Large language model (LLM) tokenizers act as structured compressors: by mapping text to discrete token sequences, they determine token count (and thus compute and context usage) and the statistical structure seen by downstream models. Despite their central role in LLM pipelines, the link between tokenization, compression efficiency and induced structure is not well understood. We empirically demonstrate that tokenizer training scale redistributes entropy: as training data grows, the token stream becomes more diverse in aggregate (higher unigram entropy) yet markedly more predictable in-context (lower higher-order conditional entropies), indicating that tokenization absorbs substantial short-range regularity although these gains degrade under train-test domain mismatch. To ground these observations, we first benchmark i) pretrained GPT-family tokenizers as black-box compressors across various domains, and ii) learned tokenizers across configurations spanning vocabulary size, training scale, and domain. Next, we study tokenization as a transform for universal compression and introduce a compression-aware BPE variant. Finally, we adopt a channel lens and introduce capacity-utilization metrics to analyze tokenizer behaviour and outline implications for downstream modeling. Put together, our results expose various trade-offs between compression, induced structure, and robustness under domain shift, and motivate principled, compression-aware tokenizer design.

Video represents the majority of internet traffic today, driving a continual race between the generation of higher quality content, transmission of larger file sizes, and the development of network infrastructure. In addition, the recent COVID-19 pandemic fueled a surge in the use of video conferencing tools. Since videos take up considerable bandwidth ($\sim 100$Kbps to a few Mbps), improved video compression can have a substantial impact on network performance for live and pre-recorded content, providing broader access to multimedia content worldwide. We present a novel video compression pipeline, called Txt2Vid, which dramatically reduces data transmission rates by compressing webcam videos (“talking-head videos”) to a text transcript. The text is transmitted and decoded into a realistic reconstruction of the original video using recent advances in deep learning based voice cloning and lip syncing models. Our generative pipeline achieves two to three orders of magnitude reduction in the bitrate as compared to the standard audio-video codecs (encoders-decoders), while maintaining equivalent Quality-of-Experience based on a subjective evaluation by users ($n=242$) in an online study. The Txt2Vid framework opens up the potential for creating novel applications such as enabling audio-video communication during poor internet connectivity, or in remote terrains with limited bandwidth. The code for this work is available athttps://github.com/tpulkit/txt2vid.git.

We consider lossless compression based on statistical data modeling followed by prediction-based encoding, where an accurate statistical model for the input data leads to substantial improvements in compression. We propose DZip, a general-purpose compressor for sequential data that exploits the well-known modeling capabilities of neural networks (NNs) for prediction, followed by arithmetic coding. DZip uses a novel hybrid architecture based on adaptive and semi-adaptive training. Unlike most NN-based compressors, DZip does not require additional training data and is not restricted to specific data types. The proposed compressor outperforms general-purpose compressors such as Gzip (29% size reduction on average) and 7zip (12% size reduction on average) on a variety of real datasets, achieves near-optimal compression on synthetic datasets, and performs close to specialized compressors for large sequence lengths, without any human input. While the main limitation of NN-based compressors is generally the encoding/decoding speed, we empirically demonstrate that DZip achieves comparable compression ratio to other NN-based compressors while being several times faster. The source code for DZip and links to the datasets are available at https://github.com/mohit1997/Dzip-torch.

COVID-19 has made video communication one of the most important modes of information exchange. While extensive research has been conducted on the optimization of the video streaming pipeline, in particular the development of novel video codecs, further improvement in the video quality and latency is required, especially under poor network conditions. This paper proposes an alternative to the conventional codec through the implementation of a keypoint-centric encoder relying on the transmission of keypoint information from within a video feed, as shown in Figure 1. The decoder uses the streamed keypoints to generate a reconstruction preserving the semantic features in the input feed. Focusing on video calling applications, we detect and transmit the body pose and face mesh information through the network, which are displayed at the receiver in the form of animated puppets. Using efficient pose and face mesh detection in conjunction with skeleton-based animation, we demonstrate a prototype requiring lower than 35 kbps bandwidth, an order of magnitude reduction over typical video calling systems. The added computational latency due to the mesh extraction and animation is below 120ms on a standard laptop, showcasing the potential of this framework for real-time applications. The code for this work is available at https://github.com/shubhamchandak94/digital-puppetry/ and the full version is available on arXiv [1].

In the emerging field of DNA storage, data is encoded as DNA sequences and stored. The data is read out again by sequencing the stored DNA. Nanopore sequencing is a new sequencing technology that has many advantages over other methods; in particular, it is cheap, portable, and can support longer reads. While several practical coding schemes have been developed for DNA storage with nanopore sequencing, the theory is not well understood. Towards that end, we study a highly abstracted (deterministic) version of the nanopore sequencer, which highlights key features that make its analysis difficult. We develop methods and theory to understand the capacity of our abstracted model, and we propose efficient coding schemes and algorithms.

MOTIVATION: Nanopore sequencing provides a real-time and portable solution to genomic sequencing, enabling better assembly, structural variant discovery and modified base detection than second generation technologies. The sequencing process generates a huge amount of data in the form of raw signal contained in fast5 files, which must be compressed to enable efficient storage and transfer. Since the raw data is inherently noisy, lossy compression has potential to significantly reduce space requirements without adversely impacting performance of downstream applications. RESULTS: We explore the use of lossy compression for nanopore raw data using two state-of-the-art lossy time-series compressors, and evaluate the tradeoff between compressed size and basecalling/consensus accuracy. We test several basecallers and consensus tools on a variety of datasets at varying depths of coverage, and conclude that lossy compression can provide 35-50% further reduction in compressed size of raw data over the state-of-the-art lossless compressor with negligible impact on basecalling accuracy (≲0.2% reduction) and consensus accuracy (≲0.002% reduction). In addition, we evaluate the impact of lossy compression on methylation calling accuracy and observe that this impact is minimal for similar reductions in compressed size, although further evaluation with improved benchmark datasets is required for reaching a definite conclusion. The results suggest the possibility of using lossy compression, potentially on the nanopore sequencing device itself, to achieve significant reductions in storage and transmission costs while preserving the accuracy of downstream applications. AVAILABILITYAND IMPLEMENTATION: The code is available at https://github.com/shubhamchandak94/lossy_compression_evaluation. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Time series data compression is emerging as an important problem with the growth in IoT devices and sensors. Due to the presence of noise in these datasets, lossy compression can often provide significant compression gains without impacting the performance of downstream applications. In this work, we propose an error-bounded lossy compressor, LFZip, for multivariate floating-point time series data that provides guaranteed reconstruction up to user-specified maximum absolute error. The compressor is based on the prediction-quantization-entropy coder framework and benefits from improved prediction using linear models and neural networks. We evaluate the compressor on several time series datasets where it outperforms the existing state-of-the-art error-bounded lossy compressors. The code and data are available at https://github.com/shubhamchandak94/LFZip.

We consider lossless compression based on statistical data modeling followed by prediction-based encoding, where an accurate statistical model for the input data leads to substantial improvements in compression. We propose DZip, a general-purpose compressor for sequential data that exploits the well-known modeling capabilities of neural networks (NNs) for prediction, followed by arithmetic coding. DZip uses a novel hybrid architecture based on adaptive and semi-adaptive training. Unlike most NN based compressors, DZip does not require additional training data and is not restricted to specific data types, only needing the alphabet size of the input data. The proposed compressor outperforms general-purpose compressors such as Gzip (on average 26% reduction) on a variety of real datasets, achieves near-optimal compression on synthetic datasets, and performs close to specialized compressors for large sequence lengths, without any human input. The main limitation of DZip in its current implementation is the encoding/decoding time, which limits its practicality. Nevertheless, the results showcase the potential of developing improved general-purpose compressors based on neural networks and hybrid modeling.

As magnetization and semiconductor based storage technologies approach their limits, bio-molecules, such as DNA, have been identified as promising media for future storage systems, due to their high storage density (petabytes/gram) and long-term durability (thousands of years). Furthermore, nanopore DNA sequencing enables high-throughput sequencing using devices as small as a USB thumb drive and thus is ideally suited for DNA storage applications. Due to the high insertion/deletion error rates associated with base-called nanopore reads, current approaches rely heavily on consensus among multiple reads and thus incur very high reading costs. We propose a novel approach which overcomes the high error rates in basecalled sequences by integrating a Viterbi error correction decoder with the basecaller, enabling the decoder to exploit the soft information available in the deep learning based basecaller pipeline. Using convolutional codes for error correction, we experimentally observed 3x lower reading costs than the state-of-the-art techniques at comparable writing costs.The code, data and Supplementary Material is available at https://github.com/shubhamchandak94/nanopore_dna_storage.

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. The images, results and additional data is available at https://compression.stanford.edu/human-compression.

Sequential data is being generated at an unprecedented pace in various forms, including text and genomic data. This creates the need for efficient compression mechanisms to enable better storage, transmission and processing of such data. To solve this problem, many of the existing compressors attempt to learn models for the data and perform prediction-based compression. Since neural networks are known as universal function approximators with the capability to learn arbitrarily complex mappings, and in practice show excellent performance in prediction tasks, we explore and devise methods to compress sequential data using neural network predictors. We combine recurrent neural network predictors with an arithmetic coder and losslessly compress a variety of synthetic, text and genomic datasets. The proposed compressor outperforms Gzip on the real datasets and achieves near-optimal compression for the synthetic datasets. The results also help understand why and where neural networks are good alternatives for traditional finite context models.

MOTIVATION: High-Throughput Sequencing 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 25× whole genome human FASTQ from Illumina's NovaSeq sequencer to less than 7 GB, around 1.6× smaller than previous state-of-the-art FASTQ compressors. SPRING achieves this improvement while using comparable computational resources. AVAILABILITY AND IMPLEMENTATION: SPRING can be downloaded from https://github.com/shubhamchandak94/SPRING. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Motivation: New Generation Sequencing (NGS) technologies for genome sequencing produce large amounts of short genomic reads per experiment, which are highly redundant and compressible. However, general-purpose compressors are unable to exploit this redundancy due to the special structure present in the data. Results: We present a new algorithm for compressing reads both with and without preserving the read order. In both cases, it achieves 1.4×-2× compression gain over state-of-the-art read compression tools for datasets containing as many as 3 billion Illumina reads. Our tool is based on the idea of approximately reordering the reads according to their position in the genome using hashed substring indices. We also present a systematic analysis of the read compression problem and compute bounds on fundamental limits of read compression. This analysis sheds light on the dynamics of the proposed algorithm (and read compression algorithms in general) and helps understand its performance in practice. The algorithm compresses only the read sequence, works with unaligned FASTQ files, and does not require a reference. Contact: [email protected]. Supplementary information: Supplementary material are available at Bioinformatics online. The proposed algorithm is available for download at https://github.com/shubhamchandak94/HARC.