Often when companies get started with data projects (especially when the initiatives come from the C-suite or from the top down), they bite off more than they can chew and end up with less-than-optimal results. But organizations can avoid this frustration by taking a queue from the lean startup thinking popularized by Eric Ries and applying the principles of Minimal Viable Product (MVP) to data projects.

Paul Pudaite points to this post by Stacy McGaugh entitled, “The dwarf galaxy NGC1052-DF2.” Pudaite writes that it’s an interesting comment on consequences of excluding one outlier.

Chinese researchers pursue state-of-the-art lip-reading with massive dataset:…What do I spy with my camera eyes? Lips moving! Now I can figure out what you are saying…Researchers with the Chinese Academic of Sciences and Huazhong University of Science and Technology have created a new dataset and benchmark for “lip-reading in the wild” for Mandarin. Lip-reading gives people a new sensory capability to imbue AI systems with. For instance, lip-reading systems can be used for “aids for hearing-impaired persons, analysis of silent movies, liveness verification in video authentication systems, and so on” the researchers write.  Dataset details: The lipreading dataset contains 745,187 distinct samples from more than 2,000 speakers, grouped into 1,000 classes, where each class corresponds to the syllable of a Mandarin word composed of one or several Chinese characters. “To the best of our knowledge, this database is currently the largest word-level lipreading dataset and the only public large-scale Mandarin lipreading dataset”, the researchers write. The dataset has also been designed to be dverse so the footage in it consists of multiple different people taken from multiple different camera angles, along with perspectives taken from television broadcasts. This diversity makes the benchmark more closely approximate real world situations whereas previous work in this domain has involved stuff taken from a fixed perspective. They build the dataset by annotating Chinese television using a service provided by iFLYTEK, a Chinese speech recognition company.  Baseline results: They train three baselines on this dataset – a fully 2D CNN, a fully 3D CNN (modeled on LipNet, research covered in ImportAI #104 from DeepMind and Google) , and a model that mixes 2D and 3D convolutional layers. All of these approaches perform poorly on the new dataset, despite having obtained performances as high as 90% on other more restricted datasets. The researchers implement their models in PyTorch and train them on servers containing four Titan X GPUs with 12GB of memory. The resulting top-5 accuracy results for the baselines on the new Chinese dataset LRW-1000 are as follows:– LSTM-5: 48.74%– D3D: 59.80%– 3D+2D: 63.50%  Why it matters: Systems for stuff like lipreading are going to have a significant impact on applications ranging from medicine to surveillance. One of the challenges posed by research like this is its inherently ‘dual use’ nature; as the researchers allude to in the introduction of this paper, this work can be used both for healthcare uses as well for surveillance uses (see: “analysis of silent movies”). How society deals with the arrival of these general AI technologies will have a significant impact on the types of societal architectures that will be built and developed throughout the 21st Century. It is also notable to see the emergence of large-scale datasets built by Chinese researchers in Chinese language – perhaps one could measure the relative growth in certain language datasets to model AI interest in the associated countries?  Read more: LRW-1000: A Naturally Distributed Large-Scale Benchmark for Lip Reading in the Wild (Arxiv).

In this work we introduce declarative statistics, a suite of declarative modelling tools for statistical analysis. Statistical constraints represent the key building block of declarative statistics. First, we introduce a range of relevant counting and matrix constraints and associated decompositions, some of which novel, that are instrumental in the design of statistical constraints. Second, we introduce a selection of novel statistical constraints and associated decompositions, which constitute a self-contained toolbox that can be used to tackle a wide range of problems typically encountered by statisticians. Finally, we deploy these statistical constraints to a wide range of application areas drawn from classical statistics and we contrast our framework against established practices. Declarative Statistics

This summer we had five interns participate in our internship program. Each intern was here for 10 weeks and worked closely with a mentor or team. Everyone jumped right in and contributed quickly. We are excited about the progress our interns made and wanted to share it with you here!

Interpretable Convolutional Filter Pruning

So you’ve decided you want to learn data science. Should you share your goals on social media or with an accountability buddy? Or should you work in silence until you’ve learned enough to call yourself a data scientist?

Welcome to week 6 of the Data Science with R Course Series.