Srikanth Kuthuru grew up in the port city of Vizag in south east India before studying for his B.Tech in Electrical Engineering at IIT Madras, where he became interested in machine learning and signal analysis. He’s now a graduate student in Electrical and Computer Engineering at Rice University and is close to completing his masters thesis on computational drug discovery using machine learning. Outside work, Srikanth plays squash and volleyball and enjoys hiking when he has the chance to get outdoors.
HP: What are you working on this summer?
I'm working on two projects in the Emerging Compute Lab. The main one is on acoustic echo cancellation systems. These are used in mobile phones, teleconferencing systems, concert halls etc. to prevent the loudspeaker sound feeding back into the microphone and creating an echo. My HP Labs mentor, Sunil Bharitkar is creating a pipeline for a new kind of immersive audio conferencing system and my challenge has been to build an acoustic echo cancellation algorithm to integrate in this system.
HP: How’s that been going?
I began with an extensive review of the research literature and from that I created a set of algorithms that represent the state-of-the art and that work pretty well on our systems. Now I'm trying to improve upon what we have. In particular, I’m looking into algorithms that are more computationally efficient. I’m also working on improving echo cancellation on low-end loudspeaker systems – these add a layer of complexity to the challenge because the loudspeaker nonlinear distortions can create problems for acoustic echo canceller (AEC) in terms of ERLE (echo return loss enhancement) since typical AEC filters operate on the linear part of the mixed speech.
HP: What research approach are you using for the loudspeaker work?
One thing we are looking at is whether we can model these distortions using recurrent neural networks. Neural networks can work well to approximate the kind of non-linear relationships that you see between the original audio and the distortions created by the speakers.
HP: And what about the other project?
This is about convolutional neural networks, a subset of deep learning algorithms, and this work comes under the edge computing research the Lab does. I’ve been doing a survey of all the existing fast inference methods for convolutional neural networks. Generally, convolutional neural networks take a lot of time to train and a lot of processing power to work well – but they do tasks like helping us recognize people and objects, things that we’d like relatively low-power mobile and edge devices to be able to do. Fast inference algorithms speed up the process and are a major area of industry interest, so there are a lot of algorithms out there. My job has been to survey the algorithms that exist in the academic literature and test them against some lab benchmarks. Then I’m creating a pipeline of promising algorithms that HP Labs researchers might want to explore as they develop next-generation edge devices. In a related side-project, I’ve also been making contributions to a lab project that is developing new machine learning-based data extraction and analysis algorithms for audio content classification.
HP: What has struck you in particular about working at HP Labs?
I’ve definitely learned a lot. The research atmosphere is very much like a university, which I didn’t expect in a company. And my advisors are very supportive and always encouraging me to try new ideas, which is great. The discussions I’m part of also often take place at a very theoretical level – sometimes more than you see even at the university.
HP: Is the internship changing your idea of what you might want to do for a PhD?
Yes, it’s giving me an idea of which direction to go in for a PhD topic and which programs would be good to apply for. I’d definitely like to do something in a field related to the work I’m doing here.