<div><strong>Who are you and what’s your role in the company?<br></strong><br></div><div>I am Oguzhan Gencoglu. I am the Co-founder and Head of Data Science of Top Data Science. My role is to lead the machine learning parts of the projects, guide our data scientists as well as implement ML/DL solutions on a daily basis.<br><br></div><div><strong>How did you get in touch with deep or machine learning in general?<br></strong><br></div><div>My machine learning journey started during the last year of my B.Sc. studies, accompanied with lots of signal processing. Then during my M.Sc. studies in 2012, I started getting involved in deep learning and eventually wrote my thesis on one of the first DL approaches to Sound Event Classification. During my Ph.D. studies in Finland, I have further utilized machine learning/deep learning techniques in numerous solutions involving multimodal (tabular, image, text, time-series) health data analysis. My journey with DL continues at Top Data Science where we solve prominent problems of the world with it.<br><br></div><div><strong>Describe what your project/business does. Which problem are you solving?<br></strong><br></div><div>At Top Data Science, we provide AI as a Service. The problems we solve vary from “prostate cancer detection from magnetic resonance images” to “industrial process optimization in forestry industry”.<br><br></div><div><strong>Can you describe your tech stack, what libraries hardware / architecture are you using and why?<br></strong><br></div><div>Our tech stack is strictly consisting of open source tools and libraries. Most of our problems are solved and delivered in Python and a small portion in R or C++. On a daily basis we utilize Numpy, Scipy, pandas, scikit-learn for generic ML and TensorFlow, PyTorch and keras for DL. We have our own workstations with GPUs but use AWS as well, especially for CPU-heavy analysis.<br><br></div><div><strong>What were some of the biggest challenges technical or business related you encountered and how did you overcome them?<br></strong><br></div><div>In my experience, the business challenges have been more difficult to overcome than the technical ones. Technical challenges are more predictable, i.e., one usually has an idea of which parts of the project may be a trouble-maker. On the other hand, business challenges such as establishing a clear communication (with the client) of what is possible to achieve with the data at hand in a given timeframe has been extremely time-consuming. The ultimate caveat to overcome these problems is to patiently explain the whole development pipeline and emphasize the difference of state-of-the-art machine learning solutions with standard software development projects over and over.<br><br></div><div><strong>How has the industry response been to your solution?<br></strong><br></div><div>It depends. Experts in the industry have been extremely open and positive on our solutions. They can immediately understand that AI can enhance their working pipelines and decrease their workload. Administrative and legal staff have been more resistant to these technologies involving AI, especially in healthcare, but we believe this is going to change.<br><br></div><div><strong>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?<br></strong><br></div><div>Semi-supervised learning combined with deep architectures excite me the most. Most industrial problems rely heavily on lots of high quality, annotated data.<br><br></div><div><strong>How do you keep up with new papers and publications?<br></strong><br></div><div>I regularly follow the ML subreddit on reddit. We also have regular internal seminars in the company in which one of us presents interesting publications, libraries, blogs etc. at each session.<br><br></div><div><strong>What do you wish to see on a candidate’s resume who is applying for an engineering position involving deep learning?<br></strong><br></div><div>I really appreciate when a candidate goes beyond the standard benchmark tasks when it comes to experience. Object recognition from images and news article classification is all cool and jazz but how about tasks involving multi-modal data for example both images and text. These kinds of problems are extremely common in the industry.<br><br></div><div><strong>Can you give us some facts about your current team and point out, why a candidate may want to join you?<br></strong><br></div><div>We are a Helsinki-based AI startup of around 10 people with 6 different nationalities (so far). We are not a single-product company and our projects usually do not last for years. Therefore, you will not only gain the knowledge of applying ML/DL solutions in numerous industries but will also have the chance to acquire knowledge of that field itself from the experts. I did not know much about prostate cancer, production parameters affecting paper quality or route optimization of nautical vessels before working on those problems. There is a unique joy in that sort of work.</div>

Top Data Science

AI and Machine Learning for Healthcare and Industrial Internet

<div><strong>Who are you and what’s your role in the company?<br></strong><br></div><div>As the Chief Scientist at Infinia ML, I oversee all of our company’s technical operations and research. I’m also the Vice Provost for Research at Duke University, where I’ve spent over two decades as a Professor of Electrical and Computer Engineering. My research has been widely published, and I was the most-published AI researcher at NIPS 2017.<br><br></div><div>I earned my BS, MS, and Ph.D. in electrical engineering from the University of Maryland.<br><br></div><div><strong>How did you get in touch with deep or machine learning in general?<br></strong><br></div><div>My initial interest was in electromagnetics, but in my early 30s I realized most of the field’s major innovation had already happened. So I did something unusual – I pivoted to where I could have greater impact which, for me, was machine learning.<br><br></div><div><strong>Describe what your project/business does. Which problem are you solving?<br></strong><br></div><div>We are a team of machine learning experts focused on delivering business impact. We help companies solve some of their toughest data challenges and unlock their biggest data opportunities. Many teams don’t have the expertise or the bandwidth to tackle these challenges – in fact, they may not fully realize what’s possible.<br><br></div><div>We bring deep machine learning expertise to projects across a number of industries including healthcare, human resources, legal, and manufacturing. Our algorithms help companies:<br><br></div><ul><li>Prevent security breaches, product defects, and medical errors by scanning images and videos for anomalies that humans might miss.</li><li>Extract meaning from large numbers of complex documents, automating a repetitive task in a fraction of the time.</li><li>Perform biomedical analysis to detect genetic risk patterns without invasive procedures.</li><li>Automatically align unique data into a standardized industry framework, saving manual effort and allowing more accurate analysis.</li><li>Visualize reality by modeling 3D images from 2D counterparts.</li></ul><div>In every case, our goal is to help lower costs, increase efficiency, and/or achieve a new breakthrough. Many of us come from academia, but everything we do must have a tangible business impact.<br><br></div><div><strong>Can you describe your tech stack, what libraries hardware / architecture are you using and why?<br></strong><br></div><div>We look for prospective data scientists who have experience with Python, including libraries such as NumPy, SciPy, pandas, scikit-learn, matplotlib, TensorFlow, and NLTK. We work with deep learning models including CNN and RNN architectures.<br><br></div><div>We also look for experience working with large datasets, including NoSQL and relational databases, experience with Linux, and experience with cloud computing and Amazon EC2.<br><br></div><div><strong>What were some of the biggest challenges technical or business related you encountered and how did you overcome them?<br></strong><br></div><div>Getting the right client data is of course an essential and ongoing challenge. So is helping clients understand what is actually possible. They may not realize all the ways that their data could help them lower costs, increase efficiency, and achieve new breakthroughs for their internal operations, or external products and services.<br><br></div><div>Machine learning problems that appear superficially different may actually have similar solutions. So we might spend time showcasing a variety of different use cases in different industries just to help clients spark their own ideas.<br><br></div><div><strong>How has the industry response been to your solution?<br></strong><br></div><div>I was getting a lot of inbound requests at Duke to help companies solve machine learning and data challenges.&nbsp; That’s why we started the company.<br><br></div><div>Demand for our services at Infinia ML is very strong. We are working with a variety of Fortune 500 companies on a number really exciting projects.<br><br></div><div><strong>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?<br></strong><br></div><div>I’m excited to see our field, which has been around for decades, making a real, tangible impact in the world.<br><br></div><div>I compare its place in society to the mobile phone. In the 1990s, many people were aware of mobile phones, but not everybody had them or felt they needed them. Today, mobile phones are ubiquitous - they touch every industry and we all wonder how we ever lived without them. That’s how people are going to feel about machine learning/deep learning in 10 years.<br><br></div><div><strong>How do you keep up with new papers and publications?<br></strong><br></div><div>I’m reading (and writing) all the time! Advising the up-and-coming researchers in my machine learning lab at Duke is a key way I keep on top of the latest developments.<br><br></div><div><strong>What do you wish you had known before getting into deep learning?<br></strong><br></div><div>I wish I’d understood just how meaningful and interesting the work would be – I would’ve gotten into it sooner!<br><br></div><div><strong>What do you wish to see on a candidate’s resume who is applying for an engineering position involving deep learning?<br></strong><br></div><div>Beyond the obvious technical proficiency, we’re looking for people who can thrive in a fast-moving startup environment. Unlike a large company, team members may work on multiple projects at once and may be asked to pitch in on other projects that are going on. We work in a number of industries, everything from healthcare to legal to marketing to people analytics, so we appreciate candidates that want that variety.<br><br></div><div>To find the right mentality, we have a preference for candidates who have successfully worked at fast-moving companies before.<br><br></div><div><strong>What do you recommend candidates do to make themselves more attractive to employers looking to hire a deep learning professional?<br></strong><br></div><div>At Infinia ML, we’re focused on the practical application of academic research. The more you can show us that you are too, the more employable you’ll appear. That might mean having done internships or even interesting side projects that show your ability to put theory into practice.<br><br></div><div><strong>Can you give us some facts about your current team and point out, why a candidate may want to join you?<br></strong><br></div><div>We’re growing rapidly and are working with some of the largest companies and organizations in the world. This means that new team members can have a massive impact not just in our company, but in our industry.</div>

Infinia ML

Machine Learning Solutions for Business

<div><strong>Who are you and what’s your role in the company?<br></strong><br></div><div><strong>Timon</strong>: My name is Timon Ruban and I am the Co-Founder and CTO of luminovo.ai.<br><br></div><div><strong>Sebastian</strong>: I am Sebastian Schaal, the second Co-Founder and CEO of luminovo.ai.<br><br></div><div><strong>How did you get in touch with deep or machine learning in general?<br></strong><br></div><div><strong>Timon</strong>: I first studied electrical engineering at ETH Zurich and had actually been pretty oblivious to the whole machine learning hype before I arrived at Stanford for my master’s degree. There I took Andrew Ng’s ML class and quickly realized that all the math I learned at ETH could actually be put to good use to solve real-world problems using machine learning.<br><br></div><div><strong>Sebastian</strong>: My journey towards ML was in some way similar to the one of Timon. I went out to Stanford to further position myself at the intersection of Business and Tech. However, as most people there it did not take long until I got dragged into the force field of ML. During the second half of my program I tried to take all ML classes Stanford had to offer and when I returned to Munich to finish my second master’s program, I made sure to follow the classes I missed during my time in California (like Karpathy’s CS231n-class on CNNs).<br><br></div><div><strong>Describe what your project/business does. Which problem are you solving?<br></strong><br></div><div><strong>Sebastian</strong>: When we came back from the Silicon Valley, we found that the adoption of AI is still lacking in many German companies. We think this is due to the fact that 1) AI often requires a new way of thinking about problems and 2) that there are only few people with actual experience in the field able to push this adoption forward. Since Timon and me wanted to apply deep learning in the wild anyways, we decided to start luminovo.ai to help companies kickstart their AI efforts.<br><br></div><div>At luminovo.ai, we often start with helping companies get their thinking straight about which of their problems can be tackled with machine learning and the data they have. Then we deliver functional AI prototypes that can be directly deployed and hopefully generate value for our clients immediately. In contrast to others, our focus is on enabling our client’s employees to get started with AI on their own after we are gone, so instead of selling black boxes and licensing fees, we share our code, documentation and best practices and do workshops at the end of every project explaining our methods and ways of approaching their problem.<br><br></div><div><strong>Can you describe your tech stack, what libraries hardware / architecture are you using and why?<br></strong><br></div><div><strong>Timon</strong>: We are all Pythonistas at Luminovo, so we mostly use TensorFlow (I got addicted to TF Slim during my time at Google) and train our models on Google Cloud. For the occasional traditional ML problem, we usually fall back to pandas/scikit-learn.<br><br></div><div><strong>What were some of the biggest challenges technical or business related you encountered and how did you overcome them?<br></strong><br></div><div><strong>Sebastian</strong>: I think, as a young startup, getting the initial credibility is always tough. It is the classic chicken-and-egg-problem: you need references to close large clients, but you do not get good references without large clients. In this situation, having a decent network and, even more important, people who promote you, is invaluable. Most of our initial customers came from warm introductions, bringing us directly into a meeting with the stakeholders. Another thing that is always a bit of a hassle in Germany is getting the data privacy setup right. There are just a lot of things you have to consider when you work with sensitive data.<br><br></div><div><strong>How has the industry response been to your solution?<br></strong><br></div><div><strong>Sebastian</strong>: We have been thinking a lot about our offering and what would be appealing to us and our clients. I think our USP at the moment is also something our clients value a lot: we are not selling licensing fees or locking in our clients with expensive support contracts. When we take on a project, our goal always is to transfer the knowledge to the employees of our clients.<br><br></div><div><strong>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?<br></strong><br></div><div><strong>Timon</strong>: I think going even more meta than deep learning already is, with approaches like Neural Architecture Search, is a really interesting trend, but it seems like that will only be feasible for the non-Googles of this world (like luminovo.ai) once the hardware gets even faster (Graphcore? Intel? Show me what you got!).<br><br></div><div><strong>How do you keep up with new papers and publications?<br></strong><br></div><div><strong>Timon</strong>: I read newsletters like <a href="https://deeplearningjobs.herokuapp.com/blog/www.deeplearningweekly.com">Deep Learning Weekly</a> and Import AI and do a daily check on r/MachineLearning. In addition, I always just hope my friends are kind enough to let me know when they read a really cool new paper.<br><br></div><div><strong>What do you wish you had known before getting into deep learning?<br></strong><br></div><div><strong>Sebastian</strong>: Before getting into the field, I was not aware of the variety of problems that could be tackled with neural networks. Also, I feel that a lot of people are hesitating to get started since they think they are just too far behind. However, you can already have some meaningful impact with just having domain expertise and a sound understanding of how to tackle problems with DL.<br><br></div><div><strong>What do you wish to see on a candidate’s resume who is applying for an engineering position involving deep learning?<br></strong><br></div><div><strong>Timon</strong>: To me the scarcest resource in deep learning today is still experience. So if an applicant can highlight with his resume that he already used machine learning to solve real-world problems (this can even be within the scope of a really interesting class project), that is always a very good sign. In terms of online courses, I’m a big fan of CS231n and deeplearning.ai’s new 5-part deep learning course.<br><br></div><div><strong>What do you recommend candidates do to make themselves more attractive to employers looking to hire a deep learning professional?<br></strong><br></div><div><strong>Timon</strong>: Pick a problem/dataset you think is interesting and try to re-implement a relatively “state-of-the-art” model, put your code/results on GitHub and maybe write a Medium post about what you learned. That should take a few weekends at most and will give you a great conversation starter for any interview for a deep learning position. Other than that, reading papers and being able to discuss the most recent neural net architectures/trends always makes you look like you know what you are doing.<br><br></div><div><strong>Can you give us some facts about your current team and point out, why a candidate may want to join you?<br></strong><br></div><div><strong>Sebastian</strong>: Besides Timon and I, we had Sebastian Schuon on board from the very beginning as our chairman. Through his successful exit of Stylight to ProsiebenSat1 and his role as an active angel investor after that, he brings quite some entrepreneurial experience to the team. In February, Oskar Triebe, a friend we still know from Stanford, joined us as the Head of Business Development. On the tech side, we are currently working together with some highly talented students from the TU Munich.<br><br></div><div>For us, working at luminovo.ai is an almost unique learning opportunity. Just like us, you will have the chance to deploy AI in the wild in a variety of projects across different industries. This will give you a better understanding of the diversity of challenges out there than working within one company on similar problems and datasets. We think that developing an intuition will even be more valuable in the future, when using and training models will become more and more of a commodity.</div>

Luminovo.ai

Bespoke AI solutions

<div><strong>Describe what your project/business does. Which problem are you solving?<br></strong><br></div><div>Lobster is an AI-powered platform which enables brands, agencies and the press to license visual content directly from social media and cloud archives, offering an alternative to ‘staged’ stock photography.<br><br></div><div><strong>Can you describe your tech stack, what libraries hardware / architecture are you using and why?<br></strong><br></div><div>For our machine learning tasks we use Python, C++, Tensorflow, word2vec and OpenCV, and for the training models we use our computing resources. In the near future we also plan on renting cloud resources.<br><br></div><div><strong>What were some of the biggest challenges technical or business related you encountered and how did you overcome them?<br></strong><br></div><div>The biggest challenge was to handle content without metadata. For example: Photos from Instagram include tags and names, so it’s easy to search for them, but photos from Dropbox don’t feature tagging. To get around this issue, we use different algorithms of computer vision to describe photos, before forwarding this information to our search engine.<br><br></div><div><strong>How has the industry response been to your solution?<br></strong><br></div><div>The response from the industry has been very positive, and we have been featured in several important industry applications such as Techcrunch, the Next Web and AdWeek. There is a real appetite for something different from run-of-the-mill stock photography, and this has been evident in Lobster’s media coverage.<br><br></div><div><strong>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?<br></strong><br></div><div>The nature of our project means we are interested in deep models of computer vision. There are interesting approaches to describing the content of images, detecting objects and making classification. We want to create a system for selecting images using AI without the need for human input.<br><br></div><div><strong>How do you keep up with new papers and publications?<br></strong><br></div><div>We learn new industry trends through scientific papers, subscribing to profile mailings and researching the web for information.<br><br></div><div><strong>What do you wish you had known before getting into deep learning?<br></strong><br></div><div>We have encountered a great deal of difficulty with marked image datasets for training our own detectors and classifiers.</div>

Lobster

Real-time stock photos

<div><strong>Who are you and what’s your role in the company?<br></strong><br></div><div>My name is Ilya Edrenkin, I am the head of machine learning at i2x̅, a German startup specializing in speech technology.<br><br></div><div><strong>How did you get in touch with deep or machine learning in general?<br></strong><br></div><div>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&amp;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.<br><br></div><div><strong>Describe what your project/business does. Which problem are you solving?<br></strong><br></div><div>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.<br><br></div><div><strong>Can you describe your tech stack, what libraries hardware / architecture are you using and why?<br></strong><br></div><div>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.<br><br></div><div><strong>What were some of the biggest challenges technical or business related you encountered and how did you overcome them?<br></strong><br></div><div>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.<br><br></div><div><strong>How has the industry response been to your solution?<br></strong><br></div><div>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.<br><br></div><div><strong>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?<br></strong><br></div><div>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.<br><br></div><div><strong>How do you keep up with new papers and publications?<br></strong><br></div><div>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.<br><br></div><div><strong>What do you wish you had known before getting into deep learning?<br></strong><br></div><div>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.<br><br></div><div><strong>What do you wish to see on a candidate’s resume who is applying for an engineering position involving deep learning?<br></strong><br></div><div>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.<br><br></div><div><strong>What do you recommend candidates do to make themselves more attractive to employers looking to hire a deep learning professional?<br></strong><br></div><div>Practice, and given enough practice, build things that matter. You will get noticed.<br><br></div><div><strong>Can you give us some facts about your current team and point out, why a candidate may want to join you?<br></strong><br></div><div>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.</div>

i2x̅

Conversation training and analysis

<div><a href="https://www.linkedin.com/in/maltepietsch/?originalSubdomain=de">with Malte Pietsch</a><br><br><strong>Can you introduce yourself and the company?</strong></div><div><br></div><div>My name Malte, I am CTO and cofounder of deep set. We are a company in the NLP space working mostly on the open source project <a href="https://github.com/deepset-ai/haystack">Haystack</a> that helps developers build semantic search pipelines.&nbsp;</div><div><br><strong>How did you get started in the field of machine learning and deep learning in particular?</strong><br><br></div><div>I studied a mix of computer and statistics so I think it was kind of inevitable that I ended up doing machine learning, because I really always liked this mix of quantitative methods engineering and solving real world problems with it.&nbsp;</div><div><br></div><div>I was interested in physics and moved more and more towards computer science. The one key event was spending a year at Carnegie Mellon University where I soaked in the atmosphere on campus and witnessed deep learning on the rise conquering computer vision tasks one after the other. I was working on time series data from the health care domain at the time and figured out that one most important signals we got was from the doctors’ notes, not actually the original time series data and that is how I got curious about NLP.</div><div><br></div><div><strong>Did you get your start in NLP at the same time as you did with deep learning?</strong></div><div><br></div><div>Basically, we started with very simple techniques and even those yielded some very interesting results. This piqued my curiosity; we could apparently use these simple techniques, bag of word, word to vec etc. and I saw what happened on the computer vision side and was convinced that deep learning was the future for NLP as well.&nbsp;</div><div><br><strong>How did you leverage your research experience into your work at deepset?</strong><br><br></div><div>The research I did sparked my interest in NLP, but I decided not to pursue academic research, because I really wanted to see NLP in production, so I started working for a couple of startups in Berlin.&nbsp;</div><div>At some point we decided to build deepset, because we saw that on the one hand NLP was on the rise (2017) we could see early things working, for instance transfer learning just came out<a href="https://github.com/google-research/bigbird"> Big Bird</a> and transformers came out a year later and things got easier.</div><div>We also saw this gap between research and industry, when I talked to peers from industry and told them what we were doing and how we applied deep learning in production, we met with a lot of skepticism that this would not work at their scale, so they stuck to working with methods that had come out a decade earlier.</div><div>We thought there has to be something to lower the threshold for production deployment of deep learning methods, at the time it took a lot of engineering resources to make these models stable and fast.</div><div>But we believed in the power of NLP and in our ability to build tooling to bridge the gap from research to production and that is why we started deepset.</div><div><strong><br>Have you had success breaking into those industries with those deep learning based methods, in my experience as recently as a couple of years ago it proved fairly difficult to sell deep learning methods as they were considered black boxes lacking interpretability.</strong></div><div><br></div><div>For us it worked quite well and things certainly became easier after transformers were released. We were able to use transfer learning to adapt to different industries and that really started piquing people’s interest. They started seeing the need for other models, especially for the use cases that we worked on i.e. semantic search and question answering where usually when we talked to them they had already experimented with deep learning techniques, but were not really satisfied.&nbsp;</div><div>I would say this attitude has changed in the industry. The concerns right now are more about how to deploy these models in production, how to do ML-Ops, how to adapt these models to their domain rather than questioning whether they should go for deep learning at all.</div><div><br><strong>You maintain an open source project called Haystack, what is Haystack?</strong></div><div><br></div><div><a href="https://github.com/deepset-ai/haystack">Haystack is Python framework</a> for developers to build semantic search pipelines, it allows you to build information extraction pipelines in a very modular way, core building blocks you can stick together to power full search pipelines. One of these building blocks for instance can be a retriever model identifying documents from a larger set of documents that are relevant to your query or it can be something like a question answering model which you then connect, so that when you have a search query coming in you can define very granularly how it should flow through these blocks.&nbsp;</div><div>It is also tightly integrated with <a href="https://github.com/huggingface">Hugging Face</a> so you can plug it in through hugging face as well. We use it in a more specialized area of NLP, semantic search, to cover everything from end to end from labelling to picking the right model, to exposing a rest pipeline.</div><div>So with Haystack we are able to cover the whole range of a production grade pipeline, not just the neural network part of the model.&nbsp;</div><div><br></div><div>We noticed that in our own client project that&nbsp; while there is tooling for tasks such as document indexing and preprocessing, it is still a lot of work to connect them into a coherent whole. Especially if you are working with domain experts for additional input such as labelling. That is what we set out to simplify.</div><div><br><strong>Is your business case then mostly built around consulting or do you have productized services on top of Haystack?</strong></div><div><br></div><div>We do not do consulting, we do integration work around open source with clients, say custom deployments and help them with a few pieces that are missing. What we see more and more is that we are approached about solving the same kinds of problems and that is why we are building out an enterprise platform / SaaS offering. which is not a traditional open core model with premium features on top of open source that you have to pay for, but rather a complementary platform that focuses on the workflows and integration with domain experts. If you are interested in question and answer, you first need to understand whether it works in your domain and understand that you need to come up with a good dataset, some questions, some answers and for that you need domain experts from other departments.</div><div>Say you are in the pharma or finance industry, our product would bridge the gap between the data scientist and those domain experts from early prototyping to production and monitoring the system.</div><div><br><strong>What kind of tech stack would someone in a deep learning engineering role work with at deepset?</strong><br><br></div><div>The majority of the code base is Python based with some minor things in Rust. From a modelling perspective we are big fans of Pytorch, especially for training and experimentation, for our deployment we sometimes use other runtimes <a href="https://www.onnxruntime.ai/">ONNX</a> for example.</div><div><a href="https://fastapi.tiangolo.com/">Fast API</a> for the API layer, <a href="https://kubernetes.io/de/docs/concepts/overview/what-is-kubernetes/">Kubernetes</a> and <a href="https://aws.amazon.com/">AWS</a> for deploying them at scale. Lately we are experimenting a little with <a href="https://ray.io/">Ray</a> which is a framework for distributed applications that works especially well with machine learning pipelines.</div><div><br></div><div><strong>What have been some technical challenges you have had to overcome?</strong></div><div><br></div><div>One big challenge we are working on constantly is how to make question answering scaleable, I would say mostly from a technical perspective in terms of speed and large datasets, how we tackle that is mostly a mix of scouting research that is out there and hands on experimentation. Especially Facebook's <a href="https://aclanthology.org/2020.acl-tutorials.8.pdf">research on open domain QA</a> proved very useful for designing an efficient scaling mechanism for open domain QA.</div><div>But research is always a starting point to generate good ideas, the hard part is then devising practical, stable and efficient ways to implement it. That is where benchmarking comes into play. We have a whole suite of models that we benchmark so we can identify bottlenecks in the pipeline to tweak and optimize, but also to compare different building blocks that we have on top of Haystack.</div><div>We are also striving to make benchmarking a more integral part of our development workflow, so we started looking into <a href="https://github.com/iterative/cml">CML</a>, to integrate with our Github workflow and include benchmarks automatically with each pull request which allows us to track whether something improved or deteriorated.</div><div>Another kind of challenge stems from our SaaS offering where we are attempting to simplify the adaptation of semantic search or question answering for particular business use cases and industrial domains. A lot of that is related to the question of how to implement efficient transfer learning, but also how can you simplify these workflows and on a more basic level, how can you measure success, what are good metrics for such a search model.</div><div><br><strong>What excites you most about the future of deep learning? Where would you like to see the field going?</strong><br><br></div><div>I am most excited about NLP, that is also why we focused on it, because we believe that at some point language is the API of the human that we still need to rely on messy work arounds to integrate with wizards or certain query languages, because you cannot talk directly to the computer.</div><div>I am mostly looking forward to how information access will change with deep learning as chatbots and search come closer together with search becoming more conversational and more about context and semantics. Eventually natural conversational chatbots could go beyond the one hundred intents that you define, but focus more on the long tail as you start integrating dense retrieval techniques, QA and so on. <br>So I am curious to see how these parts come together. Technically I am most excited about generative models such as <a href="https://openreview.net/forum?id=Bkx0RjA9tX">generative Q&amp;A </a>which holds a lot of promise although there are still some drawbacks. For instance it is still not clear how to evaluate the merit of the answer to a question. Right now, this seems to be the biggest blocker for industry. If you can find a method that assigns certain facts that can be checked and verified to an answer that would lend it some trustworthiness.</div><div><br><strong>How can someone without a PhD in deep learning break into the industry and work in research? What do you look for on a candidate’s resume?</strong></div><div><br></div><div>One way to stand out is definitely open source contributions to projects that are relevant, this is very easy to verify and a transparent way to show excitement about and commitment to the field. It really demonstrates whether deep learning is just an interest that you occasionally read about or whether you are dedicated enough to actually build something. When I look at a resume I try to find out whether this is a person that is more on the theoretical side, following the latest research and trying out new models or whether someone is more practically inclined and wants to make things work.</div><div>The latter type of person has a problem to solve and finds the right methods to apply and a way to engineer around it and that is an exceedingly important property to have in industry.</div><div><br><strong>What is the background of your team, how varied is it?</strong><br><br></div><div>We have a wild mix of people from computational linguists and classical linguists studying Latin and Greek, who eventually ended up coding to the PhD who focused on NLP in academia as well as people with a very strong engineering background who got interested in ML and NLP during their careers.</div><div>It is incredibly important to have people of diverse backgrounds that follow the same vision and come together to build it.&nbsp;</div><div><br><strong>What makes deepset a great place to work and what makes it stand out from other employers?</strong></div><div><br></div><div>We work on some of the most interesting NLP problems and engineering problems, we want them to run at scale and that comes with a great need for innovation. Being on the front lines of NLP yet in close proximity to industry sets us apart from a lot of other places that focus on either research or industry niche.</div><div>Secondly, our open source work; you see the impact of your work directly and you have other developers using it immediately. Last but not least, we are very close to industry seeing problems as they arise when developers use our projects. We know what it takes to bring sematic search and Q&amp;A to the next level.&nbsp;</div><div><br><br></div><div><br></div>

Deep Learning in the Wild