deep learning in the wild
Deep Learning in the Wild

i2x̅Conversation training and analysis

Who are you and what’s your role in the company?

My name is Ilya Edrenkin, I am the head of machine learning at i2x̅, a German startup specializing in speech technology.

How did you get in touch with deep or machine learning in general?

For several years I was leading speech technology – speech recognition, text-to-speech, voice biometry – at Yandex, the most popular Russian web search engine and largest internet company. Later, I was in charge of R&D for self-driving cars there. Both voice technology and robotics are ML-heavy topics, and that’s how I got in touch and dived deeply into them.

Describe what your project/business does. Which problem are you solving?

Currently, I am building the machine learning stack at i2x̅. Particular technologies include automatic speech recognition, emotion detection and building models that predict the probability of a sales or service call being successful. In general, we want to build a sound data-driven understanding of the communication process: what makes people trust their partner in conversations, and how can communication be improved automatically.

Can you describe your tech stack, what libraries hardware / architecture are you using and why?

We are following a pragmatic approach, where we use the best parts of open-source technologies (given they have appropriate licenses) and either improve them or build on top of them. However, certain parts of the system – those that are most relevant to the business – are completely proprietary. A good example of an open-source library for ML is Google’s TensorFlow. It has a simple Python interface for model training and experimentation; at the same time, its well-designed C++ interface is really handy for the deployment of scalable, reliable and heavy-loaded systems.

What were some of the biggest challenges technical or business related you encountered and how did you overcome them?

The biggest challenge in machine learning is almost always the lack of data, or at least the lack of in-domain labeled data. To combat that, all kinds of tricks are used: actively researching possible data sources, building specialized labeling tools, and employing domain conversion techniques.

How has the industry response been to your solution?

We were already able to build an automatic speech recognition system for German which visibly outperforms competitors in our target domain. This is just the beginning: as we get more and more data, we are growing our capability to build very powerful and unique models, products and tools.

What excites you most about the future of deep learning? Where do you hope the technology is going / where do you see the greatest opportunities?

It’s important to consider deep learning in a more general context of machine learning. In the near future I strongly hope to see better mathematical explanations of the reasons why deep learning works so well and how we can make it work even better. I also see huge potential in biologically inspired models: if we are able to reverse-engineer the human brain one day, it will be a huge breakthrough for artificial intelligence. I strongly believe that as a community, we should pay close attention to these efforts and work hard on making that happen.

How do you keep up with new papers and publications?

I read new arxiv papers daily, and monitor the research blogs of the leading companies and institutions. Also we have a reading club at i2x̅, where we present the most prominent or practically important research papers to each other and listen to scientific talks on machine learning.

What do you wish you had known before getting into deep learning?

The deeper your background in mathematics and software engineering, the higher the output that you can produce as a machine learning specialist. It’s also important to consider real-world constraints and business needs – asymptotically optimal solutions are not always the best way to go.

What do you wish to see on a candidate’s resume who is applying for an engineering position involving deep learning?

I want to see verifiable accomplishments: open-source software that one has built or contributed to, competitive achievements (algorithmic contests/kaggle/CTFs), and well-written academic papers relevant to machine learning.

What do you recommend candidates do to make themselves more attractive to employers looking to hire a deep learning professional?

Practice, and given enough practice, build things that matter. You will get noticed.

Can you give us some facts about your current team and point out, why a candidate may want to join you?

We are always open to talk with outstanding candidates. My goal is to build one of the strongest machine learning teams in Europe, and so far I am reasonably happy with the results. We have a number of demanding tasks, which require excellence both in machine learning and in software engineering, a well-designed management system with clear goals, pragmatic and friendly atmosphere. If you want to work on speech technology, we are one of the best places to join.