Cognitive behavioral therapy and machine learning to speed up adult learning

5 minute read

Below is my original statement of purpose for my PhD. One day, I hope to have the time to conduct more academic research. Contact me (or send an email to jobs@bld.ai ) if you are interested in this topic.

1. Danny Castonguay is a co-founder of bld.ai

I am a father of three and a co-founder of bld.ai, a rapidly growing full-stack design studio and software development shop. I’m the largest shareholder of bld.ai. We were founded in December 2017 and as of November 2021, we have 150 employees, including engineers, data scientists, designers, and product managers. Our clients include five global corporations (energy, mining, tech, telecom, and defense) as well as a few dozen well-funded tech startups (e.g., venture-backed by Andreessen Horowitz).

2. bld.ai growth requires new knowledge to accelerate learning as a competitive advantage

Our clients pay us hourly to create with them human-centered software such as a) a digital twin for frontline worker safety, b) a mobile crypto wallet for foreign remittances, and c) a mobile platform for pediatric research. Our goal is to grow to 550 people in 2022 and to grow 100X every four years, reaching 1 million employees by 2030. In order to achieve this ambitious goal, we must build software that allows us to be distinctive in recruiting, training, mentoring, and retaining top design and engineering talent around the world. On-the-job training is of particular strategic interest to bld.ai’s long-term prospects as a differentiating factor and has been a lifelong mission and purpose.

3. Cognitive-behavioral therapy and artificial intelligence may speed up adult learning

As a data scientist with over a decade of professional experience (ranging from natural language processing to computer vision across multiple industries such as energy, mining, travel, finance, and tech), I aim to advance the field of research at the intersection of cognitive-behavioral therapy (CBT) and artificial intelligence, specifically reinforcement learning (RL). My goal is to study and understand the science of learning, and to develop techniques to increase the speed of learning in young adults.

4. Danny has a background in computer engineering and business

I completed my undergraduate and my master’s degrees in computer engineering at McGill University. For my master’s, I studied under Profs Shie Mannor, Doina Precup, and Benoit Boulet. I also completed my MBA at MIT, which is where I met my friend, classmate, and potential co-supervisor Prof. Chaithanya Bandi, currently an Assistant Professor in analytics and operations with NUS Business School. I’m also an alumni of the prestigious Y-Combinator incubator in Silicon Valley. Paul Graham had a profound impact on how I think about startups. My experience also includes raising millions of dollars in venture capital and serving as an associate partner at McKinsey & Company, where I helped global corporations with digital transformation.

5. Cognitive behavioral therapy seems effective at improving mental health

CBT is an evidence based treatment proven to be effective at treating conditions such as anxiety, depression or addictions such as smoking or overeating. The overall emerging theory is that what we think and what we do can influence how we feel and how we behave, in a self-reinforcing feedback mechanism. Over time, patterns can lead to undesired (or desired) ways of thinking. With CBT, we can train ourselves to change our thoughts. Typically CBT is used by mental health professionals to treat problems, and is quite costly and requires heavy calibration.

6. Examples of combining CBT with artificial intelligence

Woebot is an example of combining CBT with artificial intelligence. Dr. Alison Darcy, CEO of Woebot, was a postdoctoral researcher at Stanford University, where she ran the Health Innovation Lab with Prof. Andrew Ng, founder of Coursera. Their research over the years has explored how to integrate human-centered design thinking and big data into technology-based psychotherapy.

7. Cognitive behavioral therapy could be used to improve learning

The main problem with applying CBT to help people learn better and faster is the return on investment: it is too expensive to hire a therapist (or a trained coach) to follow a student who just wants to learn Python Pandas faster at work. The hypothesis I want to test experimentally is whether it is possible to leverage machine learning to replicate some parts of the role of a therapist, thereby lowering the bar and offering more self-serve CBT to help everyone benefit from some of these breakthrough techniques. The reinforcement learning agent will use information it has on each individual learner, decide whether to explore or exploit its current action policy, and take actions, where actions include a) what content to display (including CBT elements), b) when to display it, and c) when to notify the coach. While Woebot is fully automated, my research will also explore how CBT trained coaches can leverage AI to improve the learning speed and quality of their students.

8. Example experiment to measure the efficacy of CBT and RL to improve adult learning

I will assemble a population of young adults including: current and prospective employees of bld.ai (designers, product managers, and engineers), current employees of two or three major corporations who are taking corporate training lessons in Python, and a random selection of people including students and startup entrepreneurs. Then, I will create a set of micro-learning modules on Python Pandas, including fragment lessons and multiple choice questions. After that, I will organize those modules into an adaptive learning management system (LMS). The effectiveness of introducing CBT strategies and RL algorithms to the LMS will be evaluated against this benchmark. The experiment will be a double blinded study over 3 months to measure how the sub-populations perform with and without CBT and RL. It is my objective to publish research findings in journals such as ICML, as well as create free and open-source software that can be reused by other researchers or corporations like bld.ai who can use it commercially.

9. This research will be partly funded by bld.ai and some of its clients

If this works well for teaching people about Python Pandas, it is likely that this could generalize well to many other skills. The potential economic upsides for a company like bld.ai are a) reduced training time, b) improvement in employee retention, and c) increase quality and speed of services offered to its clients. It makes sense for bld.ai to sponsor me personally to conduct this research. In addition to bld.ai sponsoring this research, I am also seeking grants and funding from other corporations and organizations that would also benefit from this research. Please send an email to research@bld.ai to get in touch.