refer to {% cite friedrich2020hybrid %}.
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto;"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
@@ -7,11 +7,7 @@ header:
|
||||
teaser: assets/figures/6_ood_pipeline.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
One critical prerequisite for the deployment of reinforcement learning systems in the real world is the ability to reliably detect situations on which the agent was not trained. Such situations could lead to potential safety risks when wrong predictions lead to the execution of harmful actions. In this work, we propose PEOC, a new policy entropy based out-of-distribution classifier that reliably detects unencountered states in deep reinforcement learning. It is based on using the entropy of an agent's policy as the classification score of a one-class classifier. We evaluate our approach using a procedural environment generator. Results show that PEOC is highly competitive against state-of-the-art one-class classification algorithms on the evaluated environments. Furthermore, we present a structured process for benchmarking out-of-distribution classification in reinforcement learning.
|
||||
{% cite sedlmeier2020peoc %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
{:style="display:block; width:45%" .align-right}In this work, the development of PEOC, a policy entropy-based classifier for detecting unencountered states in deep reinforcement learning, is proposed. Utilizing the agent's policy entropy as a score, PEOC effectively identifies out-of-distribution scenarios, crucial for ensuring safety in real-world applications. Evaluated against advanced one-class classifiers within procedurally generated environments, PEOC demonstrates competitive performance.
|
||||
Additionally, a structured benchmarking process for out-of-distribution classification in reinforcement learning is presented, offering a comprehensive approach to evaluating such systems' reliability and effectiveness. {% cite sedlmeier2020policy %}
|
||||
|
||||
{:style="display:block; width:90%" .align-center}
|
||||
|
||||
@@ -7,9 +7,9 @@ header:
|
||||
teaser: assets/figures/5_meantime_coverage.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
Fully autonomously driving vehicles are expected to be a widely available technology in the near future. Privately owned cars, which remain parked for the majority of their lifetime, may therefore be capable of driving independently during their usual long parking periods (e.g. their owners working hours). Our analysis aims to focus on the potential of a privately owned shared car concept as transition period between the present usages of privately owned cars towards a transportation paradigm of privately owned shared autonomous vehicles. We propose two methods in the field of reachability analysis to evaluate the impact of such vehicles during parking periods. Our proposed methods are applied to a dataset of parking times of users of a telematics service provider in the Munich area (Germany). We show the impact of time and location dependent effects on the analyzed service coverage, such as business week rush hours or cover age divergence between urban and suburban regions.
|
||||
This analysis explores the concept of privately owned shared autonomous vehicles as a transitional phase towards a new transportation paradigm. It proposes two reachability analysis methods to assess the impact of utilizing privately owned cars during their typical long parking intervals, such as during an owner's work hours. By applying these methods to a dataset from the Munich area, the study reveals how time and location-dependent factors, like rush hours and urban vs. suburban differences, affect service coverage.
|
||||
{% cite illium2020meantime %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
@@ -7,9 +7,8 @@ header:
|
||||
teaser: assets/figures/7_mask_models.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
In many fields of research, labeled data-sets are hard to acquire. This is where data augmentation promises to overcome the lack of training data in the context of neural network engineering and classification tasks. The idea here is to reduce model over-fitting to the feature distribution of a small under-descriptive training data-set. We try to evaluate such data augmentation techniques to gather insights in the performance boost they provide for several convolutional neural networks on mel-spectrogram representations of audio data. We show the impact of data augmentation on the binary classification task of surgical mask detection in samples of human voice. Also we consider four varying architectures to account for augmentation robustness. Results show that most of the baselines given by ComParE are outperformed
|
||||
{% cite illium2020surgical %}
|
||||
This study assesses the effectiveness of data augmentation in enhancing neural network models for audio data classification, focusing on mel-spectrogram representations. Specifically, it examines the role of data augmentation in improving the performance of convolutional neural networks for detecting the presence of surgical masks from human voice samples, testing across four different network architectures. The findings indicate a significant enhancement in model performance, surpassing many of the existing benchmarks established by the ComParE challenge. For further details, refer to {% cite illium2020surgical %}.
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
@@ -7,7 +7,6 @@ header:
|
||||
teaser: assets/figures/8_anomalous_sound_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
The problem of Constructive Solid Geometry (CSG) tree reconstruction from 3D point clouds or 3D triangle meshes is hard to solve. At first, the input data set (point cloud, triangle soup or triangle mesh) has to be segmented and geometric primitives (spheres, cylinders, ...) have to be fitted to each subset. Then, the size- and shape optimal CSG tree has to be extracted. We propose a pipeline for CSG reconstruction consisting of multiple stages: A primitive extraction step, which uses deep learning for primitive detection, a clustered variant of RANSAC for parameter fitting, and a Genetic Algorithm (GA) for convex polytope generation. It directly transforms 3D point clouds or triangle meshes into solid primitives. The filtered primitive set is then used as input for a GA-based CSG extraction stage. We evaluate two different CSG extraction methodologies and furthermore compare our pipeline to current state-of-the-art methods.
|
||||
{:style="display:block; width:40%" .align-right}
|
||||
This study explores the use of pretrained neural networks as feature extractors for detecting anomalous sounds, utilizing these networks to derive semantically rich features for a Gaussian Mixture Model that estimates normality. It examines extractors trained on diverse data domains—images, environmental sounds, and music—applied to industrial noises from machinery. Surprisingly, features based on music data often surpass others, including an autoencoder baseline, suggesting that domain similarity between extractor training and application might not always correlate with performance improvement.
|
||||
{% cite muller2020analysis %}
|
||||
|
||||
@@ -6,10 +6,9 @@ excerpt: "Acoustic Anomaly Detection for Machine Sounds based on Image Transfer
|
||||
header:
|
||||
teaser: assets/figures/9_image_transfer_sound_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
In industrial applications, the early detection of malfunctioning factory machinery is crucial. In this paper, we consider acoustic malfunction detection via transfer learning. Contrary to the majority of current approaches which are based on deep autoencoders, we propose to extract features using neural networks that were pretrained on the task of image classification. We then use these features to train a variety of anomaly detection models and show that this improves results compared to convolutional autoencoders in recordings of four different factory machines in noisy environments. Moreover, we find that features extracted from ResNet based networks yield better results than those from AlexNet and Squeezenet. In our setting, Gaussian Mixture Models and One-Class Support Vector Machines achieve the best anomaly detection performance.
|
||||
{:style="display:block; width:45%" .align-right}
|
||||
This paper explores acoustic malfunction detection in industrial machinery using transfer learning, specifically leveraging neural networks pretrained on image classification to extract features.
|
||||
These features, when used with anomaly detection models, outperform traditional convolutional autoencoders in noisy conditions across different machine types. The study highlights the superiority of features from ResNet architectures over AlexNet and Squeezenet, with Gaussian Mixture Models and One-Class Support Vector Machines showing the best performance in detecting anomalies.
|
||||
{% cite muller2020acoustic %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:85%" .align-center}
|
||||
|
||||
@@ -7,9 +7,9 @@ header:
|
||||
teaser: assets/figures/10_water_networks_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
In industrial applications, the early detection of malfunctioning factory machinery is crucial. In this paper, we consider acoustic malfunction detection via transfer learning. Contrary to the majority of current approaches which are based on deep autoencoders, we propose to extract features using neural networks that were pretrained on the task of image classification. We then use these features to train a variety of anomaly detection models and show that this improves results compared to convolutional autoencoders in recordings of four different factory machines in noisy environments. Moreover, we find that features extracted from ResNet based networks yield better results than those from AlexNet and Squeezenet. In our setting, Gaussian Mixture Models and One-Class Support Vector Machines achieve the best anomaly detection performance.
|
||||
{:style="display:block; width:40%" .align-right}
|
||||
This study introduces a method for acoustic leak detection in water networks, focusing on energy efficiency and easy deployment. Utilizing recordings from microphones on a municipal water network, various anomaly detection models, both shallow and deep, were trained. The approach mimics human leak detection methods, allowing intermittent monitoring instead of constant surveillance. While detecting nearby leaks proved easy for most models, neural network-based methods excelled at identifying leaks from a distance, showcasing their effectiveness in practical applications.
|
||||
{% cite muller2021acoustic %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:85%" .align-center}
|
||||
|
||||
|
||||
@@ -7,9 +7,9 @@ header:
|
||||
teaser: assets/figures/11_recurrent_primate_workflow.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:40%" .align-right}
|
||||
|
||||
Wildlife monitoring is an essential part of most conservation efforts where one of the many building blocks is acoustic monitoring. Acoustic monitoring has the advantage of being noninvasive and applicable in areas of high vegetation. In this work, we present a deep and recurrent architecture for the classification of primate vocalizations that is based upon well proven modules such as bidirectional Long Short-Term Memory neural networks, pooling, normalized softmax and focal loss. Additionally, we apply Bayesian optimization to obtain a suitable set of hyperparameters. We test our approach on a recently published dataset of primate vocalizations that were recorded in an African wildlife sanctuary.
|
||||
This study introduces a deep, recurrent architecture for classifying primate vocalizations, leveraging bidirectional Long Short-Term Memory networks and advanced techniques like normalized softmax and focal loss. Bayesian optimization was used to fine-tune hyperparameters, and the model was evaluated on a dataset of primate calls from an African sanctuary, showcasing the effectiveness of acoustic monitoring in wildlife conservation efforts.
|
||||
{% cite muller2021deep %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:85%" .align-center}
|
||||
@@ -7,9 +7,9 @@ header:
|
||||
teaser: assets/figures/12_vision_transformer_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
We apply the vision transformer, a deep machine learning model build around the attention mechanism, on mel-spectrogram representations of raw audio recordings. When adding mel-based data augmentation techniques and sample-weighting, we achieve comparable performance on both (PRS and CCS challenge) tasks of ComParE21, outperforming most single model baselines. We further introduce overlapping vertical patching and evaluate the influence of parameter configurations.
|
||||
This work utilizes the vision transformer model on mel-spectrogram audio data, enhanced by mel-based data augmentation and sample weighting, to achieve notable performance in the ComParE21 challenge, surpassing many single model baselines. The introduction of overlapping vertical patching and the analysis of parameter configurations further refine the approach, demonstrating the model's adaptability and effectiveness in audio processing tasks.
|
||||
{% cite illium2021visual %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
|
||||
@@ -7,7 +7,7 @@ header:
|
||||
teaser: assets/figures/13_sr_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:80%" .align-center}
|
||||
|
||||
Self-replicating neural networks can be trained to output a representation of themselves, making them navigate towards non-trivial fixpoints in their weight space. We explore the problem of adding a secondary functionality to the primary task of replication. We find a successful solution in training the networks with separate input/output vectors for one network trained in both tasks so that the additional task does not hinder (and even stabilizes) the self-replication task. Furthermore, we observe the interaction of our goal-networks in an artificial chemistry environment. We examine the influence of different action parameters on the population and their effects on the group’s learning capability. Lastly we show the possibility of safely guiding the whole group to goal-fulfilling weight configurations via the inclusion of one specially-developed guiding particle that is able to propagate a secondary task to its peers.
|
||||
This research delves into the innovative concept of self-replicating neural networks capable of performing secondary tasks alongside their primary replication function. By employing separate input/output vectors for dual-task training, the study demonstrates that additional tasks can complement and even stabilize self-replication. The dynamics within an artificial chemistry environment are explored, examining how varying action parameters affect the collective learning capability and how a specially developed 'guiding particle' can influence peers towards achieving goal-oriented behaviors, illustrating a method for steering network populations towards desired outcomes.
|
||||
{% cite gabor2021goals %}
|
||||
|
||||
@@ -7,7 +7,8 @@ header:
|
||||
teaser: assets/figures/14_ad_rl_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
Identifying datapoints that substantially differ from normality is the task of anomaly detection (AD). While AD has gained widespread attention in rich data domains such as images, videos, audio and text, it has has been studied less frequently in the context of reinforcement learning (RL). This is due to the additional layer of complexity that RL introduces through sequential decision making. Developing suitable anomaly detectors for RL is of particular importance in safety-critical scenarios where acting on anomalous data could result in hazardous situations. In this work, we address the question of what AD means in the context of RL. We found that current research trains and evaluates on overly simplistic and unrealistic scenarios which reduce to classic pattern recognition tasks. We link AD in RL to various fields in RL such as lifelong RL and generalization. We discuss their similarities, differences, and how the fields can benefit from each other. Moreover, we identify non-stationarity to be one of the key drivers for future research on AD in RL and make a first step towards a more formal treatment of the problem by framing it in terms of the recently introduced block contextual Markov decision process. Finally, we define a list of practical desiderata for future problems.
|
||||
This work investigates anomaly detection (AD) within reinforcement learning (RL), highlighting its importance in safety-critical applications due to the complexity of sequential decision-making in RL. The study criticizes the simplicity of current AD research scenarios in RL, connecting AD to lifelong RL and generalization, discussing their interrelations and potential mutual benefits. It identifies non-stationarity as a crucial area for future AD research in RL, proposing a formal approach through the block contextual Markov decision process and outlining practical requirements for future studies.
|
||||
{% cite muller2022towards %}
|
||||
|
||||
{:style="display:block; width:50%" .align-center}
|
||||
|
||||
@@ -7,9 +7,9 @@ header:
|
||||
teaser: assets/figures/15_sr_journal_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:65%" .align-center}
|
||||
|
||||
A key element of biological structures is self-replication. Neural networks are the prime structure used for the emergent construction of complex behavior in computers. We analyze how various network types lend themselves to self-replication. Backpropagation turns out to be the natural way to navigate the space of network weights and allows non-trivial self-replicators to arise naturally. We perform an in-depth analysis to show the self-replicators’ robustness to noise. We then introduce artificial chemistry environments consisting of several neural networks and examine their emergent behavior. In extension to this works previous version (Gabor et al., 2019), we provide an extensive analysis of the occurrence of fixpoint weight configurations within the weight space and an approximation of their respective attractor basins.
|
||||
This study extends previous work on self-replicating neural networks, focusing on backpropagation as a mechanism for facilitating non-trivial self-replication. It delves into the robustness of these self-replicators against noise and introduces artificial chemistry environments to observe emergent behaviors. Additionally, it provides a detailed analysis of fixpoint weight configurations and their attractor basins, enhancing the understanding of self-replication dynamics within neural networks.
|
||||
{% cite gabor2022self %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:65%" .align-center}
|
||||
|
||||
@@ -7,9 +7,11 @@ header:
|
||||
teaser: assets/figures/16_on_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:65%" .align-center}
|
||||
|
||||
A key element of biological structures is self-replication. Neural networks are the prime structure used for the emergent construction of complex behavior in computers. We analyze how various network types lend themselves to self-replication. Backpropagation turns out to be the natural way to navigate the space of network weights and allows non-trivial self-replicators to arise naturally. We perform an in-depth analysis to show the self-replicators’ robustness to noise. We then introduce artificial chemistry environments consisting of several neural networks and examine their emergent behavior. In extension to this works previous version (Gabor et al., 2019), we provide an extensive analysis of the occurrence of fixpoint weight configurations within the weight space and an approximation of their respective attractor basins.
|
||||
{% cite illium2022constructing %}
|
||||
This work delves into the concept of self-replicating neural networks, focusing on how backpropagation facilitates the emergence of complex, self-replicating behaviors.
|
||||
{:style="display:block; width:45%" .align-right}
|
||||
By evaluating different network types, the study highlights the natural emergence of robust self-replicators and explores their behavior in artificial chemistry environments.
|
||||
A significant extension over a previous version, this research offers a deep analysis of fixpoint weight configurations and their attractor basins, advancing the understanding of neural network self-replication.
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
For more detailed insights, refer to {% cite illium2022constructing %}.
|
||||
@@ -7,11 +7,10 @@ header:
|
||||
teaser: assets/figures/17_vp_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:85%" .align-center}
|
||||
|
||||
Overfitting is a problem in Convolutional Neural Networks (CNN) that causes poor generalization of models on unseen data. To remediate this problem, many new and diverse data augmentation methods (DA) have been proposed to supplement or generate more training data, and thereby increase its quality. In this work, we propose a new data augmentation algorithm: VoronoiPatches (VP). We primarily utilize non-linear recombination of information within an image, fragmenting and occluding small information patches. Unlike other DA methods, VP uses small convex polygon-shaped patches in a random layout to transport information around within an image. Sudden transitions created between patches and the original image can, optionally, be smoothed. In our experiments, VP outperformed current DA methods regarding model variance and overfitting tendencies. We demonstrate data augmentation utilizing non-linear re-combination of information within images, and non-orthogonal shapes and structures improves CNN model robustness on unseen data.
|
||||
{% cite illium2022voronoipatches %}
|
||||
This study introduces VoronoiPatches (VP), a novel data augmentation algorithm that enhances Convolutional Neural Networks' performance by using non-linear recombination of image information. VP distinguishes itself by utilizing small, convex polygon-shaped patches in random layouts to redistribute information within an image, potentially smoothing transitions between patches and the original image. This method has shown to outperform existing data augmentation techniques in reducing model variance and overfitting, thus improving the robustness of CNN models on unseen data. {% cite illium2022voronoipatches %}
|
||||
|
||||
:trophy: This paper won the conference's [Best Poster Award](https://icaart.scitevents.org/PreviousAwards.aspx?y=2024#2023), which is a special honor. :trophy:
|
||||
:trophy: Our work was awarded the [Best Poster Award](https://icaart.scitevents.org/PreviousAwards.aspx?y=2024#2023) at ICAART 2023 :trophy:
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:90%" .align-center}
|
||||
|
||||
@@ -7,10 +7,9 @@ header:
|
||||
teaser: assets/figures/18_surprised_soup_teaser.jpg
|
||||
---
|
||||
|
||||
{: .align-right style="padding:2em; width:20em"}
|
||||
|
||||
A recent branch of research in artificial life has constructed artificial chemistry systems whose particles are dynamic neural networks. These particles can be applied to each other and show a tendency towards self-replication of their weight values. We define new interactions for said particles that allow them to recognize one another and learn predictors for each other’s behavior. For instance, each particle minimizes its surprise when observing another particle’s behavior. Given a special catalyst particle to exert evolutionary selection pressure on the soup of particles, these ‘social’ interactions are sufficient to produce emergent behavior similar to the stability pattern previously only achieved via explicit self-replication training.
|
||||
{:style="display:block; width:40%" .align-right}
|
||||
This research explores artificial chemistry systems with neural network particles that exhibit self-replication. Introducing interactions that enable these particles to recognize and predict each other's behavior, the study observes emergent behaviors akin to stability patterns previously seen in explicit self-replication training. A unique catalyst particle introduces evolutionary pressure, demonstrating how 'social' interactions among particles can lead to complex, emergent outcomes.
|
||||
{% cite zorn23surprise %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto;"}
|
||||
{:style="display:block; width:90%" .align-center}
|
||||
|
||||
|
||||
@@ -7,11 +7,10 @@ header:
|
||||
teaser: assets/figures/19_binary_primates_teaser.jpg
|
||||
---
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
|
||||
In the field of wildlife observation and conservation, approaches involving machine learning on audio recordings are becoming increasingly popular. Unfortunately, available datasets from this field of research are often not optimal learning material; Samples can be weakly labeled, of different lengths or come with a poor signal-to-noise ratio. In this work, we introduce a generalized approach that first relabels subsegments of MEL spectrogram representations, to achieve higher performances on the actual multi-class classification tasks. For both the binary pre-sorting and the classification, we make use of convolutional neural networks (CNN) and various data-augmentation techniques. We showcase the results of this approach on the challenging ComparE 2021 dataset, with the task of classifying between different primate species sounds, and report significantly higher Accuracy and UAR scores in contrast to comparatively equipped model baselines.
|
||||
{:style="display:block; width:40%" .align-right}
|
||||
This study advances machine learning applications in wildlife observation by introducing a sophisticated approach to audio classification. By meticulously relabeling subsegments of MEL spectrograms, it significantly refines the process of multi-class classification, crucial for identifying various primate species from audio recordings. Employing convolutional neural networks alongside innovative data augmentation techniques, the methodology showcases remarkable enhancements in classification performance. When applied to the demanding ComparE 2021 dataset, this approach not only achieved substantially higher accuracy and UAR scores over existing baselines but also marked a significant stride in the field of bioacoustics research, demonstrating the potential of machine learning to overcome challenges presented by datasets with weak labeling, varying lengths, and poor signal-to-noise ratios.
|
||||
{% cite koelle23primate %}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:70%" .align-center}
|
||||
|
||||
{:style="display:block; margin-left:auto; margin-right:auto"}
|
||||
{:style="display:block; width:70%" .align-center}
|
||||
|
||||
Reference in New Issue
Block a user