Machine intelligence for the interpretability of olfactory centered (multi-modal) data - QC-231

Desired discipline(s): Engineering - chemical / biological, Engineering, Computer science, Mathematical Sciences, Mathematics, Chemistry, Natural Sciences
Company: Stratuscent Inc.
Project Length: Longer than 1 year
Preferred start date: 11/01/2019
Language requirement: Bilingual
Location(s): Montreal, QC, Canada
No. of positions: 6

About the company: 

Stratuscent is a Montreal-based start-up that delivers business intelligence using a breakthrough electronic nose that detects and digitizes everyday scents. The company’s innovative platform combines connected chemical sensors along with novel artificial intelligence (AI) algorithms that offers customers a real-time view of chemical fingerprints of their products as they evolve.

Please describe the project.: 

Stratuscent develops a novel and powerful odor information processing system (Electronic nose).  This system is useful to digitize the world of smells, with potential applications for several commercially valuable markets.  In particular, Stratuscent’s electronic nose democratizes olfaction-based contextual information processing. Stratuscent’s system comprises of a sensor module that is based on a nano-composite material technology exclusively licensed from JPL, NASA. Using novel nano-composites, Stratuscent has built an array of artificial olfactory neural receptors. This chemical receptor array generates a unique Scentprint when it comes in contact with any chemical or smell. This Scentprint is then processed using Artificial intelligence methods to interpret various contextual information. Currently, Stratuscent is commercially focused in the food and beverages industry where smell-based information can be used to improve the entire food supply chain.

One of Stratuscent’s current challenges is to handle the large manufacturing variation across the sensor modules. They have had some initial success in learning mutual representations across devices with manufacturing variations. However, to build a scalable solution, Stratuscent is looking to explore and leverage structural learning approaches in order to improve the representation learning techniques.

The scope of the project is to develop mutual representation learning methods where leveraging the semantics across input modalities is important.  This project includes the development of few shot learning algorithms for machine-olfaction.

Stratuscent is developing reinforcement learning based systems that would leverage exploration for modelling abstract chemical semantics within the data space.  This project includes the development of a data acquisition system for enabling action-observation-reward loops.

Stratuscent is continuously developing a tool for generalizing the creation of machine intelligence-oriented solutions. Adding new mechanisms, methods, algorithms and maintaining the developments is an ongoing project which helps them to efficiently provide solutions.

Research objectives / sub-objectives

  1. Development and implementation of an automated system for collection of data with an intelligent backend.
  2. Integrating systems for processing information from other modalities including image and text.
  3. Development of reinforcement learning algorithms which can effectively explore and learn discrete or continuous odor data spaces.
  4. Development and deployment of generative models that mimic data observations in contexts that are aligned within Stratuscent’s market.

Stratuscent’s technologies are about digitizing smells by capturing and processing information provided through odors.  The semantics associated with an odor could vary given the context.  Moreover, part of the context information for a datapoint can be provided through other modalities such as text or image. Stratuscent’s methodologies are centered around learning mutual representation of semantics given observations from (i) various sensing modules aka sniffer devices as well as (ii) complementary modalities such as text.

Our approach is based on learning reduced maps that transform the input data spaces to semantic spaces where a unified and decentralized system enables collaborative intelligence.

Required expertise/skills: 

  • Intermediate knowledge of Python is a must; knowledge of C/C++, java and C-sharp is a plus.
  • Experience working on Machine learning applications/research, with a focus on computer vision, sequence modeling and natural language processing/generation etc.
  • Experience with tensorflow, keras, numpy, tfx, gym, matplotlib libraries.
  • Research skills and scientific literacy.
  • Knowledge/Previous Studies in cloud computing, databases and robotics is a plus.
  • Development and testing under the following environments: Windows, Linux, Mac OS, Android (Android studio), iOS (X code)
  • Background/knowledge in chemistry and working with chemical datasets is a plus.

We are seeking 2 PhDs for 3.5 years and 4 MSc students for 2 years.