Resume – Anders Øland

- Hide menu

Resume

Ph.D. Dissertation

Efficient Deep Learning

My dissertation will be on the topic of efficient deep learning. This captures the essence of my work at Carnegie Mellon University over the past 5-6 years. Multiple subjects will be covered including, but not limited to, algorithms, theory, distributed learning, and implementation. Expected graduation: fall 2018.

Research Interests & Ambition

Algorithms & Theory for Deep Learning

I have been working with deep learning & and neural networks since 2011. Since March 2016 my work has revolved solely around developing new methods and training algorithms. In short, my mission is to show that things are sometimes simpler than we think. There is a bit of a veil of mystery surrounding neural nets. Perhaps due to the whole brain analogy? This affects how we approach problems and research. Thus, my goal is to take new and orthogonal views on the problem of training deep neural networks. This is what I am doing for my dissertation.

Software for Deep Learning

Over the past few years, I have written my own custom software package for deep learning, NEUTELLA. It is very efficient. For most workloads I have tried on a single GPU it is faster than even Nervana’s Neon (which has a reputation of being very fast), and of course much faster than Google’s TensorFlow (which seems to be built to scale, rather than getting maximum performance on individual GPUs). The speed is obtained by use of aggressive auto-tuning of CUDA kernels plus some other tricks (that I will probably put in a paper eventually).

Computational Music Production

My ambition is to establish Computational Music Production (CMP) as a new research field in academia.

With the introduction of computers, and the increasing accessibility of professional-quality music software,  music production has moved from the studios and into the homes of musicians around the world. This has prompted the record labels to expect professional sounding recordings from bands and artists, even when demoing their material, while at the same time cutting recording budgets dramatically. However, producing such recordings—even with access to the necessary tools—remains a daunting task for most.

While development of CMP technologies is actively taking place in the industry, the area remains mostly unexplored in academic research. With the increasing need for more intelligent and user-friendly tools, for both amateurs and professionals, introducing AI-boosted technologies into CMP will bring scientific, artistic and economic benefits.

As a research field CMP is closely related to other fields, such as Computer Music, Music Information Retrieval (MIR), AI, Signal Processing, and Human-Computer Interaction (HCI). Thus, many technologies already in existence today would be obvious to implement in CMP-related applications. However, unlike computer music composition, CMP deals with the automation of manipulation and enhancement of recordings. Unlike MIR, CMP deals with the creation of music rather than search and analysis. Moreover, CMP requires subjectively pleasing musical results. Hence, a critical area of CMP research is to develop models and measures for the subjective evaluation of musical features. Other areas offer useful tools and techniques, but so far have not directly addressed CMP problems. With a better understanding of evaluation techniques, a vast amount of AI, machine learning, and HCI techniques can be directly applied to many typical problems in CMP and possibly revolutionize the field of music production.

Get My Official Resume Here

cv-anders-oland.pdf 

PUBLICATIONS

BE CAREFUL WHAT YOU BACKPROPAGATE: A CASE FOR LINEAR OUTPUT ACTIVATIONS & GRADIENT BOOSTING, Anders Øland, Aayuhs Bansal, Roger B. Dannenberg & Bhiksha Raj [LINK]

VOICE DISGUISE BY MIMICRY: DERIVING STATISTICAL ARTICULOMETRIC EVIDENCE TO EVALUATE CLAIMED IMPERSONATION, Singh, R., Jimenez, A., & Øland , A. (2017, February 17). IET Biometrics. Institution of Engineering and Technology (IET). [LINK]

REDUCING COMMUNICATION OVERHEAD IN DISTRIBUTED LEARNING BY AN ORDER OF MAGNITUDE (ALMOST), Anders Øland & Bhiksha Raj, 2015 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
** NOTE: Won Best Paper Award at ICLR 2016 by proxy. That is, the paper that officially won presented exactly the same work as our paper (more or less) – only two years later. We kindly asked the authors to cite us, but apparently they could not be bothered. Hence, I know claim that our paper won… by proxy. **
[pdf]