The Fifth Elephant 2019

Gathering of 1000+ practitioners from the data ecosystem

Tickets

Let's dope it: Interoperable ML via Deep Learning

Submitted by Soma Dhavala (@dhavala) on Saturday, 13 April 2019


Preview video

Session type: Full talk of 40 mins

Abstract

One of the biggest hurdles to reducing time-to-market of an ML Product is the two language problem. Generaly speaking, the tech stacks of the Producers of the ML models and its Consumers are different. Say, a DataScientist may work with Python, but a Production Engineer may want it in a JVM language. There are multiple approaches to solving this problem. Languages like Julia offer the expressiveness of high level languages and performnance of lowe level languages like C. Distributed computing framewokrs like Spark provide a rich set of APIs in multiple languages. There were also attempts to express Models in a markup language (PMML). But one things that is certain is that entropy of the developer world is always on the rise. In such cases, how can ML made interporable and yet let the developer bring his/her own knowledge without having to re-learn new unified stacks?

In the last few years, Deep Learning has made great strides due to variety of reasons: availablity of cheap data, computing power and technology and the science itself. It is notable that industry leaders in this area have made their deep learning frameworks accessible by open sourcing them. There are also standards such as as ONNX and NNEF which can make Deep Learning interporable. That is, a Deep Learning model created in MXTNet on a laptop can be scored using, say, TensorFlow Server on a TPU in the cloud. But what about the mainstream, non-deep learning ML algorithms such as Generalized Lienar Models, Decision Trees, Colloborative Filtering algorithms. Can we, some how exploit the Deep Learning frameworks, to make them interoprable as well.

We answer this question affirmitively: It is based on the view point that 1) Deep Learning has a lego block architecture, and is object oriented in spirit and 2) Many classical algorithms can be composed using those lego blocks. It means that, algorithm’s intent is specified via any standard ML modeling libraries like sciki-learn and fullfill that intent via a backend Deep Learning modelm without any radical changes in the developer experience.

We will demo how algorithms in python’s scikit-learn, trained on a laptop, can be scored in a browser. All this by just adding one line of code. That tool is called – IMLY: Interporable Machine Learning: Yay!

Outline

  • interoporabilty in ML: challenges and current solutions
  • emergence of deep learning technologies and standardization efforts
  • deep learning: a lego block architecture for composing algorithms/models
  • a tour of equivalence between classical/popular ML algorithms and their deep learnign counter parts
  • demo of an open source project IMLY: Interporable Machine Learning: Yay!

Speaker bio

Soma S Dhavala is freelance consultant operating at the interface of Statistics, Machine Learning, Computing, and Internet of Things. His interests are in Graphs, Meta Machine Learning and their application in representing, and reasoning with information. Since last few years, he has been working on ML-as-Infrastructure.

Soma currently consults for Framewirk, is a member of the Design Council at EkStep. Since 2014 and is leading the efforts in designing Machine Learning Infrastructure and developing various DataProducts concerned with Learning such as Recommendation Engines, Auto-Tagging Content. He designed and ran a Deep Learning and Natural Language Processing immersive bootcamp for NIIT. He is also working on Deep Variational Inference with his collaborators in the academia. He is also a co-founder of VitalTicks Pvt Ltd, a start-up in the digital health care space.

In the past, Soma worked with Dow AgroSciences in the area of ML applications in Systems Biology and with General Electric Global Research Center in the area of Information Theory and Parallel Computing.

Soma obtained his Ph.D (TAMU, 2010) in Statistics where he worked on applying Bayesian Nonparametrics to systems biology (2010), has a Masters (IIT-M, 2000) in EE and a B.E (SRKREC-Andhra University, 1997) in ECE. He also worked as a post-doc close to a year in the area of Dynamical Systems at TAMU. He has over 10+ publications and multiple patents.

Links

Slides

https://docs.google.com/presentation/d/1X5W1BcRwR_RId4m-uZcxjoAoWosv2fWO46zkdPvPeFg/edit?ts=5ce1e958

Preview video

https://www.youtube.com/watch?v=p2YEZx8mjf4&feature=youtu.be

Comments

  • Zainab Bawa (@zainabbawa) Reviewer 4 months ago

    This is an interesting proposal, Soma.

    We need slides explaining:

    1. The problem statement (which is clear in your abstract, and can be mapped straightforward onto the slides).
    2. Why is this approach better than approaches for solving the problem?
    3. Deep dive in the implementation details.
    4. Why and when to pick this approach? Can organizations with running ML models and pipelines adopt this approach? What cautions should participants exercise apart from the optimism?
    5. Case studies and examples of real-life implementations that have been done using this approach.
    6. Pros and cons of the approach.

    Upload your slides with the above details, by or before 21 May. You also have to upload a two-minute preview video which is an elevator pitch explaining what your talk is about and why participants should attend it.

  • Soma Dhavala (@dhavala) Proposer 4 months ago
  • Abhishek Balaji (@booleanbalaji) Reviewer 3 months ago

    Hey Soma,

    We evaluated the proposal based on the slides. Here’s the feedback:

    • The proposal is very presriptive - talks about what one should do, instead of how should one approach a problem
    • This will need some work in restructuring the presentation to add a case study to explain the different concepts here
    • Without a case study, this talk would not be useful for someone who’s looking for solutions in their own work

    We need to hear back from you on how you want to proceeed by Jun 22, 2019.

  • Soma Dhavala (@dhavala) Proposer 3 months ago

    Abhishek,

    I recognize a problem faced in the industry, and developing a solution to address it. Few ways to solve, and the specific approach we have taken are presented, and expanded later on. I believe that we should have alternatives when it comes to technology and its practice. Having said that, we have to take “a” stance to solve any problem. Ultimately, usability truimphs. Community is not going accept “a” way doing thisn becasue, someone “presribed”. So I am not sure where “prescriptiveness” is coming into play.

    Set aside the prescriptiveness part for a moment, if it is all about presentation. I believe that audience will benefit by looking at Deep Learning as a technology from a very different view point. As it stands today, Deep Leanring and Machine Learning are seen as two different worlds. It does not have to be. Underneath building blocks are the same – is the main point of the idea. We are exploiting it to make ML interoprable by riding the Deep Learning Tech wave. Microsoft also released a new tool, along the same direction, except with some differences. Those differences are subtle but profound. I am seeking a venue to present them.

    A case-study is planned as a part of the demo. In the demo, we will show the real two language problem, and how producers and consumers can be isolated from each other. At this point in time, we are introducing Decition Tree layers, and LDA losses, and some refacttoring is planned before the Conference. Repo are public and anybody can take a look them.

    Given this, I dont think it is necessary to change the story line dramatically. We might certainly revise, though.

    thanks,
    -soma

    • Abhishek Balaji (@booleanbalaji) Reviewer 3 months ago

      Thanks for your responses, Soma.

      • We’re not asking to refrain from suggesting an approach or a solution, but ask that you compare it to other solutions available.
      • We ask to do this since not everyone in the audience will share the same experience, knowledge and journey that you have gone through.
      • I mentioned prescriptiveness since the presentation comes off as suggesting that only one approach would work. While this might be true for your use-case or work, someone in the audience might not be able to relate to your problem.

      Having said that, the proposal in its current form would not be useful for the audience and hence we cannot consider it for The Fifth Elephant. I’ll be parking this proposal for an event around the year, where the right audience would benefit from this approach. Do let me know how you watn to proceed. The deadline of 22 June still holds for this.

Login with Twitter or Google to leave a comment