Synthesizing audio for specific domains has many practical applications in creative sound design for music and film. But the application is not restricted to entertainment industry. We propose an architecture that will convert audio (human voice) to the voice owner’s preferred image – for the time being we restrict the intended images to two domains – Object Design and Human body. Many times, human beings are unable to describe a design (may be power-point presentation or interior decoration of a house) or a known person by verbally described attributes as they are able to visualise the same design or the person. But the other person, the listener may be unable to interpret the object or human descriptions from the speaker’s verbal descriptions as he/she is not visualising the same. Complete communication thus needs much of a trial and error and overall hazardous and time consuming. Examples of such situations are 1) While making presentation, an executive or manager can visualise something and an express to his/her employee to make the same. But, making the best slides from manger’s description may not be proper. Another relevant example is that a house owner or office owner wants his/her premises to have certain design which he/she can visualise and express to the concerned vendor. But the vendor may not be able to produce the same. Also, trial and error in this case is highly expensive. Having an automated Image, recommended to him/her can address this problem. 2) Verbal description of a terrorist or criminal suspect (facial description and/or attribute) may not be always available to all the security people every time, in Airports or Railway Stations or sensitive areas. Presence of a software system having Machine Generated Image with Ranked Recommendation for such suspect can immediately point to one or very few people in a crowded Airport or even Railway Station or any such sensitive place. Security agencies can then frisk only those people or match their attributes with existing database. This can avoid hazardous manual checking of every people in the same process and can help the security agencies to do adequate checking for those recommended individuals.

We can use a Sequential Architecture consisting of simple NLP and more complex Deep Learning algorithms primarily based on Generative Adversarial Network (GAN) and Neural Personalised Ranking (NPR) to help the object designers and security personnel for serving their specific purposes.

The idea to combat the problem:

I propose a combination of Deep Learning and Recommender System approach to tackle this problem. Architecture of the Solution model consists of 4 major Components – 1) Speech to Text

2) Text Classification into Person or Design; 3) Text to Image Formation; 4) Recommender System

We are trying to address these four steps in consecutive applications of effective Machine Learning and Deep Learning Algorithms. Deep Learning community has already been able to make significant progress in terms of Text to Image generation and also in Ranking based Recommender System

Brief Details about the four major pillars of this problem:

Deep Learning based Speech Recognition – Primary technique for Speech to text could be Baidu’s DeepSpeech for which a Tensorflow implementation is readily available. Also, Google Cloud Speech-to-Text enables the develop to convert Voice to text. Voice of the user needs to be converted in .wav file. Our steps for Deep-Speech-2 are like this – Fixing GPU memory, Adding Batch normalization to RNN, implement row Convolution layer and generate text.

Nowadays, we have quite a few free Speech to Text software, e.g. Google Docs Voice typing, windows Speech Recognition, Speech-notes etc.

Text Classification of Content – This is needed to classify the converted text into two classes – a) Design Description or b) Human Attribute Description because these two applications and therefore image types are different. This may be Statistically easier part, but its importance is immense. A Dictionary of words related to Designs and Personal Attributes can be built using online available resources. Then, a supervised algorithm using tf-idf and Latent Semantic Analysis (LSA) should be able to classify the text into two classes – Object and Person. These are very much traditional and proven techniques in many NLP research

Text to Image Formation – This is our main component for this proposal. Today, one of the most challenging problems in the world of Computer Vision is synthesizing high-quality images from text descriptions. In recent years, GANs have been found to generate good results. Samples generated by existing text-to-image approaches can roughly reflect the meaning of the given descriptions, but they fail to contain necessary details and vivid object parts. There have been a few approaches to address this problem, all using GAN. One of those is given as Stacked Generative Adversarial Networks (StackGAN). Heart of such approaches is Conditional GAN which is an extension of GAN where both generator and discriminator receive additional conditioning variables c, yielding G(z, c) and D(x, c). This formulation allows G to generate images conditioned on variables c.

In our case, we train deep convolutional generative adversarial network (DC-GAN) conditioned on text features. These text features are encoded by a hybrid character-level convolutional-recurrent neural network. Overall, DC-GAN uses text embeddings where the context of a word is of prime importance. Class label determined in the earlier step will be of help in this case. This will simply help DC-GAN to generate more relevant images than irrelevant ones. Details will be discussed during the talk

The most straightforward way to train a conditional GAN is to view (text, image) pairs as joint observations and train the discriminator to judge pairs as real or fake. The discriminator has no explicit notion of whether real training images match the text embedding context. To account for this, in GAN-CLS, in addition to the real/fake inputs to the discriminator during training, a third type of input consisting of real images with mismatched text is added, which the discriminator must learn to score as fake. By learning to optimize image/text matching in addition to the image realism, the discriminator can provide an additional signal to the generator. (details are in talk)

Image Recommender System – In the last step, we propose personalised image recommendation for the user from the set of images generated by GAN-CLS architecture. Image Recommendation brings down the number of choice of images to a top N (N=3, 5, 10 ideally) with a rank given to each of those and therefore user finds it easier to choose. In this case, we propose Neural Personalized Ranking (NPR) – a personalized pairwise ranking model over implicit feedback datasets – that is inspired by Bayesian Personalized Ranking (BPR) and recent advances in neural networks. We like to mention that, now NPR is improved to contextual enhanced NPR. This enhanced Model depends on implicit feedbacks from the users, its contexts and incorporates the idea of generalized matrix factorization. Contextual NPR significantly outperforms its competitors

In the presentation, we shall describe the complete sequence in detail


Outline/Structure of the Talk

  • Introduction : Discussion on the problem statement
  • Objectives : Discussion on the target that can be achieved
  • Technical discussion on the three components of the problem:
  • Speech Processing using Deep-Speech, with glimpse of RNN
  • Text Classification using LSA
  • Image Generation using DCGAN and GAN-CLS
  • Image Recommendation using NPR
  • Complete Architecture and Implementation Overview
  • Innovation – Effective use of Different Algorithm for unique purpose
  • Practical Applications – Two Applications – one for business presentation, another for curbing illegal and terrorist activity
  • Future Scope of Improvements – Scopes of improvement and constant innovation to increase accuracy, accommodate instantaneous user preference and computation
  • Brief Demonstration

Learning Outcome

The audience is expected to receive appreciable knowledge on the following topic:

  1. Speech Processing with Deep Neural Net – Various Speech Recognition Software
  2. Image generation with GAN
  3. Different types of GANs and Best techniques with Architecture for text to image generation
  4. Image Recommendation with NPR

Also, some of the applications of this solution can be extended to other use cases as well which will be discussed during the talk

Target Audience

ML Researchers and Professionals, Image Processing and Computer Vision Professionals, AI Enthusiasts, Government and Private Security Agencies, Automated Computer Graphics and Design Professionals, Respected Professors teaching in the related fields

Prerequisites for Attendees

  1. Basics of Deep Learning and its Applications
  2. Basics of Recommendation Techniques and Applications
  3. Basics of Computer Vision
  4. Basics of Natural Language Processing
  5. Some idea on Speech Processing and Recognition
  6. An intuitive mind


schedule Submitted 3 years ago

  • 45 Mins

    Since we originally proposed the need for a first-class language, compiler and ecosystem for machine learning (ML) - a view that is increasingly shared by many, there have been plenty of interesting developments in the field. Not only have the tradeoffs in existing systems, such as TensorFlow and PyTorch, not been resolved, but they are clearer than ever now that both frameworks contain distinct "static graph" and "eager execution" interfaces. Meanwhile, the idea of ML models fundamentally being differentiable algorithms – often called differentiable programming – has caught on.

    Where current frameworks fall short, several exciting new projects have sprung up that dispense with graphs entirely, to bring differentiable programming to the mainstream. Myia, by the Theano team, differentiates and compiles a subset of Python to high-performance GPU code. Swift for TensorFlow extends Swift so that compatible functions can be compiled to TensorFlow graphs. And finally, the Flux ecosystem is extending Julia’s compiler with a number of ML-focused tools, including first-class gradients, just-in-time CUDA kernel compilation, automatic batching and support for new hardware such as TPUs.

    This talk will demonstrate how Julia is increasingly becoming a natural language for machine learning, the kind of libraries and applications the Julia community is building, the contributions from India (there are many!), and our plans going forward.

  • Dr. Vikas Agrawal

    Dr. Vikas Agrawal - Non-Stationary Time Series: Finding Relationships Between Changing Processes for Enterprise Prescriptive Systems

    45 Mins

    It is too tedious to keep on asking questions, seek explanations or set thresholds for trends or anomalies. Why not find problems before they happen, find explanations for the glitches and suggest shortest paths to fixing them? Businesses are always changing along with their competitive environment and processes. No static model can handle that. Using dynamic models that find time-delayed interactions between multiple time series, we need to make proactive forecasts of anomalous trends of risks and opportunities in operations, sales, revenue and personnel, based on multiple factors influencing each other over time. We need to know how to set what is “normal” and determine when the business processes from six months ago do not apply any more, or only applies to 35% of the cases today, while explaining the causes of risk and sources of opportunity, their relative directions and magnitude, in the context of the decision-making and transactional applications, using state-of-the-art techniques.

    Real world processes and businesses keeps changing, with one moving part changing another over time. Can we capture these changing relationships? Can we use multiple variables to find risks on key interesting ones? We will take a fun journey culminating in the most recent developments in the field. What methods work well and which break? What can we use in practice?

    For instance, we can show a CEO that they would miss their revenue target by over 6% for the quarter, and tell us why i.e. in what ways has their business changed over the last year. Then we provide the prioritized ordered lists of quickest, cheapest and least risky paths to help turn them over the tide, with estimates of relative costs and expected probability of success.

  • 45 Mins
    Case Study

    With the big boom in Data Science and Analytics Industry in India, a lot of data scientists are keen on learning a variety of learning algorithms and data manipulation techniques. At the same time, there is this growing interest among data scientists to give back to the society, harness their acquired skills and help fix some of the major burning problems in the nation. But how does one go about finding meaningful datasets connecting to societal problems and plan data-for-good projects? This session will summarize our experience of working in Data-for-Good sector in last 5 years, sharing few interesting datasets and associated use-cases of employing machine learning and artificial intelligence in social sector. Indian social sector is replete with good volume of open data on attributes like annotated images, geospatial information, time-series, Indic languages, Satellite Imagery, etc. We will dive into understanding journey of a Data-for-Good project, getting essential open datasets and understand insights from certain data projects in development sector. Lastly, we will explore how we can work with various communities and scale our algorithmic experiments in meaningful contributions.

  • Badri Narayanan Gopalakrishnan

    Badri Narayanan Gopalakrishnan / Shalini Sinha / Usha Rengaraju - Lifting Up: How AI and Big data can contribute to anti-poverty programs

    45 Mins
    Case Study

    Ending poverty and zero hunger are top two goals United Nations aims to achieve by 2030 under its sustainable development program. Hunger and poverty are byproducts of multiple factors and fighting them require multi-fold effort from all stakeholders. Artificial Intelligence and Machine learning has transformed the way we live, work and interact. However economics of business has limited its application to few segments of the society. A much conscious effort is needed to bring the power of AI to the benefits of the ones who actually need it the most – people below the poverty line. Here we present our thoughts on how deep learning and big data analytics can be combined to enable effective implementation of anti-poverty programs. The advancements in deep learning , micro diagnostics combined with effective technology policy is the right recipe for a progressive growth of a nation. Deep learning can help identify poverty zones across the globe based on night time images where the level of light correlates to higher economic growth. Once the areas of lower economic growth are identified, geographic and demographic data can be combined to establish micro level diagnostics of these underdeveloped area. The insights from the data can help plan an effective intervention program. Machine Learning can be further used to identify potential donors, investors and contributors across the globe based on their skill-set, interest, history, ethnicity, purchasing power and their native connect to the location of the proposed program. Adequate resource allocation and efficient design of the program will also not guarantee success of a program unless the project execution is supervised at grass-root level. Data Analytics can be used to monitor project progress, effectiveness and detect anomaly in case of any fraud or mismanagement of funds.

  • 90 Mins

    This will be a hands-on workshop how to build a custom interactive dashboard application on your local machine or on any cloud service provider. You will also learn how to deploy this application with both security and scalability in mind.

    Powerful Data visualization software solutions are extremely useful when building interactive data visualization dashboards. However, these types of solutions might not provide sufficient customization options. For those scenarios, you can use open source libraries like D3.js, Chart.js, or Bokeh to create custom dashboards. While these libraries offer a lot of flexibility for building dashboards with tailored features and visualizations.

  • 45 Mins

    In this digital era when the attention span of customers is reducing drastically, for a marketer it is imperative to understand the following 4 aspects more popularly known as "The 4R's of Marketing" if they want to increase our ROI:

    - Right Person

    - Right Time

    - Right Content

    - Right Channel

    Only when we design and send our campaigns in such a way, that it reaches the right customers at the right time through the right channel telling them about stuffs they like or are interested in ... can we expect higher conversions with lower investment. This is a problem that most of the organizations need to solve for to stay relevant in this age of high market competition.

    Among all these we will put special focus on appropriate content generation based on targeted user base using Markov based models and do a quick hack session.

    The time breakup can be:

    5 mins : Difference between Martech and traditional marketing. The 4R's of marketing and why solving for them is crucial

    5 mins : What is Smart Segments and how to solve for it, with a short demo

    5 mins : How marketers use output from Smart Segments to execute targeted campaigns

    5 mins: What is STO, how it can be solved and what is the performance uplift seen by clients when they use it

    5 mins: What is Channel Optimization, how it can be solved and what is the performance uplift seen by clients when they use it

    5 mins: Why sending the right message to customers is crucial, and introduction to appropriate content creation

    15 mins: Covering different Text generation nuances, and a live demo with walk through of a toy code implementation

  • Maryam Jahanshahi

    Maryam Jahanshahi - Applying Dynamic Embeddings in Natural Language Processing to Analyze Text over Time

    Maryam Jahanshahi
    Maryam Jahanshahi
    Research Scientist
    schedule 4 years ago
    Sold Out!
    45 Mins
    Case Study

    Many data scientists are familiar with word embedding models such as word2vec, which capture semantic similarity of words in a large corpus. However, word embeddings are limited in their ability to interrogate a corpus alongside other context or over time. Moreover, word embedding models either need significant amounts of data, or tuning through transfer learning of a domain-specific vocabulary that is unique to most commercial applications.

    In this talk, I will introduce exponential family embeddings. Developed by Rudolph and Blei, these methods extend the idea of word embeddings to other types of high-dimensional data. I will demonstrate how they can be used to conduct advanced topic modeling on datasets that are medium-sized, which are specialized enough to require significant modifications of a word2vec model and contain more general data types (including categorical, count, continuous). I will discuss how my team implemented a dynamic embedding model using Tensor Flow and our proprietary corpus of job descriptions. Using both categorical and natural language data associated with jobs, we charted the development of different skill sets over the last 3 years. I will specifically focus the description of results on how tech and data science skill sets have developed, grown and pollinated other types of jobs over time.

  • Sunil Jacob

    Sunil Jacob - Automated Recognition of Handwritten Digits in Indian Bank Cheques

    Sunil Jacob
    Sunil Jacob
    Sr. Architect
    schedule 3 years ago
    Sold Out!
    45 Mins
    Case Study

    Handwritten digit recognition and pattern analysis are one of the active research topics in digital image processing. Moreover, automatic handwritten digit recognition is of great technical interest and academic interest.

    In today’s digital realm, banks cheques are widely used around the world for various financial transactions. A rough estimate says that almost 120+ billion cheques move around the world. In the Indian banking scenario, CTS cheque clearance system has come. Even though the check is cleared quickly, there is still manual intervention needed to validate the date and amount fields. There is a lot of manual effort in this area.

    This case study, followed by a demo, will parade on how handwritten date and amount fields were extracted and validated. By adopting this automated way of recognising handwritten digits, banks can cut down the manual time and increase speed in their process. Although this is still in the proof of concept phase, this feat was achieved using computer vision and image processing techniques.

    This case study will briefly cover:

    • Detection of bounding and taking the region of interest
    • Fragment and Identify technique
    • Checking the accuracy of bounding box using Intersection over Union technique

    This case study/approach can be extended to other operative environments, where handwritten digits recognition is needed.

  • Kshitij Srivastava

    Kshitij Srivastava / Manikant Prasad - Data Science in Containers

    45 Mins
    Case Study

    Containers are all the rage in the DevOps arena.

    This session is a live demonstration of how the data team at Milliman uses containers at each step in their data science workflow -

    1) How do containerized environments speed up data scientists at the data exploration stage

    2) How do containers enable rapid prototyping and validation at the modeling stage

    3) How do we put containerized models on production

    4) How do containers make it easy for data scientists to do DevOps

    5) How do containers make it easy for data scientists to host a data science dashboard with continuous integration and continuous delivery

  • Dr. Neha Sehgal

    Dr. Neha Sehgal - Open Data Science for Smart Manufacturing

    45 Mins

    Open Data offers a tremendous opportunity in transformation of today’s manufacturing sector to smarter manufacturing. Smart Manufacturing initiatives include digitalising production processes and integrating IoT technologies for connecting machines to collect data for analysis and visualisation.

    In this talk, an understanding of linkage between various industries within manufacturing sector through lens of Open Data Science will be illustrated. The data on manufacturing sector companies, company profiles, officers and financials will be scraped from UK Open Data API’s. The work I plan to showcase in ODSC is part of UK Made Smarter Project, where the work has been useful for major aerospace alliances to find out the champions and strugglers (SMEs) within manufacturing sector based on the open data gathered from multiple sources. The talk includes discussion on data extraction, data cleaning, data transformation - transforming raw financial information about companies to key metrics of interest - and further data analytics to create clusters of manufacturing companies into "Champions" and "Strugglers". The talk showcased examples of powerful R Shiny based dashboards of interest for suppliers, manufacturer and other key stakeholders in supply chain network.

    Further analysis includes network analysis for industries, clustering and deploying the model as an API using Google Cloud Platform. The presenter will discuss about the necessity of 'Analytical Thinking' approach as an aid to handle complex big data projects and how to overcome challenges while working with real-life data science projects.

  • Krishna Sangeeth

    Krishna Sangeeth - The last mile problem in ML

    Krishna Sangeeth
    Krishna Sangeeth
    Data Scientist
    schedule 3 years ago
    Sold Out!
    45 Mins

    “We have built a machine learning model, What next?”

    There is quite a bit of journey that one needs to cover from building a model in Jupyter notebook to taking it to production.
    I would like to call it as the “last mile problem in ML” , this last mile could be a simple tread if we embrace some good ideas.

    This talk covers some of these opinionated ideas on how we can get around some of the pitfalls in deployment of ML models in production.

    We would go over the below questions in detail think about solutions for them.

    • How to fix the zombie models apocalypse, a state when nobody knows how the model was trained ?
    • In Science, experiments are found to be valid only if they are reproducible. Should this be the case in Datascience as well ?
    • Training the model in your local machine and waiting for an eternity to complete is no fun. What are some better ways of doing this ?
    • How do you package your machine learning code in a robust manner?
    • Does an ML project have the luxury of not following good Software Engineering principles?
  • Amit  Baldwa


    45 Mins

    Machine learning provides systems the ability to automatically learn and improve from experience without being explicitly programmed.

    Technical analysis shows in graphic form investor sentiment, both greed and fear. Technical analysis attempts to use past stock price and volume information to predict future price movements. Technical analysis of various indicators has been a time-tested strategy for seasoned traders and hedge funds, who have used these techniques to effective turn our profits in Securities Industry.

    Some researchers claim that stock prices conform to the theory of random walk, which is that the future path of the price of a stock is not more predictable than random numbers. However, Stock prices do not follow random walks.

    We will evaluate whether stock returns can be predicted based on historical information.

    Coupled with Machine Learning, we further try to decipher the correlation between the various indicators and identify the set of indicators which appropriately predict the value

  • Deepthi Chand

    Deepthi Chand / Shreya Agrawal - Samantar, an open assistive translation framework for Indic Languages

    45 Mins
    Case Study

    India is a land of many languages. There are 23 official and much more unofficial languages prevalently used in day-to-day conversations. Unfortunately, information dissemination to the low resource languages get difficult because of the geo-spatial distances. Popular translation platforms helped to fill this gap in major languages but their efficiency is challenged by the lack of availability of proper datasets and their generic nature. This problem is very evident when more domain information gets involved.

    We present Samantar, an open translation suggestion framework targeted at Indian languages. Samantar is built with open parallel corpora and opensource technologies. The translations can be tuned to suggest according to different target domains.