At WaveOne, I have been working on a variety of fascinating problems from across the spectrum of low-level (optical flow estimation, learned entropy coding, etc.) to high (semantic segmentation for intelligent bitrate allocation).
Previously, I was a member of the Harvard Intelligent Probabilistic Systems group, and worked with advisor Ryan Adams on a spectrum of fundamental problems in machine learning. I was an active member of the machine learning communities at MIT and Harvard. I was a frequent contributor to the Building Intelligent Probabilistic Systems blog, and participated in the Harvard Machine Learning Tea.
Learned Video Compression
Oren Rippel, Sanjay Nair, Carissa Lew, Steve Branson, Alexander G. Anderson, Lubomir Bourdev
We present a new algorithm for video coding, learned end-to-end for the low-latency mode. In this setting, our approach outperforms all existing video codecs. To our knowledge, this is the first ML-based method to do so. Our approach does not suffer from blocking artifacts and pixelation, and thus produces videos that are more visually pleasing.
International Conference on Computer Vision (ICCV), 2019. PDF | BibTeX
Real-Time Adaptive Image Compression
Oren Rippel, Lubomir Bourdev
We present a machine learning-based approach to lossy image compression which outperforms all existing codecs, while running in real-time. We supplement our approach with adversarial training specialized towards use in a compression setting: this enables us to produce visually pleasing reconstructions for very low bitrates.
International Conference on Machine Learning (ICML), 2017. PDF | BibTeX
Metric Learning with Adaptive Density Discrimination
Oren Rippel, Manohar Paluri, Piotr Dollar and Lubomir Bourdev
We propose an approach to address a number of subtle yet important issues which have stymied earlier distance metric learning algorithms. It maintains an explicit model of the distributions of the different classes in representation space. It employs this knowledge to adaptively assess similarity, and pursue local discrimination by penalizing class distribution overlap. This idea allows us to surpass existing approaches by a significant margin on a number of tasks such as classification, training efficiency and representation saliency.
International Conference on Learning Representations (ICLR), 2016. PDF | BibTeX
Spectral Representations for Convolutional Neural Networks
Oren Rippel, Jasper Snoek, and Ryan P. Adams
We argue that, beyond its advantages for efficient convolution computation, the spectral domain also provides a powerful representation in which to model and train convolutional neural networks. We propose spectral pooling, which preserves considerably more information per parameter than other pooling strategies and enables flexibility in the choice of pooling output dimensionality. We also demonstrate the effectiveness of complex-coefficient spectral parameterization of convolutional filters, which leads to significantly faster convergence during training.
Neural Information Processing Systems (NIPS), 2015. PDF | BibTeX
Scalable Bayesian Optimization Using Deep Neural Networks
Jasper Snoek, Oren Rippel, Kevin Swersky, Ryan Kiros, Nadathur Satish, Narayanan Sundaram, Md. Mostofa Ali Patwary, Prabhat and Ryan P. Adams
We explore the use of neural networks as an alternative to Gaussian processes to model distributions over functions. While this approach performs competitively with state-of-the-art GP-based approaches, it scales linearly with the number of data rather than cubically. This allows us to achieve a previously intractable degree of parallelism, which we use to rapidly search over large spaces of models.
International Conference on Machine Learning (ICML), 2015. PDF | BibTeX
MICMat Kernel Library for Xeon Phi
Oren Rippel, Nadathur Satish, Narayanan Sundaram, Md. Mostofa Ali Patwary, Prabhat and Ryan P. Adams
I developed and maintained the MICMat (MIC Matrix) kernel library, which enables interfacing with Intel's Xeon Phi Coprocessor directly from pure Python. It presents an extensive library of primitives, optimized for high performance computation, while allowing very convenient development.
Maintained 2014 - 2016. GitHub repository
Learning Ordered Representations with Nested Dropout
Oren Rippel, Michael Gelbart and Ryan P. Adams
We study ordered representations of data in which different dimensions have different degrees of importance. To learn these we introduce Nested Dropout. We rigorously show that the application of nested dropout enforces identifiability of the units. We use the ordering property to construct data structures that permit retrieval in time logarithmic in the database size and independent of the dimensionality of the representation. We also show that ordered representations are a promising way to learn adaptive compression for efficient online data reconstruction.
International Conference on Machine Learning (ICML), 2014. PDF | video lecture | BibTeX
Avoiding Pathologies in Very Deep Networks
David Duvenaud, Oren Rippel, Ryan P. Adams and Zoubin Ghahramani
To help suggest better deep neural network architectures, we analyze the related problem of constructing useful priors on compositions of functions. We study deep Gaussian process, a type of infinitely-wide, deep neural net. We also examine deep covariance functions, obtained by composing infinitely many feature transforms. Finally, we characterize the model class you get if you do dropout on Gaussian processes.
Artificial Intelligence and Statistics (AISTATS), 2014. PDF | video of 50-layer warping | BibTeX
High-Dimensional Probability Estimation with Deep Density Models
Oren Rippel and Ryan P. Adams