From the article: Abstract: An overview of neural network architectures is presented. Some of these architectures have been created in recent years, whereas others originate from many decades ago. Apart from providing a practical tool for comparing deep learning models, the Neural Network Zoo also uncovers a taxonomy of network architectures, their chronology, and traces back lineages and inspirations for these neural information processing systems.
We present a novel architecture for an AI system that allows a priori knowledge to combine with deep learning. In traditional neural networks, all available data is pooled at the input layer. Our alternative neural network is constructed so that partial representations (invariants) are learned in the intermediate layers, which can then be combined with a priori knowledge or with other predictive analyses of the same data. This leads to smaller training datasets due to more efficient learning. In addition, because this architecture allows inclusion of a priori knowledge and interpretable predictive models, the interpretability of the entire system increases while the data can still be used in a black box neural network. Our system makes use of networks of neurons rather than single neurons to enable the representation of approximations (invariants) of the output.
LINK
The recent success of Machine Learning encouraged research using artificial neural networks (NNs) in computer graphics. A good example is the bidirectional texture function (BTF), a data-driven representation of surface materials that can encapsulate complex behaviors that would otherwise be too expensive to calculate for real-time applications, such as self-shadowing and interreflections. We propose two changes to the state-of-the-art using neural networks for BTFs, specifically NeuMIP. These changes, suggested by recent work in neural scene representation and rendering, aim to improve baseline quality, memory footprint, and performance. We conduct an ablation study to evaluate the impact of each change. We test both synthetic and real data, and provide a working implementation within the Mitsuba 2 rendering framework. Our results show that our method outperforms the baseline in all these metrics and that neural BTF is part of the broader field of neural scene representation. Project website: https://traverse-research.github.io/NeuBTF/.
The project aim is to improve collusion resistance of real-world content delivery systems. The research will address the following topics: • Dynamic tracing. Improve the Laarhoven et al. dynamic tracing constructions [1,2] [A11,A19]. Modify the tally based decoder [A1,A3] to make use of dynamic side information. • Defense against multi-channel attacks. Colluders can easily spread the usage of their content access keys over multiple channels, thus making tracing more difficult. These attack scenarios have hardly been studied. Our aim is to reach the same level of understanding as in the single-channel case, i.e. to know the location of the saddlepoint and to derive good accusation scores. Preferably we want to tackle multi-channel dynamic tracing. • Watermarking layer. The watermarking layer (how to embed secret information into content) and the coding layer (what symbols to embed) are mostly treated independently. By using soft decoding techniques and exploiting the “nuts and bolts” of the embedding technique as an extra engineering degree of freedom, one should be able to improve collusion resistance. • Machine Learning. Finding a score function against unknown attacks is difficult. For non-binary decisions there exists no optimal procedure like Neyman-Pearson scoring. We want to investigate if machine learning can yield a reliable way to classify users as attacker or innocent. • Attacker cost/benefit analysis. For the various use cases (static versus dynamic, single-channel versus multi-channel) we will devise economic models and use these to determine the range of operational parameters where the attackers have a financial benefit. For the first three topics we have a fairly accurate idea how they can be achieved, based on work done in the CREST project, which was headed by the main applicant. Neural Networks (NNs) have enjoyed great success in recognizing patterns, particularly Convolutional NNs in image recognition. Recurrent NNs ("LSTM networks") are successfully applied in translation tasks. We plan to combine these two approaches, inspired by traditional score functions, to study whether they can lead to improved tracing. An often-overlooked reality is that large-scale piracy runs as a for-profit business. Thus countermeasures need not be perfect, as long as they increase the attack cost enough to make piracy unattractive. In the field of collusion resistance, this cost analysis has never been performed yet; even a simple model will be valuable to understand which countermeasures are effective.
Human kind has a major impact on the state of life on Earth, mainly caused by habitat destruction, fragmentation and pollution related to agricultural land use and industrialization. Biodiversity is dominated by insects (~50%). Insects are vital for ecosystems through ecosystem engineering and controlling properties, such as soil formation and nutrient cycling, pollination, and in food webs as prey or controlling predator or parasite. Reducing insect diversity reduces resilience of ecosystems and increases risks of non-performance in soil fertility, pollination and pest suppression. Insects are under threat. Worldwide 41 % of insect species are in decline, 33% species threatened with extinction, and a co-occurring insect biomass loss of 2.5% per year. In Germany, insect biomass in natural areas surrounded by agriculture was reduced by 76% in 27 years. Nature inclusive agriculture and agri-environmental schemes aim to mitigate these kinds of effects. Protection measures need success indicators. Insects are excellent for biodiversity assessments, even with small landscape adaptations. Measuring insect biodiversity however is not easy. We aim to use new automated recognition techniques by machine learning with neural networks, to produce algorithms for fast and insightful insect diversity indexes. Biodiversity can be measured by indicative species (groups). We use three groups: 1) Carabid beetles (are top predators); 2) Moths (relation with host plants); 3) Flying insects (multiple functions in ecosystems, e.g. parasitism). The project wants to design user-friendly farmer/citizen science biodiversity measurements with machine learning, and use these in comparative research in 3 real life cases as proof of concept: 1) effects of agriculture on insects in hedgerows, 2) effects of different commercial crop production systems on insects, 3) effects of flower richness in crops and grassland on insects, all measured with natural reference situations
Predictive maintenance, using data of thousands of sensors already available, is key for optimizing the maintenance schedule and further prevention of unexpected failures in industry. Current maintenance concepts (in the maritime industry) are based on a fixed maintenance interval for each piece of equipment with enough safety margin to minimize incidents. This means that maintenance is most of the time carried out too early and sometimes too late. This is in particular true for maintenance on maritime equipment, where onshore maintenance is strongly preferred over offshore maintenance and needs to be aligned with the vessel’s operations schedule. However, state-of-the-art predictive maintenance methods rely on black-box machine learning techniques such as deep neural networks that are difficult to interpret and are difficult to accept and work with for the maintenance engineers. The XAIPre project (pronounce “Xyper”) aims at developing Explainable Predictive Maintenance (XPdM) algorithms that do not only provide the engineers with a prediction but in addition, with 1) a risk analysis on the components when delaying the maintenance, and 2) what the primary indicators are that the algorithms used to create inference. To use predictive maintenance effectively in Maritime operations, the predictive models and the optimization of the maintenance schedule using these models, need to be aware of the past and planned vessel activities, since different activities affect the lifetime of the machines differently. For example, the degradation of a hydraulic pump inside a crane depends on the type of operations the crane performs. Thus, the models do not only need to be explainable but they also need to be aware of the context which is in this case the vessel and machinery activity. Using sensor data processing and edge-computing technologies that will be developed and applied by the Hanze UAS in Groningen, context information is extracted from the raw sensor data. The XAIPre project combines these Explainable Context Aware Machine Learning models with state-of-the-art optimizers, that we already developed in the NWO CIMPLO project at LIACS, in order to develop optimal maintenance schedules for machine components. The optimizers will be adapted to fit within XAIPre. The resulting XAIPre prototype offers significant competitive advantages for companies such as Heerema, by increasing the longevity of machine components, increasing worker safety, and decreasing maintenance costs. XAIPre will focus on the predictive maintenance of thrusters, which is a key sub-system with regards to maintenance as it is a core part of the vessels station keeping capabilities. Periodic maintenance is currently required in fixed intervals of 5 years. XPdM can provide a solid base to deviate from the Periodic Maintenance prescriptions to reduce maintenance costs while maintaining quality. Scaling up to include additional components and systems after XAIPre will be relatively straightforward due to the accumulated knowledge of the predictive maintenance process and the delivered methods. Although the XAIPre system will be evaluated on the use-cases of Heerema, many components of the system can be utilized across industries to save maintenance costs, maximize worker safety and optimize sustainability.
