David C. Poole

Matthew Dirks, David Poole

Neural networks are configured by choosing an architecture and hyperparameter values; doing so often involves expert intuition and hand-tuning to find a configuration that extrapolates well without overfitting. This paper considers automatic methods for configuring a neural network that extrapolates in time for the domain of visible and near-infrared (VNIR) spectroscopy. In particular, we study the effect of (a) selecting samples for validating configurations and (b) using ensembles. Most of the time, models are built of the past to predict the future. To encourage the neural network model to extrapolate, we consider validating model configurations on samples that are shifted in time similar to the test set. We experiment with three validation set choices: (1) a random sample of 1/3 of non-test data (the technique used in previous work), (2) using the latest 1/3 (sorted by time), and (3) using a semantically meaningful subset of the data. Hyperparameter optimization relies on the validation set to estimate test-set error, but neural network variance obfuscates the true error value. Ensemble averaging - computing the average across many neural networks - can reduce the variance of prediction errors. To test these methods, we do a comprehensive study of a held-out 2018 harvest season of mango fruit given VNIR spectra from 3 prior years. We find that ensembling improves the state-of-the-art model's variance and accuracy. Furthermore, hyperparameter optimization experiments - with and without ensemble averaging and with each validation set choice - show that when ensembling is combined with using the latest 1/3 of samples as the validation set, a neural network configuration is found automatically that is on par with the state-of-the-art.

Matthew Dirks, David Poole

We consider energy-dispersive X-ray Fluorescence (EDXRF) applications where the fundamental parameters method is impractical such as when instrument parameters are unavailable. For example, on a mining shovel or conveyor belt, rocks are constantly moving (leading to varying angles of incidence and distances) and there may be other factors not accounted for (like dust). Neural networks do not require instrument and fundamental parameters but training neural networks requires XRF spectra labelled with elemental composition, which is often limited because of its expense. We develop a neural network model that learns from limited labelled data and also benefits from domain knowledge by learning to invert a forward model. The forward model uses transition energies and probabilities of all elements and parameterized distributions to approximate other fundamental and instrument parameters. We evaluate the model and baseline models on a rock dataset from a lithium mineral exploration project. Our model works particularly well for some low-Z elements (Li, Mg, Al, and K) as well as some high-Z elements (Sn and Pb) despite these elements being outside the suitable range for common spectrometers to directly measure, likely owing to the ability of neural networks to learn correlations and non-linear relationships.

David Poole

Artificial intelligence seems to be taking over the world with systems that model pixels, words, and phonemes. The world is arguably made up, not of pixels, words, and phonemes but of entities (objects, things, including events) with properties and relations among them. Surely we should model these, not the perception or description of them. You might suspect that concentrating on modeling words and pixels is because all of the (valuable) data in the world is in terms of text and images. If you look into almost any company you will find their most valuable data is in spreadsheets, databases and other relational formats. These are not the form that are studied in introductory machine learning, but are full of product numbers, student numbers, transaction numbers and other identifiers that can't be interpreted naively as numbers. The field that studies this sort of data has various names including relational learning, statistical relational AI, and many others. This paper explains why relational learning is not taking over the world -- except in a few cases with restricted relations -- and what needs to be done to bring it to it's rightful prominence.

Bahare Fatemi, Siamak Ravanbakhsh, David Poole

Knowledge graphs are used to represent relational information in terms of triples. To enable learning about domains, embedding models, such as tensor factorization models, can be used to make predictions of new triples. Often there is background taxonomic information (in terms of subclasses and subproperties) that should also be taken into account. We show that existing fully expressive (a.k.a. universal) models cannot provably respect subclass and subproperty information. We show that minimal modifications to an existing knowledge graph completion method enables injection of taxonomic information. Moreover, we prove that our model is fully expressive, assuming a lower-bound on the size of the embeddings. Experimental results on public knowledge graphs show that despite its simplicity our approach is surprisingly effective.

Nandini Ramanan, Gautam Kunapuli, Tushar Khot et al.

We consider the problem of learning Relational Logistic Regression (RLR). Unlike standard logistic regression, the features of RLRs are first-order formulae with associated weight vectors instead of scalar weights. We turn the problem of learning RLR to learning these vector-weighted formulae and develop a learning algorithm based on the recently successful functional-gradient boosting methods for probabilistic logic models. We derive the functional gradients and show how weights can be learned simultaneously in an efficient manner. Our empirical evaluation on standard and novel data sets demonstrates the superiority of our approach over other methods for learning RLR.