AI research & applications from Google Brain, Netflix, Amazon & more: What did you miss in Boston today?
"If you work in machine learning today, you almost certainly work in Python, but that doesn't mean it should end there. The web world has a lot to offer for ML, and with Tensorflow.js, we are building a new ecosystem of Machine Learning in browser." - Daniel Smilkov, Google Brain
"Human +Machine> Human/Machine" - Amir Tahmasebi, Philips Cambridge Innovation
Today saw RE•WORK return to Boston for the annual Deep Learning Summit and Deep Learning in Healthcare Summit for a day full of presentations, workshops, interviews and networking with leading minds from the likes of Google Brain, Netflix, GSK and more.
The day kicked off with eager attendees registering and grabbing breakfast with plenty of experts sharing their thoughts on the current landscape of DL, as well as what they’re hoping to get out of the two days here in Boston:
‘We’re most excited about collaborating with like-minded individuals to bridge the gap between healthcare and technology’. Brandon Canter, Prognos.ai
‘I’m looking forward to learning more about the approaches used by experts in deep learning across different verticals in the industry and in academia. RE•WORK Summit brings together Pioneers and experts in this field. I am really looking forward to learn from each one of them. I learned so much last year and I am pumped to learn more!’ Anisha Baidya, Research Assistant, Harvard Medical School
‘We’re going to be announcing that we have BETA running of our new suite of told, ao we’re looking forward to sharing it with attendees and getting feedback, and hopefully some BETA users!’ Diana Yuan, Indico
On the deep learning track this morning’s sessions started with Ziming Zhang, Research Scientist at MERL sharing his work on bringing deep learning to global optimality in deep learning via branch and pruning. He explained to the packed room that they are proposing a novel approximation algorithm, the BPGrad, towards optimising deep models globally via branch and pruning, based on the assumption of Lipschitz continuity in D. ‘We have to keep pruning until we find a global solution. We can use branching in Deep Learning with the Lipschitz Continuity to find the difference between the two functions and therefore the utmost solution.’
Madison May @pragmaticml First up -- Ziming Zhang from @MitsubishiHVAC discussing "BPGrad -- Towards Global Optimality in Deep Learning via Branching and Pruning"! //arxiv.org/abs/1711.06959 #reworkDL
Over on the healthcare track we began the day with an introduction to how deep learning can be used in the medical industry and started by learning about machine learning in radiology. Amir Tahmasebi, Senior Research Scientist at Philips Cambridge Innovation Labs shared the successes of machine learning in healthcare over recent years in both clinical and technical aspects. ‘There are so many time consuming and menial tasks in Healthcare, and easily we can automate such laborious tasks. At Philips we’re providing a way to bring information from different domains together and also fuse and correlate them to help interpret information cross domain.’
Ankit Gupta @gankit If AI algorithms in health care do not take legal liability, they just add cost to the system and might be only useful in the developing world. My take on the intro session at #reworkhealth
We next heard from Nikhil Thorat and Daniel Smilkov, Software Engineers at Google Brain who shared their work and the recent launch of Learning.js. ‘We announced a new ecosystem of libraries and tools which have 3 major use cases: to model directly in a browser, to import pre-trained models for inference, and also import retained models. There’s the core API and the layer API, and we have automatic tools that allow you to take pre-trained models to automatically run in the browser’. The duo showed us a video demonstration on how they trained their model on 400 images to work in a game like scavenger hunt. This was done on a web browser and has huge potential for real-world applications from industries such as healthcare, to gaming.
As the sessions continued, the RE•WORK team were busy behind the scenes recording expert interviews with some of the speakers. We were fortunate enough to sit down with Anoop Deoras, Lead Researcher at Netflix who was explaining how Netflix are using both shallow and latent models for search recommendation. Anoop is currently working at the intersection of deep learning, natural language processing and recommender systems and shared with us how they relate to traditional collaborative filtering techniques. The interview will be available online post-event, and you can view the videos here.
Other presentation highlights:
Michael Sollami, Salesforce started off by demonstrating NVIDIA’s example of image generation of fake celebrities. We also looked at neural networks and imaging technology and translation. Michael showed an example of a picture of his cat to show how we can segment images with Deep Learning and style transfer. ‘What is computational photography? It’s rich and multifunctional.’
@KDnuggets: Amazing progress in computational photography with #DeepLearning. By next year people or algorithms will not be able to determine if a photo or video was real or generated. #AI #reworkdl @msollami
Chen Lin from Boston Children’s Hospital on the Healthcare track demonstrated how they’re using clinical NLP to process and recognise all the entities in a sentence and identify their semantic types and translate them into useful information. ‘We want to build a temporal model to identify the key points in a sentence in relation to the time stamps.’
@HealthcareWen In medicine, TEMPORALITY/timing important (e.g. long term effects of meds, inferring causes, research treatments), but not easy w/ EHR data because unstructured text with jargon & billing codes sparse. Chen Lin @BostonChildrens using #DeepLearning & #GPU's.
Throughout the day we enjoyed chatting with attendees and hearing about their highlights:
“I think it’s a great event where people come together and share ideas, especially in Healthcare, which is something very new.” Nachiketh Prabhakar, Student, Case Western Reserve University
“The main reason I’m here is because I want to learn more about how Machine Learning has been developed, particularly in Healthcare. I’m enjoying meeting people and gaining more understanding about which specific area I would like to persue” Boyang Fu, Rutgers
’It’s Fascinating to see how DL can be applied across such a wide spectrum. I just saw Cory McLean’s talk on deep variants, and watching CNNs be applied to genome processing is amazing.’ Elliot Swart, CTO & Liz Asai, CEO, 3derm
‘I really enjoyed the emoji scavenger hunt demo presented by Google Brain. I think it showcases the power of inbrowser computing and giving access to the end user’ - Mitesh Gadgil, Associate Data Scientist, ID Analytics
‘This is the best organised conference I’ve been to. Great job RE•WORK every time. And the food is fantastic!’ Gregory Patinskey, KDnuggets
As well as video interviews, we also recorded several episodes of the Women in AI Podcast and today spoke to Sergul Aydore, Machine Learning Scientist from Amazon who’s currently working on forecast recommendation. We spoke about her current work as well as encouraging more women into the field. Sergul explained that in order to encourage more diversity in the AI, it’s important to have role models from lots of different backgrounds.
Alongside the presentations, interviews and networking, were interactive workshops and panel discussions. We first heard from Neil Tenenholtz, Director of Machine Learning at the MGH & BWH Center for Clinical Data Science who spoke about Distributed Tensorflow: Scaling Model Training to Multiple GPUs. Neil explained that ‘thousands and thousands of GPUs are trying to communicate with a parameter server, but this gets too much, so what can you do? Get another. But then the question is how many do you want? You can keep getting more, can we do better? Fundamental to the solution is the ring all reduce where you say ‘ok I have this piece of data, let me send it to the person next to me’ if you create enough loops you get a full update. You can’t be more efficient than this.’ Neil walked through the code line by line to explain how this works and took several questions from attendees throughout the workshop.
Next up was a panel discussion looking into The ROI of Deep Learning Application - Is it Worth it? “If you try and do a project with people who don’t have Deep Learning expertise it’s going to cost a lot more, so the ROI is really about bringing up people’s capabilities. This is the way our company is going, the way the future is going, we need to have these capabilities. We need to be relevant, not like dinosaurs” Zac Kriegman, Thomson Reuters Innovation Lab. Throughout the session, we covered so many more questions: What types of applications do benefit and which ones not so much? What factors drive development timelines when it comes to deep learning? How does an increase in algorithm performance realistically translate into a gain in return-on-investment?
Madison May @pragmaticml Perhaps even more important than today's technical sessions @reworkdl is this evenings panel on the intersection of AI and ethics w/ @g_fariello @ccanso @kathytpham @sim0nmueller
Back in the deep learning room, the ethics discussion kicked off with our four experts introducing themselves. We heard from Gabriele Fariello from Harvard University, Cansu Canca from AI Ethics Lab, Kathy Pham from Harvard Berkman Klein Center, and Simon Mueller from the Future Society. There was a lot of discussion about regulations such as GDPR, limiting sensitive data, and ensuring that ethics and development are intertwined throughout the whole process of AI. Gabriel explained that 'ethics needs to be woven into the foundations of STEM.' Simon asked Kathy whether she thought that ethics is preventing progressions of AI? ‘No, it's not a watercolour conversation. It's not like we're talking about how to scale something and people say 'but ethics!' or even on a softer level as we're talking about launching something, it's almost like the same as the tension between users and product, but ethics isn't even close to being there yet, so I don't think we need to worry about that.’ Cansu continued by explaining that there are a few things being done to help infuse ethics, ‘this starts with policy and GDPR. It’s all about having a proactive approach on policy level. That's more important than playing catch up. Policy tries to follow, but now in companies there's more interest with DeepMind and Microsoft working on ethics etc. But it's always that they're separate from the developers, it's always on the side that there are 'ethics people'. I think what should happen is ethics should be integrated from the start.’
With plenty of food for thought, we’re tying up this evening with networking, drinks and more interviews with our speakers in advance of day 2 tomorrow. Keep up to date with tomorrow’s activities by following us on Twitter at @reworkDL and @reworkHEALTH.