Talking AI in Ghana and the rest of the globe: An interview with Dr. Paul Azunre

Opinions and insights from a world-class data scientist and AI practitioner

Hello everyone,

Today, we continue our segment titled The Insider’s Corner, where we talk to AI experts and industry players from around the globe. Dr. Paul Azunre is a Ghanaian-American computer scientist, ML engineer, and AI entrepreneur with a PhD from MIT. Splitting time between his native Ghana and the US, Dr. Azunre is at the helm of various initiatives and projects across the African continent. He has kindly agreed to talk with us.

• We understand that you come from a rather unique background with parents from different parts of the planet. Could you briefly describe your upbringing and academic studies, including your time at MIT? How did you get into machine learning?

My dad is Ghanaian, and my mom is Russian. They met when my dad was pursuing his PhD in Russia. The first 10 years of my life were spent in Russia and the next 10 in Ghana. After completing my high school education in Ghana, I took the SAT. I did well, which — coupled with my grades and extracurriculars — landed me a full scholarship at Swarthmore College in Pennsylvania, USA. 

I continued to work hard and amassed solid research experience doing work in areas such as acoustic analysis, programming DSP chips to perform signal processing routines, and running different types of experiments. I ended up going into algorithms and coding early in my undergraduate career, but it wasn’t “machine learning” in the current understanding of the term (i.e., learning insights from vast amounts of data). My PhD was on the applications of optimization algorithms to solar energy. 

After completing my PhD, I started work on designing chips at Oracle in Austin (former Sun Microsystems division). When they shuttered the division, I found a job through the networks I’d been building in the Austin startup scene. The startup that hired me was tackling issues related to online information integrity, which was fundamentally an NLP problem. So, you can say I’m doing NLP partly as a consequence of fortune, or alternatively just following my interests.

• Not all of our readers are familiar with the technical aspects of AI model training and computational linguistics. Could you lay out the basics of Natural Language Processing (NLP)? To put it simply, what do you do exactly and how?

NLP deals with human speech: either text, audio, or sign language. This could mean making deductions about human speech, such as classifying it as some category or generating it (e.g., text-to-speech models). A major aspect of my work deals specifically with Ghanaian and other African languages. I work on problems in translation, LLMs, chatbots, and automatic speech recognition where such tools often don’t exist. 

I also work on problems in business analytics such as multilingual news analysis. The lines between NLP and, for example, computer vision have blurred — nowadays you often find yourself working within fully multimodal contexts. 

• You’re the author of Transfer Learning for Natural Language Processing, published by Manning. Could you briefly discuss the main themes of the book?

The book addresses the practical aspects of using NLP model architectures such as recurrent neural nets and transformers for popular tasks such as classification, language modeling, spam detection, and sentiment classification, among others. One of the key features of modern NLP architectures that have enabled the revolution we’re witnessing at the moment is the ability to reuse the same data across different tasks and even languages for different purposes. 

The term “transfer learning” refers to this process of knowledge transfer, and the book discusses some ways to do that with modern systems and tools. Although the field has advanced enormously since the release of the book, it remains a great reference for understanding the fundamentals and origins of all major AI models and chatbots. 

• Initially you specialized in electrical engineering and computer science as it relates to solar power. Do you believe AI can significantly enhance this sector? We also know that you collaborated with DARPA and the US Department of Energy, so what prompted your pivot to NLP? 

Some of my thesis work around solar panels was funded by the US Department of Energy. I cooperated with DARPA when I first started out in machine learning. The projects revolved around building tools that could automate away the tedious work a data scientist does.

We did things like create automatic machine learning (AutoML) libraries for simple classification problems and pipelines that could automatically clean, parse, and type data. The findings told me that you can build a lot of useful tools and make things easier, but obviously you aren’t going to automate away a programmer’s job completely. Honestly, I still laugh when I come across such claims.

I definitely think there’s an opportunity to leverage AI to discover new materials for more efficient and cheaper solar panels. However, solar is an area that requires massive investment to do anything worthwhile. When I began my studies, investment in solar research was high, but by the time I completed them, the financial interest had fallen. Discovering a more practical application of my skills — switching to NLP — made sense.

• We first became aware of your work through the GhanaNLP initiative, which we highlighted in one of our weekly AI Bulletins. But you’ve actually been expanding your efforts beyond Ghana with projects like Khaya AI and your research lab, Algorine. Could you elaborate on these undertakings?

Ghana NLP was formed at a time when not a single Ghanaian language had a translation app or solution available. For many of our languages, there wasn’t even a simple embedding model such as word2vec or a proper phone keyboard available. This was back when the BERT transformers disruption was in full effect, which, you may recall, is how we got to more modern tools such as ChatGPT. 

A few Ghanaians with visibility into these developments felt we had to do something to ensure our languages weren’t left behind, and our people weren’t excluded from AI’s benefits. We got together to form Ghana NLP and have since established it into a brand that’s recognized as a research leader in this space. A big part of what we do involves launching apps and APIs that ignite technological breakthroughs, often as building blocks by other developers. 

Together with communities and organizations in other countries, we’ve grown to provide solutions across Africa, among them Kikuyu, Kimeru, and Luo for Kenyans and Yoruba for Nigerians. We mentor students through our nonprofit GhanaNLP, put out research datasets for African languages in partnership with other organizations, and innovate at the bleeding edge of ML research. We hope to contribute whatever we can to the Ghanaian AI community and ecosystem. Algorine is part of the same ecosystem and tackles the exact same objectives, but from the perspective of a private company.

• What about the current state of AI in Ghana and more broadly across the African continent? What potential do you see for the future, especially with policies like the recently adopted Continental AI Strategy? The term “empowerment” is used often these days — do you think it accurately reflects the regional benefits AI is bringing to Africa?

Things are certainly picking up — there’s a lot of interest. Top universities are training skilled professionals, and we’re beginning to see some investment and jobs. We’re also seeing impressive raises by AI startups across the continent. The investment in AI in Ghana is obviously substantially lagging behind other African nations, such as Morocco, Nigeria, South Africa and others, but I guess this is something to be expected. 

The policy initiatives you mention are great in that they show intent to establish the right frameworks for AI ecosystems to thrive, in which case “empowerment” is possible. With that said, unfortunately, as of now, the reality on the ground often looks dangerously closer to one where AI is just another dimension where obscene inequalities get amplified. In short, there’s still much work to be done.

• Launching a successful AI company is a formidable challenge, with many tech startups failing to become profitable before depleting their runway capital. It appears you’ve managed to mitigate this risk by keeping your day job and also diversifying your income sources, which include music production (as per your radio interview on 3FM 92.7). What advice would you offer to any aspiring AI startupreneurs out there?

The entrepreneurship mantra in the US is all about “raise huge capital, move quickly, try to disrupt, and fail if you must, so you can do it again working on another problem… until you find one that is really profitable.” I don’t know if that approach is the right one for building Ghanaian language tech. In my case, it’s not profit per se that I’m chasing. I’d like to think that I’m embracing entrepreneurial spirit or even passion, employing creativity every step of the way to address my long-term vision. 

I realized quite early on that no one was going to give me millions of dollars for this. So, if I wanted to get it done, I should focus more on “bootstrapping.” I’m happy with that choice since I get more freedom to carve out the path I want. My music work is a manifestation of exactly that. It’s certainly a longer journey, but it’s one I’m committed to and I also enjoy. I encourage other entrepreneurs to define for themselves what’s important to them. A multimillion dollar VC raise is great, but it’s just one way to define success: one that may not even be feasible for every valuable problem that needs solving.

• Based on your extensive experience in the AI industry, could you share any observations or even predictions that might not be apparent to non-engineers or business executives?

There’s a lot of hype nowadays about the impending “AGI” and “existential” threats. I hear outlandish claims about capabilities of AI products on a daily basis that are proven to be false virtually within hours. I encourage everyone to take a deep breath and treat any such claim with skepticism — take it with a pinch of salt, so to speak. 

Yes, this is really powerful tech that can be super useful, but it lacks independent agency, so it isn’t going to replace us. Be that as it may, certain bad actors are already utilizing AI to advance their own interests, resulting in unethical behaviors such as privacy violations, resource disparities, and manipulation of public opinion.

This is a much more pressing matter, and I worry about it far more than the supposedly impending AGI fantasy. It’s not AI we should be concerned about — it’s people misusing AI. Which is why I believe leveraging AI in a conscientious manner is solely our responsibility, not the machine’s.