# Stanford Compression Workshop 2019

February 15, 2019 • Bechtel Conference Center • 616 Serra Street • Stanford University

### Program

 8:30 am - 9:00 am Registration + Breakfast 9:00 am - 9:15 am Tsachy Weissman, Stanford UniversityOpening Remarks [video] 9:15 am - 9:45 am Anne Aaron, Netflix Tidying up (bits on the internet) with AnneAaron [video] ▸ Abstract:  In a video streaming service, the goal is to deliver the best video quality possible at the lowest bitrates necessary, for any type of content or viewing condition. We continuously work towards this by questioning the encoding industry status quo, conducting targeted applied research, and productizing the results. In this talk we give an overview of our work on perceptual video quality metrics (VMAF), rate control (per-title encoding and dynamic optimization) and codecs and how they impact our production system. Bio:  Anne Aaron is Director of Video Algorithms at Netflix and leads the team responsible for video analysis, processing and encoding in the Netflix cloud-based media pipeline. The team is tasked with generating the best quality video streams for millions of Netflix members worldwide. The team is also actively involved in defining next-generation video through academic research collaboration and standardization work. Prior to Netflix, Anne had technical lead roles at Cisco, working on the software deployed with millions of Flip Video cameras, Dyyno, an early stage startup which developed a real-time peer-to-peer video distribution system, and Modulus Video, a broadcast video encoder company. During her Ph.D. studies at Stanford University, she was a member of the Image, Video and Multimedia Systems Laboratory, led by Prof. Bernd Girod. Her research was one of the pioneering work in the sub-field of Distributed Video Coding. Anne is originally from Manila, Philippines. She holds B.S. degrees in Physics and Computer Engineering from Ateneo de Manila University and M.S. and Ph.D. degrees in Electrical Engineering from Stanford University. Anne was recognized by Forbes as one of America’s Top 50 Women In Tech in 2018. Anne recently binged on Ozark, Bodyguard and Bojack Horseman. 9:45 am - 10:15 am Jim Bankoski and Jan Skoglund, GoogleLow rate speech coding using deep generative models [video] ▸ Abstract:  Traditional parametric coding of speech facilitates low bit rate but provides poor reconstruction quality because of the inadequacy of the model used. In the last few years, machine learning has facilitated the development of speech synthesis systems that are able to produce excellent speech quality by generative neural network models using deep learning. In this talk we describe how such generative models can be used to produce high quality speech from the bit stream of a parametric coder operating at low rates. 10:15 am - 10:30 am Elaina Chai, Stanford UniversityQuantization Error in Neural Networks [video] 10:30 am - 10:45 am Break 10:45 am - 11:15 am Mikel Hernaez, University of Illinois Urbana-Champaign An overview of the ISO-based MPEG-G standard for genomic information representation [video] ▸ Abstract:  The MPEG-G standardization project is the largest coordinated international effort to specify a compressed data format that enables large-scale genomic data processing, transport, and sharing. It is the first ISO/IEC standard that addresses the problems and limitations of current genomic data formats towards a truly efficient and economical handling of genomic information. It provides the means to implement leading-edge compression technology achieving more than 10x improvement over the BAM format. The standard also provides a set of currently-needed functionalities, such as selective access, application programming interfaces to the compressed data, support of data protection mechanisms, and support for streaming applications. Furthermore, ISO/IEC is also engaged in supporting the maintenance of the standard to guarantee the long-support of applications using MPEG-G. Finally, interoperability and integration with existing genomic information processing pipelines is enabled by supporting conversion from/to the FASTQ/SAM/BAM file formats. In this talk we will describe the MPEG-G standard in more detail, as well as the main advantages and functionalities offered by it. Bio:  Mikel Hernaez is the current Director for Computational Genomics at the Carl R. Woese Institute for Genomic Biology at the University of Illinois at Urbana-Champaign (UIUC). From 2013-2016 we was a postdoctoral Scholar at Electrical Engineering Department at Stanford University in the group of Prof. Tsachy Weissman. Prior to Stanford, he was the Director of Research at Enigmedia, a start-up developing crypto solutions based in Spain. In 2012 he graduated from his PhD in Electrical Engineering from the University of Navarra under the supervision of Prof. Pedro Crespo. During his time at Stanford, he was the organizer of the Stanford Compression Forum, and an active member of the Stanford Data Science Initiative. He was also among the few selected individuals that participated in the Pear Garage, an entrepreneurial mentoring program from Pear VC in Palo Alto, CA. Currently, he is one of the leaders of the upcoming International Standard for Genomic Information Representation, MPEG-G, which will change how genomic data is stored, processed and analyzed. His research interests are Data Compression, Computational Biology, Machine Learning and Information Theory and Coding. His Postdoc was partially funded by the Stanford Data Science Initiative, and during his masters he was awarded the Telefonica Scholarship. 11:15 am - 11:45 am Hanlee Ji, Stanford University Advances in High Density Molecular Data Storage [video] 11:45 am - 12:15 pm Victoria Popic, SambaNova SystemsLeveraging domain knowledge for compressive, lightweight, and privacy-preserving genome analysis [video] 12:15 pm - 1:30 pm Lunch 1:30 pm - 1:45 pm Sadjad Fouladi, Stanford University Salsify: Low-Latency Network Video Through Tighter Integration Between a Video Codec and a Transport Protocol [video] 1:45 pm - 2:00 pm Leighton Barnes, Stanford University Learning Distributions from their Samples under Communication Constraints [video] ▸ Abstract:  We consider the problem of learning high-dimensional, nonparametric and structured (e.g. Gaussian) distributions in distributed networks, where each node in the network observes an independent sample from the underlying distribution and can use $k$ bits to communicate its sample to a central processor. We consider three different models for communication. Under the independent model, each node communicates its sample to a central processor by independently encoding it into $k$ bits. Under the more general sequential or blackboard communication models, nodes can share information interactively but each node is restricted to write at most $k$ bits on the final transcript. We characterize the impact of the communication constraint $k$ on the minimax risk of estimating the underlying distribution under $\ell^2$ loss. We develop minimax lower bounds that apply in a unified way to many common statistical models and reveal that the impact of the communication constraint can be qualitatively different depending on the tail behavior of the score function associated with each model. A key ingredient in our proof is a geometric characterization of Fisher information from quantized samples. This is joint work with Ayfer Ozgur and Yanjun Han. Bio:  Leighton Pate Barnes received a B.S. in Mathematics '13, B.S. in Electrical Science and Engineering '13, and M.Eng. in Electrical Engineering and Computer Science '15, all from the Massachusetts Institute of Technology. While there, he received the Harold L. Hazen Award for excellence in teaching. He is currently a Ph.D. candidate in the Department of Electrical Engineering at Stanford University, where he studies geometric extremal problems applied to information theory, communication, and estimation. 2:00 pm - 2:15 pm Kristy Choi, Stanford University Neural joint-source channel coding [video] ▸ Abstract:  For reliable transmission across a noisy communication channel, classical results from information theory show that it is asymptotically optimal to separate out the source and channel coding processes. However, this decomposition can fall short in the finite bit-length regime, as it requires non-trivial tuning of hand-crafted codes and assumes infinite computational power for decoding. In this work, we propose to jointly learn the encoding and decoding processes using a new discrete variational autoencoder model. By adding noise into the latent codes to simulate the channel during training, we learn to both compress and error-correct given a fixed bit-length and computational budget. We obtain codes that are not only competitive against several separation schemes, but also learn useful robust representations of the data for downstream tasks such as classification. Finally, inference amortization yields an extremely fast neural decoder, almost an order of magnitude faster compared to standard decoding methods based on iterative belief propagation. 2:15 pm - 2:30 pm Shubham Chandak, Stanford UniversitySPRING: a next-generation compressor for FASTQ data [video] ▸ Abstract: Motivation: High-Throughput Sequencing (HTS) technologies produce huge amounts of data in the form of short genomic reads, associated quality values, and read identifiers. Because of the significant structure present in these FASTQ datasets, general-purpose compressors are unable to completely exploit much of the inherent redundancy. Although there has been a lot of work on designing FASTQ compressors, most of them lack in support of one or more crucial properties, such as support for variable length reads, scalability to high coverage datasets, pairing-preserving compression and lossless compression.Results: In this work, we propose SPRING, a reference-free compressor for FASTQ files. SPRING supports a wide variety of compression modes and features, including lossless compression, pairing-preserving compression, lossy compression of quality values, long read compression and random access. SPRING achieves substantially better compression than existing tools, for example, SPRING compresses 195 GB of 25x whole genome human FASTQ from Illumina's NovaSeq sequencer to less than 7 GB, around 1.6x smaller than previous state-of-the-art FASTQ compressors. SPRING achieves this improvement while using comparable computational resources.Availability: SPRING can be downloaded from Github. Bio:  Shubham Chandak received a B.Tech. degree in Electrical Engineering from IIT Bombay, India in 2016 and a M.S. degree in Electrical Engineering from Stanford University in 2018. He is currently a Ph.D. candidate in the Department of Electrical Engineering at Stanford University, where he works on genomic data compression and DNA storage. 2:30 pm - 2:45 pm Break 2:45 pm - 3:15 pm Song Han, Massachusetts Institute of TechnologyDeep Compression and Hardware-Centric AutoML for Efficient Deep Learning Computing [video] 3:15 pm - 4:00 pm Panel: Compression via and for machine learning [video]Hyeji Kim, Samsung AI Center Cambridge, UK ▸ Bio:  Hyeji Kim is a researcher at Samsung AI Center Cambridge in United Kingdom. She received her Ph.D. in Electrical Engineering from Stanford University in 2016. Following her Ph.D., she worked as a postdoctoral research associate at the Coordinated Science Laboratory at University of Illinois at Urbana-Champaign. She is a recipient of Stanford Graduate Fellowship and participated in the Rising Stars in EECS Workshop in 2015. Her research interests are in information theory, machine learning, and the interdisciplinary research across the two areas. Dmitri Pavlichin, Stanford University▸ Bio:  Dmitri Pavlichin is a postdoctoral scholar advised by Tsachy Weissman at Stanford. He specializes in compression of databases and genomic data, as well as information theory and statistics. He holds a PhD in physics from Stanford, where he received a Stanford Graduate Fellowship, and a BS in physics from Harvard. He is currently trying to make it easier for more people to benefit from high-performing but hard-to-develop compression schemes for tabular data and DNA. Oren Rippel, WaveOne ▸ Bio:  Oren is a co-founder and the CTO of WaveOne. He holds a Ph.D. from MIT. He was a research fellow at Harvard, and worked at Facebook AI Research as part of the computer vision team. He specializes in machine learning, with expertise in deep learning and Bayesian optimization. Oren holds a B.Sc. in Combined Honours in Mathematics and Physics from UBC, where he was the valedictorian of his class. Oren is a recipient of the MIT Presidential Graduate Fellowship, was recognized as a Wesbrook Scholar, and is a recipient of the Millennium Excellence Award. George Toderici, Google Moderator:Kedar Tatwawadi, Stanford University 4:00 pm - 4:15 pm Break 4:15 pm - 4:30 pm Ashutosh Bhown, Palo Alto High SchoolSoham Mukherjee, Monta Vista High SchoolSean Yang, Saint Francis High SchoolHumans are still the best lossy image compressors [video] ▸ Abstract:  Lossy image compression has been studied extensively in the context of typical loss functions such as RMSE, MS-SSIM, etc. However, it is not well understood what loss function might be most appropriate for human perception. Furthermore, the availability of massive public image datasets appears to have hardly been exploited in image compression. In this work, we perform compression experiments in which one human describes images to another, using publicly available images and text instructions. These image reconstructions are rated by human scorers on the Amazon Mechanical Turk platform and compared to reconstructions obtained by existing image compressors. In our experiments, the humans outperform the state of the art compressor WebP in the MTurk survey on most images, which shows that there is significant room for improvement in image compression for human perception. 4:30 pm - 5:00 pm Jonathan Dotan, HBO, Stanford University Compression, Self-Sovereign Data and a Decentralized Internet [video] 5:00 pm - 6:00 pm Closing remarks + Poster session + RefreshmentsList of Posters

• Parking permit machines: You may pay for parking at any machine in the lot or structure ($2.00 -$2.50 per hour). Most of the machines accept Visa, MasterCard, or Discover.