Helping Developers Build Privacy-Friendly, Location-Powered Experiences Across iOS, Android, and the Web

By Dan Abelon, Frances Schwiep, and Mickey Graham

When it comes to the device we keep by our side 90% of the time, we have very little understanding of the location-related context of our smartphones. To be sure, certain companies have found success by taking advantage of the inherently mobile nature of our devices by adding location-based features — think Yelp showing nearby restaurants, Google Maps showing your GPS coordinates, or Uber & Lyft connecting you to nearby drivers — but this is a challenging engineering and data management problem for developers at large, especially at scale and across a growing number platforms.

Companies who want to offer context-aware situations — for example, imagine a grocery chain being able to show their most recent deals when a customer enters a store — have to dedicate multiple engineers and ongoing development resources to the project, often troubleshooting through updates and new versions of each new iOS and Android release. Frustrating the problem, solution providers in this market have historically been dominated by AdTech companies and data monetization companies or resellers, who aren’t afraid to hold your users’ data hostage for pricing leverage, or to sell your customers’ data behind your back. Simply put, we’ve found there is a lack of easy-to-use and privacy-oriented platforms that enable developers to add location-based awareness and functionality to their mobile applications, which is why we were so excited when we met Radar.

Radar is a modern, privacy-first solution that allows mobile application developers to build location-powered experiences on any platform, including iOS, Android, and web. Radar does this by enabling mobile app developers to start generating events based on geofences, places, and insights with a few lines of code, with a dedication to privacy. Unlike other players in the space, Radar is not building any AdTech offerings, thereby aligning their privacy-first approach with the growing security demands developers face in a post GDPR world. Their focus is on empowering developers to easily create better location-tailored experiences, while ultimately making it simple for companies to put privacy back into the hands of their end users.

Founded by Nick Patrick and Coby Berman who met while working at Foursquare, Radar is already trusted by leading companies such as Walmart, Burger King, trivago, and others. Someday soon, we believe that every mobile application will be powered by a new, modern infrastructure stack comprising of an analytics tool, a marketing automation tool, a customer data platform, and now, a location context platform (Radar). It’s an exciting glimpse into what the future of mobile infrastructure could look like when you see the power of all four platforms combined, and it’s just the beginning.

We’re beyond thrilled to be partnering with the team making that future a reality, and we’re excited to announce that we’re leading Radar’s $8M Series A financing, with participation from our friends at Accel and Heavybit, along with Scott Belsky (Behance Founder and Adobe Chief Product Officer), and Ashley Smith (the former VP of Marketing at GitHub). Congrats to the whole team on the milestone. We’re super excited about this next chapter for Radar.

• If you’re a mobile app developer, check out Radar’s SDK here.
• Help share the news on Twitter or LinkedIn.
• Make sure to see Radar’s post here.

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Ventures, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

Announcing Our Investment in Radar’s $8 Million Series A was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

We recently hosted our inaugural Debate{X} at Two Sigma Ventures, where we invited 5 of our Two Sigma colleagues for a Churchill Club style debate on what the most impactful & non-obvious trends are that will shape the security industry over a 5 year timeframe. Frances Schwiep & Mickey Graham from our team co-MC’ed a night of lively debate between Two Sigma employees across engineering management, security infrastructure, platform reliability engineering, security architecture, and security engineering teams.

Each panelist presented their thesis one at a time. The rest of the speakers then had the chance to agree or disagree, offering their rebuttals or support. At the end, we had an audience vote on the most compelling thesis presented. Here were the five theses our security experts offered up for debate:

1. Passwords Will Be Gone — Christos Zoulas, Engineering Management

The traditional means of authentication will gradually disappear given how easily they can be subverted and how the power of what a user can do from his online accounts increases. They will be replaced by multifactor authenticators that combine biometrics (something you are) with hardware tokens (something you have) and will work seamlessly.

2. Your Data Will (Again) Be Yours: Decentralization of Personal Information — Ethan Christ, Security Infrastructure

You will regain control of your personal data and with whom it is shared. Shifts in privacy concerns, regulation, and the developer ecosystem are fostering the creation of backend platforms that will decouple data from apps and diminish the monopolistic control companies like Facebook have over your personal information.

3. Back to the Future: Data-Centric Security — Tad Taylor, Security Architecture

In the past, approaches to computer security focused on protecting data (Mandatory Access Control a lá Bell & LaPadula, Information Flow Models) but with the increasing fluidity of data in the coming years (Cloud, loss of the perimeter, mobile), data-centric security will become paramount. Data will have to be self-protecting and self-describing. It will be possible to set conditions on a data object to limit where and how it can be accessed. For example, a subject may have read access rights to a piece of data, but the conditions for access may depend on the strength of the user’s authentication and attestation of the user’s environment..

4. If You Want to Keep a Secret, Don’t Tell Anyone — Thor Simon, Security Engineering

An increasing pace of high-consequence data breaches, coupled with increased awareness of the susceptibility of online data to hacking, will become a serious drag on consumer adoption of new products and services. A surprising number of consumers may even attempt — unsuccessfully — to withdraw their most sensitive personal information from electronic storage entirely, reverting to paper records and in-person transactions. Such efforts will fail, but this in turn may lead to a reduction in brand loyalty, and paradoxical behavior such as customer flight to smaller players with privacy and security practices that are even less well understood.

5. IoT Devices as a Service —Matthew Selsky, Platform Reliability Engineering

Connected devices (e.g. mobile phones, smart devices, autonomous cars, and IoT sensors) will move from a one-time purchase to a leasing/pay-as-you-go model in the next 5 years. The reason is twofold: consumers will increasingly have difficulty properly securing devices through the lifecycle of ownership and updates, and in the future vulnerabilities may be discovered after a device’s release that are unable to be secured with firmware or software patches (thus rendering them useless).

It was a close count in overall votes (with a lot of back & forth and changing leads), but in the end the audience was most convinced by the thesis that “Passwords Will Be Gone.” We’re going to be hosting a second Debate{X} on AI/ML trends later in March, so make sure to sign up our newsletter if you’re interested in receiving future invites. Additionally, we’d love to hear from you if you have any topics you’d like to see discussed, or if you have any panelists you want to see go head to head for the next Debate{X} crown.

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Ventures, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

5 Security Trends We’re Following: A Recap of our Inaugural Debate{X} was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

The beginning of the year is always one of our favorite times at Two Sigma Ventures. For us, it’s time of reflection, planning, and looking ahead at what’s to come. Judging by all the out of office responses we’ve seen over the last two weeks, it can also be a welcome respite and a well-deserved break, hopefully full of rest and the joys of family and friends.

Every year during this time, we like to share some of the top tech trends we’re most excited about as a team, along with a few predictions for 2019. We’ll revisit as the year progresses to see how we’re doing. So without further ado, here is what we think will shape 2019:

1. We’ll see an increase in the number of companies using data science and advanced engineering to fight climate change — Dan Abelon

It’s no longer a question of whether climate change is happening, but rather a question of how quickly, at what scale, and what can be done to stop it. With a seeming unwillingness for nations to step in to address the problem, I think it’s likely we’ll see more companies using data science and advanced engineering to fight the challenges facing our planet. From lab-grown food and cellular agriculture, which lead to substantially less greenhouse emissions, to optimizing the electrical grid and agriculture production utilizing data science, there are some great opportunities to leverage technology to fight climate change.

2. Quantum computing applications will see slow development and adoption, while hardware tries to live up to the hype — Kyra Durko

The launch of IBM Q and Rigetti’s Quantum Cloud Services made headlines and increased the promise that quantum applications will be deployed in the near future. However, error rates remain high and hardware still has technical milestones to achieve before these applications are proven valuable and become worth financial and human capital investment. This year, I think we will continue to see new hardware approaches, such as photonics and quantum dots to address current hardware shortfalls.

3. 2019 marks a resurgence of interest in space-focused startups and exploration — Mickey Graham

It’s already been a busy year on the space exploration front. With SpaceX paving the way for cheaper satellites and launches (and an upcoming test of their spacecraft, Dragon Crew, that will take astronauts to the ISS aboard a U.S spacecraft for the first time since 2011), I think we can expect a flurry of space-related startup and exploration activity this year. It was an auspicious start to the year with Nasa’s New Horizons’ whizzing by Ulitma Thule at 31,000 miles per hour on New Years’ Eve, and just last week, China’s Chang’e-4 landed on the far side of the moon. We’ll also be celebrating the 50 year anniversary of Apollo 11, so mark your calendars for July 21.

4. It will be a year of positive achievements and milestones for autonomous vehicles — Colin Beirne

Last year we saw reality deliver a setback to the autonomous vehicle industry. The general population got a glimpse of how incredibly difficult it is to marry driving and planning software with the sensors available today to produce reliable, safe driving in all conditions. We saw tech troubles, exposés on previously unreported crashes, and a general industry-wide tempering of expectations. Nevertheless, it was an important incremental year that focused on improvements in core technology, which I believe the world will start to see the benefits of in 2019. We expect to see more successful deployments in limited scenarios focused primarily on goods (e.g. grocery delivery, sidewalk robots), while pilot programs in Phoenix and Pittsburgh make credible steps toward proving the wider viability of the technology for transport of people.

5. We’ll see regulatory action against one of the large tech companies — David Winton

Elected officials will continue their attacks on the tech industry, spurred on by the general unease with these companies and a tenuous grasp of the technologies they deploy. These platforms will continue to be blamed for societal ills and simultaneously get slammed for exerting too much and too little control over their platforms and the messages promulgated through them. Sooner or later we may see a deeper regulatory approach from government focusing on how Silicon Valley conducts business, either in the form of new regulations/laws or emboldened regulators. One potential side effect is that this increased regulatory scrutiny could be the catalyst for a new wave of competitors trying to take advantage of the perceived weakness in consumer trust in the existing brands.

6. Tech-enabled and hyper-focused customer experiences separate the winners from the losers — Erhan Soyer-Osman

Rising costs associated with both reaching new customers (digital advertising) and servicing existing ones (customer support) could make it difficult for enterprises to ignore software tools that can create improved efficiency in both repeatable and scalable ways. Companies that started utilizing smart tech over the last few years, such as a customer data platform (unified customer views) or customer support platform (using AI to better manage service requests while making humans more productive), could begin outperforming competitors, creating increased opportunities for early-stage companies focused on providing these hyper-focused customer experience tools.

7. User/consumer scrutiny of companies’ policies and ethics reaches an all-time high, while reputation management becomes vitally important - Colin Beirne

In their attempt to personalize and differentiate ads and content, technology companies will continue to wrestle with where users’ privacy red lines are. Consumer scrutiny of companies’ policies and ethics could reach an all-time high, with trust taking precedence, especially as the technology to create deep fakes and deceive the public becomes more pervasive. Good, thoughtful PR management to engage with consumers may become more important than ever and could be a big driver of whether companies thrive or fail.

Thanks for taking a look through. We’d love to hear your thoughts. Let us know if you agree or disagree, and feel free to reach out directly if you’re working on a company tackling any of these trends.

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Ventures, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice. Any forward looking statements herein constitute only subjective views or beliefs, should not be relied upon, are subject to change due to a variety of factors, which are beyond the control of Two Sigma Ventures and/or its affiliates and agents.

7 Predictions and Tech Trends That We’ll Be Watching in 2019 was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

By Matt Jacobus & Erhan Soyer-Osman

Back in the day, if the master of coin wanted to transact he had to send a raven to Bravos, to have a guard enter the bank vault, and wait for a ship to return across the Narrow Sea with his silver stags and copper pennies. Fast forward to today’s cryptosphere, and things have improved only slightly. If you want to transact securely in cryptocurrency, you send a message to your custodian asking them to: physically enter a vault, collect your keys, walk them to an internet enabled device, sign your transaction and finally, send your crypto. Your electronic trustless future is neither. It is being slowed to the pace of a human worker in the (fantastical) middle ages, who you must also trust! Fortunately, now there is hopefully a better solution.

We’re happy to announce that Two Sigma Ventures is leading an $8M seed investment in Trustology, a new platform for safeguarding digital assets. Their first product, TrustVault, strives to streamline this process and authenticate using biometrics, encrypt instructions using a user device’s secure enclave, sign transactions in an infrastructure where no human can access and no computer other than your device can message, and finally write transactions to the blockchain.

At Two Sigma Ventures, we’re intimately familiar with the power of electronic and systematic approaches to finance, so we recognized Trustology’s TrustVault solution as a potential game changer for any institution or enterprise that needs immediate access to their digital assets to manage investments, risk, or just plainly to transact real-time. TrustVault also hopes to have a strong and flexible authentication solution for institutions that require multiple sign-offs, supporting strong user authentication mechanisms for m-of-n transactions, customizable smart contract design, account transacting controls, just to name a few.

Being a crypto-first, high-functioning team that is focused on security is just one aspect of what we think makes the Trustology team great. Equally important we believe, especially in the case of institutional-grade custody, is the expertise in regulatory, compliance and risk measures that can only come from deep industry experience. Alex Batlin, Founder and CEO, brings the combination of these capabilities, having formerly led technical blockchain innovation at both UBS and BNY Mellon (one of the largest custodians in the world) while also serving as a founding member of the Enterprise Ethereum Alliance and R3 consortium.

We’re excited to join ConSensys in backing the Trustology team as they set out to realize a better way for enabling digital asset security. Huge congrats to the team on this exciting milestone, and we’re looking forward to supporting them through this next phase of growth. For coverage of this new round of financing, please also take a look at Business Insider, VentureBeat, and Reuters here.

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Ventures, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

Announcing Our Investment in Trustology’s $8M Seed Round was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

By Mickey Graham and Hannah Yoo

We recently hosted our fourth annual portfolio Summit at Two Sigma Ventures. We brought together founders and key leaders from 30+ portfolio companies from seven states for a full day of celebration, collaboration, and shared learning.

Over the past few years we’ve seen venture funds take various approaches to how they structure these summits — some have headline-worthy keynote speakers, others have lots of panels, and some take trips to exotic locations. While we don’t think there’s a wrong way to build programs and resources to support portfolio companies, our team’s approach has always been slightly different. We believe in connecting founders and teams with each other and with experts (both from within the Two Sigma network of talented engineers, data scientists, and operators and outside vendors), and then getting out of the way to help facilitate open and introspective conversations.

Colin Beirne on Twitter

At @TwoSigmaVC Summit, we ask our founders to be the content. Talk to each other, in groups big and small. Here's @RapidSOS CEO Michael Martin schooling us on how startups can secure partnerships with huge tech cos.

Our conference spanned 12 workshops, breakout sessions, and lightning talks, covering our most requested support, content, and connections from throughout the year. In the afternoon, we featured a fireside chat with Dennis Mortensen of x.ai and Boris Sofman of Anki and a talk from Nir Eyal on his book Hooked: How to Build Habit-Forming Products.

We offered tailored tracks for roles in human resources, engineering & data science, and sales in order to provide a space for conversations about challenges, solutions, and goals amongst people doing similar roles across a diverse group of companies. These discussions revolved around a variety of challenges many startups face, including the ins and outs of reliability engineering, how to build and evolve a sales organization, managing a business through cash constraints and pivots, growing your data engineering function in a scalable way, best practices for measuring people analytics for the whole employee lifecycle, among others.

Jordan Wan Provides a Crash Course on How Startups Can Build a High Performing Sales Function

Jordan Wan from CloserIQ kicked off the day with an information-packed presentation about constructing and evolving sales structures. Particularly, he pointed out the do’s and don’ts in talent architecture and sales strategy.

Michael Martin from RapidSOS Discusses How to Access the Right People in the Right Way

Michael stressed the importance of internal advocates at companies where you are seeking a partnership and understanding what potential partners care the most about as well as their weak points. That way, you can make yourself indispensable by offering a solution to their problems.

Brad Wellington and Matt Greenwood of Two Sigma Host a Discussion on Effective Information Streams

Companies of all sizes struggle with calibrating the most effective methods to share different types of information streams, which can range in complexity from a quick text message (“meeting changed to eleven”) to a detailed discussion about company-wide policies. Two Sigma Engineering Manager Brad Wellington and CIO Matt Greenwood shared their extensive experiences with and insights into developing the right communication channels, identifying common trends and behaviors of teams, and figuring out the best way to keep everyone on the same page. Both stressed the importance of critically thinking about what communication tools are most effective in different contexts: “The value of one phone is 0 because it can make no calls to anyone else, so the value of a network is how many people are plugged in.”

A Toasty Fireside Chat with Dennis Mortensen and Boris Sofman on Lessons Learned Along the Way

While AI can sometimes feel like magic, the reality is that a lot of blood, sweat, and tears go into making AI products and experiences. Dennis Mortensen from x.ai and Boris Sofman from Anki discussed the impact of AI on everyday life and their experience as entrepreneurs and experimenters bringing AI products to life. They talked about the challenge of continually adapting products to address customer input and how to educate consumers. Boris and Dennis reflected on their roles as leaders at the companies as well, sharing their thoughts about the importance of managing people effectively with transparency, building trust, and helping employees see a long term career path at the company. They concluded their chat by encouraging people to go for long walks, have lunch with new hires, and actively work on company culture.

Nir Eyal Brings the Summit to a Close with What We Can Learn From Habit-Inducing Companies

Nir Eyal, the author of Hooked: How to Build Habit-Forming Products, closed the Summit with a keynote on the psychology between the human need to seek pleasure, hope and acceptance while avoiding pain, fear and rejection. He showed how many social habit-forming products design for the internal trigger to capitalize on the need to modulate moods.

Overall, it was an inspiring day and a chance to remind ourselves about what a privilege it is to be working with such data-driven teams constantly pushing the envelope. We’re already counting down the days until next year.

Bonus: A Few More of Our Favorite Photos

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Ventures, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

Highlights from the Fourth Annual Two Sigma Ventures 2018 Summit was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

By Kat Rumbles and Hannah Yoo

College to first job. Academia to startup. Large corporation to startup. Whether deciding which job offer to take coming out of school or looking for your next career move, there are seemingly endless factors and potential outcomes to consider when evaluating new career opportunities. Career transitions are rarely simple, and they can be even more complex and challenging for women moving across male-dominated industries.

To help understand how to make the most of (and ace) these transitions, our team hosted a panel with three of our portfolio companies at Grace Hopper Celebration, the world’s largest gathering of women technologists. Ann Ruble, an Operating Partner on our team at Two Sigma Ventures and a Managing Director at Two Sigma, sat down with with Raquel Ledezma-Haight of x.ai, Maureen Teyssier of Enigma Technologies, and Renee Orser of NS1 to cover everything from your first job to negotiating a new salary.

1. On Taking the First, Unknown Step

How many of you knew what you wanted to do post-graduation before you came to college? As an exercise in understanding our career trajectories, Ann Ruble kicked off the session by asking the entire room that question. Not a single person raised their hand, showing how career paths are often nonlinear with unexpected beginnings and endings.

Raquel came to school wanting to do neuroscience research, interned in marketing, finance, and as a lab assistant. She’s now a Product Manager at x.ai. Maureen planned to major in English and Chemistry and ended up completing a PhD in Computational Astrophysics. After exploring a career in academia, she decided to join a startup and is now a Technical Lead and Manager at Enigma Technologies. Renee studied International Relations and Arabic. She worked at Americorps and as a consultant before becoming the VP of Engineering at NS1.

It was only by building experience and tenacity — through part-time jobs, coffee chats, different internships, past jobs, boot camps — that they were able to land where they are now. This experimentation, which can sometimes feel like a detour, was critical in each of their career journeys and was necessary to get them where they are today. Ask questions, be curious, learn from those around you. Don’t be afraid to take a risk and try something new. Until you try it, you’ll never know if a new role or skill might end up being a perfect fit.

2. Prioritizing Your “Must-Haves”

What’s most important to you in a role and at a company? We all have unique wants and needs at work, from wanting lateral or upward mobility, to a more flexible work schedule, or the opportunity to work on a variety of projects. If you are finding that your needs aren’t being met, Maureen advises writing down what you don’t like about your current role or company. Once you have your list, look through each one and determine whether or not the solution is within your control. For example, is the company culture, the leadership team, or something else getting in the way of your success? If so, then it’s probably time to look for your next opportunity outside of the company. If you find that the things you’d like to change are ultimately within your control, it might be time to advocate a bit harder for yourself, ask to work on new projects, or take other steps towards achieving your goals.

3. Make Yourself Indispensable; Know your Worth and Negotiate

Take on responsibility. Work on teams that have different skill sets than you. Find the intersection between your career goals and what will benefit your company’s bottom line. After being at Enigma for about 6 months, Maureen realized she could take on more responsibility, which would also give her the career growth and managerial skills she was looking to develop. She approached her manager with an outline of a plan of shifting from a Senior Data Scientist to a Technical Lead and Manager, and he agreed to give it a try. From there, Maureen continued to receive more opportunities to grow as she succeeded in delivering more value back to the company.

She and the other panelists also advocated for making your work and your skill set an indispensable part of the team, so it’s difficult for the company to imagine what they would do if you left. Once you’ve become a crucial member of the team, ensure that your boss and your team know your value and take the time during performance reviews and 1:1s to negotiate for what you want and need. These discussions go best when you can align your professional goals with the company’s goals. Make your advancement a solution to the company’s challenges, rather than an additional problem the company has to solve.

4. Find a Mentor and Be a Mentor

All of our panelists stressed the importance of finding a network and mentor. Leverage your company’s network, but don’t be afraid to look elsewhere for valuable career advice. Renee suggested reaching out to a well-connected contact and asking for introductions to others who might have been through a similar challenge. Raquel also stressed the importance of leaning on others:

“In the tech and startup world, so many of our roles are really new and haven’t existed, so I find myself leaning on peers and people just above me.”

It’s not always the most beneficial to depend on the most senior leaders; rather than getting advice from people who completed the same problem 10 years ago, someone who just dug themselves out of the trenches yesterday may prove to be more helpful. Then, pay it forward: go, talk, volunteer, and become a mentor for someone else. You know how nerve wracking it can be to send that cold email, so when you get one, take a minute to answer.

5. Build a Toolbox of Skills and Bring It Wherever You Go

If you’re an engineer, it’s important to continue defining your technical skills, but also spend time sharpening soft skills. If you’ve cultivated your soft skills, then work on building a new skill set to add to your toolbox. When Renee was moving between industries, she wanted prospective employers to understand why her previous work at Americorps and in consulting was relevant for the roles she was applying for in the tech industry. To make this clear, she added a section to her resume called ‘People Skills and Programmatic Skills,’ which included programming languages as well as facilitation and program management so potential employers would make the connection.

6. Finally, enjoy the journey

In her closing remarks, Maureen shared a piece of advice she wished she could give to her former self: stop, sit down, and enjoy the company of others with no formal strategy of negotiating or networking. This is how you build real relationships; it will bring joy to your work and to your life. Finally, and perhaps most importantly, take time to reflect on what you’ve learned from each experience, including the steps in your career that felt like setbacks at the time. Relish in the lessons that happened between, and inevitably, you’ll continue learning and growing as you consider your next career transition.

Enigma Technologies is an operational data management and intelligence company. NS1 is an intelligent domain name system (DNS) and traffic-management platform that uses a data-driven architecture to make routing decisions in real-time. X.ai is an artificial-intelligence-driven personal assistant that combines machine learning and human intelligence to schedule meetings.

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Ventures, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

Taking the Leap: 6 Lessons in Making the Most out of Career Transitions was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

One of the core benefits Two Sigma Ventures provides to our portfolio companies is the deep expertise and extensive technical network of Two Sigma. With over 1,500 employees, including over 200 PhDs, Two Sigma has an extended network of engineers and data scientists that can give our companies the extra bandwidth and support in bringing applied AI and ML products to market faster. In keeping with a belief of embracing state of the art technology and challenging methodological assumptions, several Two Sigma researchers and engineers recently attended the 35th International Conference on Machine Learning (ICML 2018) in Stockholm.

ICML 2018 featured 621 excellent papers, for which we are grateful not only to the authors, but to the 160 area chairs and 1,776 reviewers who made the event possible. This year, we were impressed not only with the strong keynotes and tutorials but also with the large number of diverse workshops, thanks to co-location with International Joint Conferences on Artificial Intelligence (IJCAI) and International Conference on Autonomous Agents and Multiagent Systems (AAMAS). Enthusiasm surrounding machine learning and related topics remains palpable, and it continues to grow.

Given the wealth of high quality information on offer at ICML 2018, absorbing and evaluating everything such a comprehensive conference has to offer can be a challenge. Below, Two Sigma researchers Xiang Zhou, Travis Wang, and Steven Kim highlight several papers they found particularly novel, practical, or otherwise compelling.

Adversarial

Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples — Anish Athalye, Nicholas Carlini, David Wagner

This paper won one of the best paper awards at this year’s ICML. It is widely acknowledged that although neural networks have achieved great performance (in some tasks even surpassing human performance), they are still susceptible to various types of adversarial examples. There has been significant recent interest in constructing defenses in order to increase the robustness of neural networks. This paper introduces the notion of “obfuscated gradients” as a phenomenon that leads to a false sense of security in defenses against adversarial examples.

This paper is in some sense a review paper. It reviewed seven out of nine defenses from ICLR 2018 relying on obfuscated gradients, and then constructed successful attacks to circumvent those defenses. The authors consider three types of obfuscated gradients, either caused intentionally or unintentionally:

1. Shattered gradients are caused when a defense is non-differentiable, and can be attacked by the proposed Backward Pass Differentiable Approximation (BPDA) method, which essentially replaces the non-differentiable layer with a differentiable approximation;
2. Stochastic gradients are caused by randomized defenses, and can be attacked by Expectation Over Transformations (EOT);
3. Exploding and vanishing gradients are caused by defenses that consist of multiple iterations of neural network evaluation, and can be attacked by re-parameterization.

The paper also gives high-level guidelines about building and evaluating defenses. A good defense should make the threat model clear (i.e., the setting under which the defense is meant to work). A good defense should also make precise and testable claims. These guidelines will hopefully benefit future attack and defense researches.

Architectures

Conditional Neural Processes — Marta Garnelo, Dan Rosenbaum, Chris J. Maddison, Tiago Ramalho, David Saxton, Murray Shanahan, Yee Whye Teh, Danilo J. Rezende, S. M. Ali Eslami

Deep neural networks excel at function approximation, yet they typically require training “from scratch” for each new function. Bayesian methods, on the other hand, exploit prior knowledge to quickly infer the shape of a new function at test time. These are computationally expensive, however, and designing appropriate priors can be difficult.

This paper by DeepMind proposes a family of neural models, Conditional Neural Processes (CNPs), which are inspired by the flexibility of stochastic processes such as Gaussian processes but are structured as neural networks and trained via gradient descent. Thus CNPs combine the benefits of both deep neural networks and Bayesian methods.

The experiments conducted on a range of canonical machine learning tasks (including regression, classification, and image completion) indicate that CNPs make accurate predictions after observing only a handful of training data points, yet scale to complex functions and large datasets.

This work is a step towards learning high-level abstractions, one of the grand challenges of contemporary machine learning. Functions learned by most conventional deep learning models are tied to a specific, constrained statistical context at any stage of training. A trained CNP is more general. It encapsulates the high-level statistics of a family of functions and thus can be reused for multiple tasks. The work could help in tackling the many key machine learning problems that seem to hinge on abstraction, such as transfer learning, meta-learning, and data efficiency.

Efficient Neural Architecture Search via Parameter Sharing–Hieu Pham, Melody Y. Guan, Barret Zoph, Quoc V. Le, Jeff Dean

This paper from a group at Google Brain is built on a previous work on (Zoph & Le, 2017) for automatic model design. Despite its impressive empirical performance, NAS is both computationally expensive and time consuming.

Recent work on transfer learning and multitask learning (Razavian et al., 2014; Zoph et al., 2016; Luong et al., 2016), established that parameters learned for a particular model on a particular task can be used for other models on other tasks, with little to no modification. Encouraged by this progress, this paper proposes Efficient Neural Architecture Search (ENAS), which improves the efficiency of NAS by forcing all child models to share weights, eschewing the lengthy process of training each child model from scratch to convergence.

The paper claims that ENAS delivers strong empirical performance, while using far fewer GPU-hours than existing automatic model design approaches; specifically, ENAS is 1000x less expensive than standard NAS. This work could certainly speed up Google’s Cloud AutoML. For automatic architecture selection, many questions still remain, such as working with small datasets, or proving optimality of the search result.

Differentiable Dynamic Programming for Structured Prediction and Attention — Arthur Mensch, Mathieu Blondel

Empirical successes in deep learning and modern computational infrastructure have inspired a push for end-to-end learning, but these efforts have largely been limited to learning models with stacked layers of elementary linear algebraic operations (due to the ease of automatic differentiation). The authors of this paper observe a growing desire for modeling pipelines that can process the optimal value and/or argument of a dynamic program (in domains like sequence prediction, time-series alignment, and machine translation).

Unfortunately, the optimal value and argument of a dynamic program are typically non-differentiable, hence cannot be directly inserted as a layer of a neural network. There is a wealth of prior literature on probabilistic graphical models that instead approaches dynamic programs by moving from the so-called (max, +) semiring to the (+, x) semiring; here, the sum-product algorithm (also known as belief propagation, message-passing, etc.) is the typical computational workhorse.

In this paper, the authors observe that existing probabilistic methods can be reframed as a particular instance of a general class of max operators with strongly convex regularizers (examples include entropy, l2 norm). Elementary results from convex duality show that this regularized max is smooth and differentiable (with the gradient as the maximizing argument). The authors go further first by showing that the optimal value and argument of the regularized max are differentiable. Next, they demonstrate that their gradients are explicitly computable. Finally, they prove bounds on the distance between the solutions to the regularized max and the original dynamic program. The results in this paper serve as a substantial primer on how to reliably incorporate non-trivial dynamic programming layers into a neural network architecture.

Generative Adversarial Networks (GANs)

GAIN: Missing Data Imputation using Generative Adversarial Nets — Jinsung Yoon, James Jordon, Mihaela van der Schaar

This paper presents a novel method for missing data imputation inspired by generative adversarial networks (GANs) (Goodfellow et al., 2014). The method is built under the assumption that data is missing completely at random. If the model-imputed data is drawn from a distribution close to the underlying distribution of the observed original data, it should not be possible to discriminate between the imputed data and the original data.

Different from the standard GAN architecture, which uses a generator to create fake samples and a discriminator to distinguish between fake and real samples, the new architecture, generative adversarial imputation network (GAIN), uses a generator to impute the missing component for samples, and a discriminator to predict which components of a sample is fake. A hint mechanism is introduced in order to reveal to the discriminator partial information about the missingness of the original sample, which the discriminator uses to focus its attention on the imputation quality of particular components. This hint ensures that the generator does in fact learn to generate according to the true data distribution.

Experiments reported in the paper show that GAIN significantly outperforms state-of-art data imputation methods. However, as noted above, the test is based on the assumption that data is missing completely at random, so an open question is how GAIN-based imputation will behave when the completely at-random assumption is not true.

A Two-Step Computation of the Exact GAN Wasserstein Distance — Huidong Liu, Xianfeng David Gu, Dimitris Samaras

Generative Adversarial Networks (GANs) (Goodfellow et al., 2014) have received tremendous attention in recent years. The idea is to generate samples by deep networks that cannot be distinguished from the real samples. However, previous researches have noted that the original GAN architecture suffers from the issue of vanishing gradients, training instability, and mode collapse. The invention of Wasserstein GANs (WGANs) (Arjovsky et al., 2017) was a breakthrough to motivate researches into computing the appropriate distance metric.

While largely successful in solving the above issues, the WGAN architecture still requires a technical nuance in order to be able to use a deep network to approximate a Lipchitz function class. Specifically, a WGAN needs to clip its network weights in every iteration in order to guarantee the boundedness of the neural network function. However, weight-clipping limits the functional space of the WGAN “critic” and can also cause gradients to explode or vanish if the clipping parameters are not carefully chosen.

To mitigate these problems, this paper proposes a two-step method to compute the Wasserstein distance in WGAN. In the first step, a linear programming problem is formed to compute the exact Wasserstein distance, without the requirement of setting a weight clip parameter. In the second step, the authors apply a deep neural network to approximate the optimal solution of the first step by a differentiable trainable function. The new algorithm is driven by a new formulation of Wasserstein distance based on the discrete Monge-Kantorovich duality. Results on both synthetic data and real data show that the proposed method computes the Wasserstein distance both faster and more accurately.

Reinforcement Learning

SBEED: Convergent Reinforcement Learning with Nonlinear Function Approximation — Bo Dai, Albert Shaw, Lihong Li, Lin Xiao, Niao He, Zhen Liu, Jianshu Chen, Le Song

Reinforcement learning (RL) continued to be a hot topic at ICML 2018. Recent years have witnessed the successful application of reinforcement learning in many real-world challenges, including Atari games, AlphaGo, and much more. Without a doubt, this is a result of a deepening understanding of the RL framework and important progress in developing RL algorithms.

In this paper, the authors first revisit the Bellman equation (Puterman, 2014), which is central in RL problems, and reformulate it into a primal-dual optimization problem using the Nesterov smoothing technique and the Legendre-Fenchel transformation. Then a new algorithm, called Smoothed Bellman Error Embedding (SBEED), is proposed to solve this optimization problem, based on stochastic mirror descent (Nemirovski et al., 2009). The algorithm allows for any differentiable function class, and a theoretical guarantee is provided for general nonlinear function approximation (the first, to the best of our knowledge).

The authors decompose the estimation error of value function into four sources: optimization error from mirror descent, generalization error from using finite samples, bias induced by the Nesterov smoothing, and approximation error induced by using function approximation. In addition, the proposed algorithm is closely related to other successful algorithms such as trust region policy optimization (TRPO) (Schulman et al., 2015) and natural policy gradient (Kakade, 2002).

Theory and Optimization

On the Optimization of Deep Networks: Implicit Acceleration by Overparameterization — Sanjeev Arora, Nadav Cohen, Elad Hazan

Traditional theory implores caution when increasing model complexity, whether it be through generalization bounds in learning theory or variance bounds in statistics, both of which largely depend on bounding dimension (for a definition of “dimension” that suitably encodes expressiveness). Nonetheless, practitioners often find that increasing model complexity improves test error. Surprisingly, this can be the case even when the practitioner is fitting a model from a perfectly specified class of models. That is, suppose one has a synthetic dataset of labels generated by applying a particular depth-D, width-W, fully-connected network to random input data; even in the setting of known D and W, one can often attain better test error by optimizing over networks in the more complex class of depth D+1 and width W+1.

In this work, the authors attempt to shine some theoretical understanding on this practical phenomenon, by quantitatively demonstrating how increased model complexity can ease model training. Note that past work has suggested that some machine learning models may suffer from underfitting due to insufficient optimization, hence accelerated optimization procedures can yield not only faster training but also better model performance. This work indicates that increasing model complexity (termed “overparameterization” here) is itself a form of acceleration. Specifically, gradient descent on an overparameterized network induces an “extra” term in each gradient step that amounts to acceleration. Moreover, this paper provides a proof that this extra acceleration term cannot be attained by any sort of weight-based regularization of the loss function, as the line integral (with respect to weights) of the gradient of the overparameterized network does not vanish.

From a theoretical perspective, this paper offers one more piece of evidence that model formulation and model optimization are crucially linked, rather than being distinct components of data analysis. For the practitioner, the upshot is that model complexity can be embraced, not only for the statistical benefit of expressiveness, but also for the computational benefit of improved trainability.

Dynamical Isometry and a Mean Field Theory of CNNs: How to Train 10,000-Layer Vanilla Convolutional Neural Networks — Lechao Xiao, Yasaman Bahri, Jascha Sohl-Dickstein, Samuel S. Schoenholz, Jeffrey Pennington

Deep architectures have allowed increasingly sophisticated modeling in diverse domains, but practical data analysts must often wrestle with the numerical pathologies that arise during training of deep models. Specifically, as depth increases, gradients may vanish or explode, a computational instability that the authors characterize in terms of the well-conditionedness of the input-output Jacobian of the network. Traditionally, practitioners will combat obstacles to training by employing architectural “tricks” like residual/skip connections or batch normalization; while these “tricks” can be effective, it is not clear whether they are crucial from a computational perspective (by allowing for reasonably fast training) or from a statistical perspective (by enforcing desirable constraints on the model class).

Prior work in the fully-connected setting (Pennington, Schoenholz and Ganguli, 2017) has analyzed the input-output Jacobian by considering the layer-to-layer transformations of a network in “mean-field” limit (i.e., width to infinity), and then interpreting the fixed point of these dynamics. It was shown that well-posedness of the input-output Jacobian can be attained by delicately constructing initial weights to be orthogonal (in a suitable sense), which induces norm-preserving (hence the term “dynamical isometry”) transformations in each layer. In the fully-connected setting, orthogonal weights can simply be constructed via the Gram-Schmidt process. Note that the desired condition here is on initialization; this differs from other work on orthogonal weights involving penalties on weights in order to induce orthogonality throughout training (e.g., Balestriero and Baraniuk, 2018).

This paper offers a substantial extension of the concept of orthogonal initialization to the setting of convolutional neural networks (CNNs), where orthogonal initial weights are constructed through a non-trivial (but explicitly detailed) scheme drawn from wavelets theory. Empirically, the authors’ orthogonal initialization allows for training of very deep CNNs, without any architectural tricks. This work (along with related work on recurrent neural networks due to Chen, Pennington, and Schoenholz, 2018) further distills the effect of dynamical isometry during training of networks, and ultimately offers some concrete guidance for practitioners: better initialization induces faster training!

Not All Samples Are Created Equal: Deep Learning with Importance Sampling — Angelos Katharopoulos, François Fleuret

This paper presents an efficient algorithm for accelerating the training of deep neural networks using importance sampling.

Previous works (Zhao & Zhang, 2015; Needell et al., 2014) have shown the optimal distribution (to sample from) is proportional to the per-sample gradient norm. However, computing this distribution is very expensive.

This paper derives a tractable upper bound to the per-sample gradient norm, as well as an estimate of the variance reduction achieved with importance sampling. Doing so enables the importance-sampling mechanism to be “switched on” when it will result in an actual speedup. This work can result in reduced computational requirements of more than an order of magnitude compared to a recent work (Alain et al., 2015).

Experiments reported in this paper show that the algorithm is effective in reducing the training time for several tasks, both on image and sequence data. The results also show that not all data points matter equally in the duration of training, a fact that can be exploited to gain a computational speedup or better quality gradients, or both. However, the image data used in the experiments has a high signal-to-noise ratio, so it would be interesting to see how much importance-sampling can help for datasets with greater levels of noise.

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Ventures, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

ICML 2018: Our Favorite Papers on Deep Learning, GANs, Optimization, and more was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

Announcing Our Investment in Comet.ml’s $2.3M Seed Round

By Matt Jacobus

The probability that two data scientists use the same tools is nearly zero. They like what they like, and across a team they might use every tool in the book: open-source analytics engines, various languages and frameworks, and many editors and IDEs. It can be difficult to get them to give up their favorite workflows, and it can be an even bigger burden to ask them to adopt an entirely new heavier-weight platform for model development and collaboration.

Just like everyone else, data scientists want to use the tools that make them the most productive. The problem is, as an organization scales, it’s unrealistic to have developers and data scientists across machine learning teams using homebrewed solutions and reverting to email to manage complex machine workflows, datasets, and production models.

While some organizations may get great mileage out of tightly integrated workflow solutions and standalone platforms, we’ve seen that in other cases machine learning teams just need better tools that match their preferred workflows — not entirely new ones. Some data scientists and developers love the flexibility of different languages, tools, and workflows, which is why we were so excited when we first met Gideon, Nimrod, and the team at Comet.ml.

Comet.ml is the first infrastructure- and workflow-agnostic machine learning platform. With 15 seconds and one line of code, Comet.ml seamlessly integrates into sets of disparate tools and libraries, allowing teams to quickly start tracking their experiments and optimizing their models without compromising on the flexibility and fluidity of their preferred workflows.

Every time you run an experiment, Comet snapshots your model code, parameters, code, configuration, and tracks it, allowing data scientists to visualize and organize their work at virtually no cost. Comet is dedicated to automating large parts of the machine-learning process, from tracking and comparing experiments to collaboration across teams to optimizing hyperparameters — with no heavyweight platform or workflow changes required.

When our team saw that adding a single line of code could give us real-time metrics and charts detailing how our models converged while saving our experiments, we realized how quickly Comet could accelerate the modeling process and catch on. This technology, and the amazing team that is building it, is why we’re excited to announce we’re investing in Comet.ml’s $2.3M seed round alongside Trilogy Ventures, Founders Co-Op, Fathom Capital, Techstars Ventures, and other angel investors.

See more coverage in TechCrunch, and make sure to share on Twitter!

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Investments, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

Meet Comet.ml, a New Way to Supercharge Machine Learning was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

By Dusan Perovic

We’re excited to announce today that Two Sigma Ventures is joining Euclidean Capital, Lux Capital, The Column Group, Polaris Partners, Illumina Ventures, and other investors in Kallyope’s $66M Series B.

When we first met Nancy, Rob, Ian, Justin, and the Kallyope team, we were hooked. Founded by Charles Zuker, Ph.D., Lasker Award winner Tom Maniatis, Ph.D., and Nobel laureate Richard Axel, M.D., Kallyope marks what we believe to be one of the most important trends of the past decade, and one we’ve been passionate about investing in over the past few years: the intersection of data science with biology and genomics.

Kallyope is pioneering a new approach to drug discovery by targeting and mapping the gut-brain axis. In only two years since launching, we believe that Kallyope has emerged as one of the most promising companies in New York City’s growing biotech community. With over 45 employees and growing, they’ve developed a state-of-the-art platform that uses single-cell sequencing, computational biology, and circuit-mapping technologies to develop new classes of drugs targeting metabolic and neurological disorders.

Kallyope is capitalizing on advances in a number of technologies: 1) new sequencing techniques that make it possible and affordable to conduct high-throughput single-cell RNA profiling; 2) the decrease in cloud computing costs that allow Kallyope to apply computational biology approaches at scale in a cost-efficient manner; and 3) advances in biotechnologies, such as availability of genetic models, organoids, and optogenetics, that make it possible for Kallyope to test their circuit-mapping hypothesis and quickly iterate on early drug targets.

Kallyope’s goal of mapping the gut-brain circuits fits perfectly with Two Sigma’s thesis that biology will increasingly be elucidated using computational approaches and insights from novel datasets, and participating in their Series B posed a great opportunity for us to make an investment in what we believe to be a highly promising space that has the possibility to fulfill massive unmet needs and improve human health.

We look forward to harnessing the resources of Two Sigma, including the expertise of our hundreds of data scientists and quantitative engineers, to help support Kallyope through this next phase of growth as they continue identifying new compounds and work on bringing them closer to the clinic.

See more coverage here and make sure to share on Twitter!

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Investments, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

Kallyope Announces a $66M Series B to Develop Drugs Targeting Gut-Brain Circuits was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.

By Kathryn Rumbles

At Two Sigma Ventures, we believe inspiring young people to pursue careers in tech is one of the most important things we can do. In keeping with that belief, this winter we welcomed three CUNY students to our team through the Women in Technology in New York (WiTNY) program, a partnership between Cornell Tech and CUNY universities. The goal of the program is to encourage more women to join tech companies, and to provide them with valuable work experience to prepare them for careers in the tech industry.

Our three winterns, Julia, Ji, and Tasnia, built a marketing analytics dashboard to help our team measure our digital web traffic since the launch of our new website in October. Thanks to their hard work, the dashboard is now live in our office. It displays a stream of portfolio company news, website and social media analytics, and even our team’s top influencers, which has created some friendly competition amongst our team.

After talking to Julia, Ji, and Tasnia about their experience, we wanted to share their thoughts on getting into tech, what their experience working at Two Sigma Ventures was like, and what their plans are next.

How did you first get involved in technology?

Julia: Oddly enough, Wikipedia! I started writing articles about music groups back in 2009. To write for Wikipedia, you needed to use HTML, so that was the first programing language I learned. After that, I was hooked. I went on to learn about front-end development, CSS, and JavaScript.

Ji: When I was studying economics, I started collecting large amounts of data and turning it into usable information by using a data sorting program called Gratl. The data’s ability to explain trends and future events fascinated me.

How did you get involved with Two Sigma Ventures?

Julia: During the application process for the WiTNY program, I expressed interests in working in an environment where everyone — not only upper management — has a voice, and where team members motivate each other to feel a sense of purpose. I conveyed my passion for technology and its potential for solving problems our society faces, such as poverty, education and healthcare. These responses linked me to Two Sigma Ventures, where I became immersed in a work environment and philosophy that matched every expectation I had for a tech company.

What was your favorite part about working here?

Julia: I loved how Two Sigma Ventures has the startup feel and mentality, but is equipped with the resources of a large organization. Working in Two Sigma Ventures’ open workplace, sitting right next to some great startups in the incubator gave me an opportunity to immerse myself in the tech-work culture. Two Sigma Ventures not only provides valuable resources to their teams, but creates an environment where people can collaborate and thrive.

Tasnia: One of the best parts about working at Two Sigma Ventures was the inclusivity. Two Sigma offers three employee resource groups that help give those in different minorities a voice. ouTSigma, BE@TS, and W@TS are run by full-time Two Sigma employees who volunteer their free time to organize events and encourage communication. From attracting more LGBTQ-A talent into the tech world by offering them ways to hone their skills and develop their talents to creating a racially and ethnically diverse environment and supporting women in furthering their careers at Two Sigma, these groups are dedicated to improving inclusivity. They provide everyone with a familiar face to look up to and aspire to emulate. I personally found all three resource groups to be extremely informative and helpful.

What are your plans next?

Ji: My priority will be to graduate with an engineering degree in two years. During my upcoming breaks, I will look for opportunities to spend more time working in tech companies, so that I can narrow down my field of interest. I will try to read and meet people across different fields to learn different perspectives and gain career advice. And while plans will certainly change throughout my experiences, one thing that I am sure of is that I will keep questioning myself and never stop learning.

Julia: This spring, I will be starting my third year in college. Besides taking my regular college classes, I plan on doing some fun tech and art projects. I’ve always liked using technology in artistic ways (drawing CSS images is one of my favorite pastimes). The first thing I want to learn after the internship is how to draw SVG images. Additionally, I plan on using my digital drawing skills on projects which will display data in engaging and beautiful ways, as I’m very interested in data analysis and visual representation of data.

Tasnia: My main goal is to master HTML, JavaScript and CSS, then move on to some backend languages. I also intend to reconfigure my educational plans. I came to this internship as a finance major with an interest in technology, but I am leaving as someone who would potentially like to pursue computer science to a higher level. I intend to search for more internships at tech companies as I really enjoyed the tech company culture at Two Sigma Ventures and hope to obtain a career somewhere with a similar environment.

Thanks, Julia, Ji, and Tasnia! We can’t wait to see what you do next!

The views expressed herein are solely the views of the author(s) and are not necessarily the views of Two Sigma Investments, LP or any of its affiliates. They are not intended to provide, and should not be relied upon for, investment advice.

Thank You to Our Three Interns from Women in Technology in New York (WiTNY) was originally published in Two Sigma Ventures on Medium, where people are continuing the conversation by highlighting and responding to this story.