Recursive Octree Network for Efficient 3D Processing
Juncheng Liu (University of Otago)
We introduce a deep recursive octree network for general-purpose 3D voxel data processing. Our network compresses a voxel grid of any size down to a very small latent space in an autoencoder- like network. We show results for compressing 32^3, 64^3 and 128^3 grids down to just 80 floats in the latent space. We demonstrate the effectiveness and efficiency of our proposed method on several publicly available datasets with four experiments: 3D shape classification, 3D shape reconstruction, shape generation and semantic segmentation. Experimental results show that our algorithm maintains accuracy while consuming less memory with shorter training times compared to existing methods, especially in 3D reconstruction tasks.
Semi-Supervised Learning for Delayed Partially Labelled Data Streams
Heitor Murilo Gomes (Victoria University of Wellington)
Unlabelled data appear in many domains and are particularly relevant to streaming applications, where even though data is abundant, labelled data is rare. To address the learning problems associated with such data, one can ignore the unlabelled data and focus only on the labelled data (supervised learning); use the labelled data and attempt to leverage the unlabelled data (semi-supervised learning); or assume some labels will be available on request (active learning).
The first approach is the simplest, yet the amount of labelled data available will limit the predictive performance. The second relies on finding and exploiting the underlying characteristics of the data distribution. The third depends on an external agent to provide the required labels in a timely fashion. In this talk, I will pay special attention to methods that leverage unlabelled data in a semi-supervised setting, and discuss the delayed labelling issue, which impacts both fully supervised and semi-supervised methods. On top of that, I will introduce a unified problem setting, discuss existing methods, and explain the differences between related problem settings. Finally, I will review the current benchmarking practices and propose adaptations to enhance them.
SOKNL: A novel way of integrating K-nearest neighbours with adaptive random forest regression for data streams
Bernhard Pfahringer (University of Waikato)
Self-Optimising K-Nearest Leaves (SOKNL) is a novel forest-based algorithm for regression problems in data streams. It extends Adaptive Random Forest Regression, a state-of-the-art online regression algorithm, in the following way: each leaf keeps the centroid of all its data points as a compressed representation of these data points. For prediction, each tree returns the centroid of the leaf an example is sorted to. The final ensemble prediction is then the result of a K-nearest neighbour computation over this specific set of centroids. The value of K is chosen adaptively during stream processing by monitoring the performance of all possible values for k, ranking from 1 up to the size of the ensemble. An empirical evaluation shows very promising results, for only small increases in runtime and memory consumption over the original Adaptive Random Forest Regressor.
Supporting Scaled Science with Strong AI
Michael Witbrock (Waipapa Taumata Rau Natural Artificial and Organisational Intelligence Institute Strong AI Lab)
The vast volume of the scientific literature, and the compex incentives around its production and maintenance, present significant potential barriers to scientific and technological progress. Adding to this complexity is the need to communicate about scientific processes and outcomes with the many cultural, economic and social groups comprised within human civilisation. Many of the difficulties presented seem to be driven, at least in part, by the need to run many scientific processes on organisational platforms that are rate and capacity limited by human cognitive, communicative, and social limitations. These concerns are even more present in Aotearoa, where our capabity to mitigate limitations with scale and proximity are reduced. A stong theme of NAOI is to understand diverse possibilities for intelligence, and where possible mitigate limitations of human individual and organisational intelligences in ways that are beneficial for the humans involved, and human civilisation, with the Strong AI Lab focussed on developing AI techniques that can help humans the most. Within that grandiose vision, we aim to "eat our own dogfood", by applying these techniques to improving the processes of science in building large, sceptical, neutral models of the world, and connecting them to technological means. To that end, we're working on question answering, theory formation and interaction, explanation generation, causal reasoning, scientific document understanding, amongst other themes, with a specific intent of enabling computers to help humans in comprehensively understanding large scale scientific knowledge sets. In this talk, I'll discuss that work in a little detail, ask for your collaboration, and wonder how we can use such techniques to build on our effective research and technology in Aotearoa to support human civilisational and ecological progress.
Unsupervised feature selection for systematic time-series engineering
Andreas Kempa-Liehr (The University of Auckland)
Clustering of time-series or streaming data can either operate on the raw data (naïve feature engineering) or on problem-specific time-series features, which utilise domain knowledge to characterise and discriminate different clusters of time-series data. Developing and implementing the algorithms for time-series feature extraction is usually tedious unless libraries for automated time-series feature engineering are used. A prominent Python implementation is tsfresh (https://github.com/blue-yonder/tsfresh). This Python module extracts a wide variety of time-series features using algorithms from statistics, time-series analysis, financial time-series analysis, signal processing, and nonlinear dynamics. While the broad range of time-series features guarantees that at least some features are suitable for clustering the time-series samples, most features are likely to be statistically independent with respect to forming informative clusters. In systematic time-series feature engineering, the automated feature extraction is combined with univariate feature selection and control of false discovery rate to reduce the dimensionality of the data set to a representative subset. Because no labels are given for the unsupervised feature selection task, these need to be generated from the internal ordering of the time-series samples. Suitable feature selection targets, which can be generated automatically, are generated from the perspective of time-series forecasting. The forecasting target is a representative statistic of a certain fraction of the time-series sample. The presentation closes with an outlook on applications in streaming data, e.g. anomaly detection.
Autonomous 3D sub-mm surveying for accurate models and precision tool use
Richard Green (University of Canterbury)
Drones (aerial and underwater) and robots use various deep learning and other machine learning optimisation approaches to autonomously acquire petabytes per hectare for agriculture and aquaculture - and then convert these sub-mm point-clouds into accurate models. Applications include drones and robots using these models to enable the use of tools with mm precision.
Ethical Considerations with AI - An holistic view
Dr Karaitiana Taiuru (Taiuru & Associates)
An holistic and high level overview Māori ethical considerations when developing and using AI, concluding with a discussion if an AI claims to be sentient and Māori - Māori lore considerations.
Navigating the Multiverse
Elle Archer (Te Matarau - The Māori Tech Association)
In Kotahitanga, the unification kaupapa, strategically weaving our ecosystems together has never been more vital. AI can play an active role in this. I will share ‘My Multiverse’ along with a couple of things that aid me in navigating and co-constructing converging spaces.