ABSTRACT
Recent work has shown potential in using Mixed Integer Programming (MIP) solvers to optimize certain aspects of neural networks (NNs). However the intriguing approach of training NNs with MIP solvers is under-explored. State-of-the-art-methods to train NNs are typically gradient-based and require significant data, computation on GPUs, and extensive hyper-parameter tuning. In contrast, training with MIP solvers does not require GPUs or heavy hyper-parameter tuning, but currently cannot handle anything but small amounts of data. This article builds on recent advances that train binarized NNs using MIP solvers. We go beyond current work by formulating new MIP models which improve training efficiency and which can train the important class of integer-valued neural networks (INNs). We provide two novel methods to further the potential significance of using MIP to train NNs. The first method optimizes the number of neurons in the NN while training. This reduces the need for deciding on network architecture before training. The second method addresses the amount of training data which MIP can feasibly handle: we provide a batch training method that dramatically increases the amount of data that MIP solvers can use to train. We thus provide a promising step towards using much more data than before when training NNs using MIP models. Experimental results on two real-world data-limited datasets demonstrate that our approach strongly outperforms the previous state of the art in training NN with MIP, in terms of accuracy, training time and amount of data. Our methodology is proficient at training NNs when minimal training data is available, and at training with minimal memory requirements-which is potentially valuable for deploying to low-memory devices.
ABSTRACT
BACKGROUND: Using knowledge representation for biomedical projects is now commonplace. In previous work, we represented the knowledge found in a college-level biology textbook in a fashion useful for answering questions. We showed that embedding the knowledge representation and question-answering abilities in an electronic textbook helped to engage student interest and improve learning. A natural question that arises from this success, and this paper's primary focus, is whether a similar approach is applicable across a range of life science textbooks. To answer that question, we considered four different textbooks, ranging from a below-introductory college biology text to an advanced, graduate-level neuroscience textbook. For these textbooks, we investigated the following questions: (1) To what extent is knowledge shared between the different textbooks? (2) To what extent can the same upper ontology be used to represent the knowledge found in different textbooks? (3) To what extent can the questions of interest for a range of textbooks be answered by using the same reasoning mechanisms? RESULTS: Our existing modeling and reasoning methods apply especially well both to a textbook that is comparable in level to the text studied in our previous work (i.e., an introductory-level text) and to a textbook at a lower level, suggesting potential for a high degree of portability. Even for the overlapping knowledge found across the textbooks, the level of detail covered in each textbook was different, which requires that the representations must be customized for each textbook. We also found that for advanced textbooks, representing models and scientific reasoning processes was particularly important. CONCLUSIONS: With some additional work, our representation methodology would be applicable to a range of textbooks. The requirements for knowledge representation are common across textbooks, suggesting that a shared semantic infrastructure for the life sciences is feasible. Because our representation overlaps heavily with those already being used for biomedical ontologies, this work suggests a natural pathway to include such representations as part of the life sciences curriculum at different grade levels.
