We recently caught up with Petar Veličković, a research scientist at DeepMind. Together with his co-authors, Petar presents his paper The CLRS Algorithmic Reasoning Benchmark at ICML2022 in Baltimore, Maryland, USA.
My Journey to DeepMind…
During my undergraduate days at Cambridge, the inability to play Go skillfully was seen as clear evidence of the shortcomings of contemporary deep learning systems, and I always wondered how mastering such games could possibly be beyond the realm of possibility.
But in early 2016, when I was starting my PhD in machine learning, everything changed. DeepMind faced off against one of the best Go players in the world in Challenge Matchwhich I spent several sleepless nights on. DeepMind won, creating groundbreaking gameplay (e.g. “Move 37”) in the process.
From that moment on, I thought of DeepMind as a company that could make seemingly impossible things happen. So I focused my efforts on joining the company one day. Shortly after submitting my PhD in early 2019, I began my journey as a research scientist at DeepMind!
My role…
My role is a virtuous cycle of learning, research, communication and consulting. I am always actively trying to learn fresh things (lately Category theorya fascinating way to study computation Structure), read relevant literature, watch lectures and seminars.
Then, using these insights, I brainstorm with my teammates about how we can extend this knowledge to positively impact the world. Ideas are generated during these sessions, and we apply a combination of theoretical analysis and programming to establish and test our hypotheses. If our methods bear fruit, we typically write a paper sharing our insights with the broader community.
Research results are not as valuable without communicating them properly and enabling others to apply them effectively. For this reason, I spend a lot of time presenting our work at conferences like ICML, giving talks, and mentoring students. This often leads to fresh contacts and the discovery of fresh scientific results to investigate, which starts the virtuous cycle again!
At ICML…
We present a presentation of our work, CLRS Algorithmic Reasoning Testwhich we hope will support and enrich activities in a rapidly developing field algorithmic neural reasoningIn this study, we ask graph neural networks to perform thirty different algorithms with Introduction to algorithms Coursebook.
Many recent research efforts aim to construct neural networks capable of performing algorithmic computations, primarily to endow them with reasoning capabilities—which neural networks typically lack. Importantly, each of these papers generates its own dataset, making it hard to track progress and raising the barrier to entry into the field.
CLRS benchmark with readily available dataset generators and publicly available codeis working to address these challenges. We have already seen a high level of enthusiasm from the community and hope to build on that even more at ICML.
The Future of Algorithmic Reasoning…
A central dream of our research in algorithmic reasoning is to capture the computation of classical algorithms inside high-dimensional neural executors. This would allow us to implement these executors directly on raw or loud data representations, thus “applying the classical algorithm” to inputs on which it was never designed to execute.
What is thrilling is that this method has the potential to enable data-efficient reinforcement learning. Reinforcement learning is full of examples of powerful classical algorithms, but most of them cannot be applied in standard environments (such as Atari) because they require access to a wealth of privileged information. Our design would enable such an application by capturing the computations of these algorithms inside the neural executors, after which they could be directly implemented in the agent’s internal representations. We even have a working prototype that has been published on NeurIPS 2021. I can’t wait to see what happens next!
Can’t wait…
Can’t wait ICML Workshop on Human-Machine Collaboration and Teamworka topic close to my heart. I fundamentally believe that the greatest applications of AI will come from synergies with human experts. This approach is also very much in line with our recent work on enhancing the intuition of pure mathematicians with artificial intelligencewhich appeared on the cover of the journal Nature tardy last year.
The workshop organizers have invited me to a panel discussion to discuss the broader implications of these efforts. I will be speaking alongside a fascinating group of co-panelists, including Sir Tim Gowerswho I admired during my undergraduate studies at Trinity College, Cambridge. Needless to say, I am really excited about this panel!
Looking to the future…
For me, huge conferences like ICML are a moment to pause and reflect on diversity and inclusion in our field. While hybrid and virtual conference formats make events accessible to more people than ever before, we need to do much more to make AI a diverse, equitable, and inclusive field. AI interventions will affect all of us, and we need to make sure underrepresented communities remain a vital part of the conversation.
That’s why I’m running a course on this topic Geometric deep learning on African Master of Science in Machine Intelligence (AMMI) – the subject of my recent co-authored work proto-book. AMMI offers top-notch machine learning courses to Africa’s brightest emerging scientists, building a vigorous ecosystem of AI practitioners in the region. I am very joyful to have recently met several AMMI students who have joined DeepMind for internships.