Project Astrl

Bridging the future of AnalogML đź’»
Machine Learning
Signal Processing
Internet of Things
Embedded Systems
Jan 2022 - Aug 2022
Raga Kavari
Sofia Kirkman
Linda Xue
Tammy Zhou
Lead Product Designer
& Design Technologist

Team

My Role

Domain

Timeline

Executive Summary

Solution

Astrl is a platform that Aspinity would provide to their client companies to enable their ML engineers to be able to program AnalogML applications on Aspinity’s hardware architecture. The platform helps ML engineers who are not familiar with this new technology, be able to discern the difference between Analog and standard ML, be able to make their own application in a software simulation, and then fine-tune their project to be deployable on Aspinity’s hardware chip. 

Problem

Aspinity has developed a way to do machine learning on analog data using special hardware architecture that is 20 times more power-efficient than standard machine learning performed on digital data. Currently, Aspinity’s clients rely on Aspinity to develop end-to-end applications for them. But clients want to keep their data in-house and have more control over the software development process when using Aspinity’s hardware. Moreover, AnalogML is a new niche technology that ML engineers at client companies know nothing about. 

Prototype and Design System

My Role

Design Leader

I am leading the product design vision of our platform. This role includes involving my team in brainstorming and ideation sessions, consolidating ideas into wireframes, and dividing work among my teammates. I also am leading the handoff of the design to the development team.

Technical Bridge

Being the only person on the team with a CS background, I translate the technical details of users' workflow and working of analog tech to my teammates, faculty, and other stakeholders. I employ tools such as visual modeling for communicating technical processes in an easily consumable manner.

Design Evangelist

I am also involved in educating colleagues in engineering divisions at Aspinity about the design process. I help colleagues understand the user-centered design process and answer questions related to the concepts like user research and testing.

Project Context

Through their AnalogML technology, Aspinity brings extremely low power solutions to problem spaces in embedded systems and IoT devices. This technology increases battery life by 20x for smart assistants, home security systems, cyber-physical systems, or any other technical systems that can employ event-detection applications on incoming data.
‍
Currently, engineers at Aspinity do the heavy lifting of understanding client requirements, creating the required application, and shipping it to the client.

Through Project Astrl, we are helping Aspinity grow by providing its clients with the tools and freedom to craft their own custom Machine Learning applications on the blank canvas of Aspinity’s analog chip technology.

Target Users

Through the project brief and interactions with our client, we discerned that the target users for our project would be - ML Engineers at Aspinity's client companies who would be tasked to program an ML application to be run on Aspinity's chip. Hence, someone like Robin.

Robin would interact with co-workers like Cam to understand the analog signals they'll be working on to effectively extract features out of the data. Robin would also interface with co-workers like Dani to make sure their model runs with the same performance on the actual hardware chip.

Understanding the User

Research Goals

  • To understand how the ML Engineer goes about their task of programming an ML model on the chip.
  • To know what methods and tools they currently use in their process and identify opportunities where the experience can be improved.
  • To identify other stakeholders with which they interface and realize how their needs can be integrated into our solution.

Who We Talked To

4 ML Engineers working in the Signal Processing Team at Aspinity.

‍

Methods Used

Conducted semi-structured interviews with 4 ML Engineers at Aspinity. Later, conducted contextual inquiry sessions with 2 ML Engineers who were asked to think-aloud while going about their tasks that informed technical user-journey maps of their workflow.

AnalogML is a very niche technical domain and concepts related to analog signal processing are not commonly known to ML Engineers. There is a need to bridge this knowledge-gap.
1
3
2
ML Engineers would need to collaborate with other engineers on their team to be able to create the ML model end-to-end. There needs to be smooth communication between engineers on the team.
Visualizations are a crucial analysis tool for ML engineers. They need to have access to methods that allow quick and easy tweaking and generating effective data visualizations rapidly.

Insights

Collaboration between Engineers

I created this model to illustrate how different engineers are involved in the AnalogML development process from end to end. The model describes their interdependencies and the data channels among them.

Modeling the workflow of an ML Engineer

I abstracted my understanding into this model to communicate to my team the technical processes involved in the model development cycle. I also identified the user's goals and critical needs at each of these phases.

Understanding the Needs of the User

Research Goals

  • Validate needs of users at different points in their application development process that surfaced while understanding their workflow.
  • Discern needs that are the most pressing and pain points that cause the most friction in the workflow.
  • Gauge opinions on potential solutions that can be employed to meet the needs of the users.

Who We Talked To

8 Engineers working at Aspinity (internal) and 10+ engineers at other companies (external) working on problems employing ML applications.

‍

Methods Used

Competitive Think-Aloud: Internal engineers at Aspinity were asked to use an existing ML programming platform for programming a wake-word ML application, and think-aloud as they go about performing this task. The aim was to gauge the specific differences that would surface in the application development process when working with a general-purpose ML platform as opposed to specifically for AnalogML.

Speed Dating: Both internal and external engineers were asked to share their thoughts on various storyboards of concepts that could potentially satisfy identified needs of the engineers at each step of the application development process. The aim was to gauge technical feasibility and engineers' preferences on these different solutions.

There are special challenges associated with working with analog data and there are technical limitations to data-science methods that can be applied to this domain. ML Engineers need to be equipped with tools to handle these differences.
Engineers have limited bandwidth and time to edify educational resources. Hence, it is important to provide them with just enough information so that they can hit the ground running.
1
3
4
2
Trust in data is always a concern. Users need transparency in the form of detailed information about sourcing if datasets are provided, and literature backing if any data science functions are provided.
While support from experts at Aspinity will be needed, mechanism should be in place to reduce this interfacing and enable users to asynchronously help themselves.

Insights

Overarching Problems

“So these concepts will be new to [clients]. And to the next person we hire.”
“They don't [] need to understand how it works. They just need to know [] they've been handheld so they don't make mistakes...”
- Senior Algorithm Architect at Aspinity
- Engineering Manager at Client company of Aspinity
1

High Effort to Learn and Integrate Technology

AnalogML is a new niche within the domain of ML that most practitioners have little awareness of, let alone an understanding. Currently, clients sit down with experts from Aspinity to learn the intricacies of the technology to develop a fundamental grasp over AnalogML.

Client engineers don't need to know all the little details that will not affect them in their work, but they should know how to work with the technology. ML engineers will need help in understanding the high level relationships and requirements of analog and hardware. They may also need support in deciphering unexpected issues with their model.

“There is no such thing as a user that will know how to do it all.”
“Different steps of the application process require different sets of skills and there are currently distinct hand-offs.”
- ML Research Expert
- Engineer at a Client Company of Aspinity
2

Cross Functional Teamwork Lacks Effective Collaboration

There exists no individual in the team who will know how to do it all - yet everyone must be aware of each step of the process because different steps require specific skill sets with handoffs to one another. If one step of the process is impacted, it creates a ripple effect for the rest of the team. 

ML Engineers often have to tackle complex problems solo, with whatever tools and knowledge they have. Effective team collaboration between diverse individuals needs to be aided.

“Having something like this you can on the fly just quickly change some parameters and get those plots...it doesn’t really take a lot of time whereas if you’re doing it on code you have to wait until it plots and store and then you have to rerun and replot...it’s more time consuming and a little cumbersome.”
- Engineer at Aspinity
3

Manual Tasks are Inefficient, Inflexible, and Inevitable

Currently, when Aspinity engineers develop an application, they perform several tasks that require lots of manual labor. They do literature searches of ML models, listen to each data point and label it, change parameters, develop figures and run simulations. Several of these steps, particularly anything involving code, requires the developer to restart processes every time they make a change. Sometimes users do not know what to change to make a model better which requires them to do more research and experimentation.

Right now, these steps are inevitable because there is no platform to make these processes smoother. There is a need to automate steps like these and provide recommendations which would ultimately save developers time.

Key Value Proposition

Through my research, I proposed that as Aspinity continues to grow, it will not be just a chip company. Moving forward, Aspinity should perform as an AnalogML service company to its clientele. Product design thinking would bring this key value into Aspinity's current roadmap.
Drawn by the Chip.
→
Stayed for the Service.

Identifying Opportunities of Improvement in Aspinity's Service

My team identified how Aspinity currently works with its potential partners from initial discovery to final deployment. The blueprint visualizes the organizational processes that deliver the current customer experiences.

Key Design Implications

1

Providing Educational Support & Guidance

  • To traverse the steep learning curve of AnalogML technology, we need to provide contextual guidance and sample use cases for ML engineers.
  • Providing users with the necessary tools to explore and learn about the data requirements for all target use scenarios for their application.
  • Bridge the knowledge gaps between different domains of expertise so that engineers in signal processing, ML modeling, and hardware can all build a shared understanding of their individual role during the solution development.
2

Encouraging 
Strategic Collaboration

  • Include methods to strengthen collaboration when a user needs help from a team member with complementary expertise. Helping communicate expectations and streamlining their workflow strategically. 
  • Providing mechanisms which enable users to seek contextual support within the platform from experts at Aspinity.
  • Tools for users to easily communicate and share current progress with their team effectively.
3

Reducing Speed Bumps 
in Application Development

  • Providing a workspace which reduces manual ways of doing repetitive tasks, and unnecessary back and forth between their IDE, terminal and other application windows. 
  • Integrating visual analysis tools that users are familiar with, while providing them with knobs to easily tweak parameters in the platform. A space to store and come back to these visualizations conveniently.
  • Curating a library of relevant example projects, and research literature that helps users quickly contextualize and start on their current project.

The Astrl Design Vision

To come up with the platform that would enable Aspinity to be the leading AnalogML service, my team envisioned a platform that...

How might we enable ML engineers to feel confident with AnalogML before they start the dev process?

I decided to focus the first design sprint on understanding how I can help ML engineers, who are starting their journey with this platform, feel supported and be able to establish a mental model of what they can do with this tool.

The team prototyped ideas at different levels of technical feasibility and implementation effort that would help users understand the scope of their project and feel confident to start working on it.
These prototypes were tested with 4 representative participants by asking them to think aloud as they experienced interfacing with these concepts.
Users preferred to dive into the dev process rather than go through a walk-through tutorial despite their knowledge gap of AnalogML.
1
3
2
Users wished the platform was tailored to their specific learning levels. Users preferred to have the context of the skill required and an overview of the learning before execution so they can orient themselves before trying out the new process.
When starting, users wished there was guidance towards one starting point to begin their learning experience. If there are multiple tools, it is needed to explicitly highlight the first point of interaction.

Insights

How might we help ML engineers learn AnalogML through experience with an example use-case?

Through this design phase, my aim was to curate the users' experience through the actual application development process end-to-end. Building on learnings from the last sprint, I pushed on employing a pre-filled example use-case as it provided users with a starting point for reference but also gave them autonomy to take ownership of their learning process.
The team evaluated the success of our prototypes by measuring pre-and-post levels of user confidence rating in developing an AnalogML application. In addition, we also analyzed users' reported behaviors and actual behaviors with task-based scenarios during the think-aloud test. This helped us understand users’ mental model of when and where they expected to find contextual support and interactive assistance from the platform.
While example projects are a helpful resource, they are not the only 
point of entry into the application development process. 3/6 users preferred not to start a project by walking through an example use case project. The entire user learning experience can not be centered around the example project.
1
3
4
2
3/6 of users expressed concerns over sharing data and their developed projects publicly. Due to the nature of some work our users are doing, projects must prioritize privacy and viewing permissions.
Users need clear visibility and control over tuning parameters and selecting metrics for comparing models. The prototype needs to accurately reflect the workflow of the ML application development process.
Users were confused regarding the actual differences between Analog and Standard ML. They could go through the process because of enough scaffolding but wanted more context about the technicalities of AnalogML.

Insights

How might we help ML engineers understand the difference between Analog and Standard ML processes?

In the next sprint, my aim was to increase the technical fidelity of the prototype to delineate the difference between standard and analog ML workflow better.

Being in charge of client communication this sprint, I also set up a co-design session with our clients at Aspinity that provided us with different perspectives and helped us collaborate & bond better.

Reflection

This project has been one of the most demanding, but rewarding design challenges I've undertaken. Being the only team member with a technical background, it has been my responsibility to support my team and bridge their understanding towards this convoluted research space.

Being an engineer myself helps me empathize with our users, and understand their vocabulary. Another key skill this project has taught me is to effectively understand the human side of the technical interactions, translate them to simpler concepts, and communicate them to an audience that may not have deep understanding of this complex domain.

This capstone project is especially meaningful because Aspinity plans to actually put our work into production at the end of this project. This platform will help bring AnalogML technology to a wider audience that will lead to its employment in various everyday devices and make them much more power-efficient. I look forward to the next few months of this fascinating and impactful project.
Linkedinemailresume