I am Dhaval Adjodah, a machine learning and policy researcher working on pushing the limits of modern machine learning while maximizing their social good. Currently, I am a research scientist at the MIT Quest for Intelligence whose purpose is to make advances towards building machine intelligence. Previously, I was a research scientist at the MIT Media Lab where I developed new machine learning algorithms using insights from social and cognitive science. I am also a consultant for the World Bank, where I am helping build new machine learning pipelines to track and implement SDG policy.
In addition to my day-to-day work (software engineering, running live online human experiments and writing papers), I enjoy being of service to the computer science and social science communities, most recently as a member of the organizing committee of the AI for Social Good conference workshop series, and on the program committees of the Black in AI initiative and Theoretical Foundations of Reinforcement Learning workshop.
I did my PhD at the MIT Media Lab with Profs. Alex Pentland, Neil Lawrence, Tim Klinger and Esteban Moro on my PhD committee. I also interned with Prof. Yoshua Bengio at MILA, was a member of the Harvard Berkman Assembly on Ethics and Governance in Artificial Intelligence, and was a fellow at the Dalai Lama Center For Ethics And Transformative Values. Previously, I worked as a data scientist in banking and insurance, consulted for the Veterans Health Administration, and founded two startup incubators. I hold a Masters degree from the Technology and Policy Program now under the MIT Institute for Data, Systems and Society, and a Bachelors in Physics (with a focus on Non-Linear Systems Modeling), also from MIT.
There has been a lot of recent work showing that sparsity in neural network structure can lead to huge improvements, such as through the Lottery Ticket Hypothesis. Coming from a computational social science background, we know that humans self-organize into sparse social networks. My hypothesis was that organizing the communication topology (social network) between agents might lead to improvements in learning performance. This is especially important because some machine learning paradigms—especially reinforcement learning—are becoming more and more distributed in order to parallelize learning, similar to how human society balances exploration and exploitation. Well, we find huge improvements: 10–798% improvement on the state-of-the-art robotics simulators! Accepted at AAMAS 2020. [Code, paper]
There has been a proliferation of research investigating whether neural networks can learn compositionally and hierarchically. One of the significant challenges in this line of research is the lack of benchmarks that are easily extensible and modifiable by researchers. In this work, we develop a new domain specific language that allows researchers to painlessly develop tasks for classification by neural networks. For example, a classification tasks might involve distinguishing 3 types of “concepts” such as “all objects in the image are green”, “some images are blue and square”, “blue crosses are next to red F-pentamino”. Each of these concepts are defined using high-level boolean-logical functions, which can themselves be other boolean functions, allowing for arbitrary depth and width of hierarchies. Using 4 simple tasks, we demonstrate that even state-of-the-art models do not show high performance on compositionality. [paper]
How do we design and deploy crowdsourced prediction platforms for real-world applications where risk is an important dimension of prediction performance? To help answer this question, we conducted a large online Wisdom of the Crowd study where participants predicted the prices of real financial assets (e.g. S&P 500). We observe a Pareto frontier between accuracy of prediction and risk, and find that this trade-off is mediated by social learning i.e. as social learning is increasingly leveraged, it leads to lower accuracy but also lower risk. We also observe that social learning leads to superior accuracy during one of our rounds that occurred during the high market uncertainty of the Brexit vote. Our results have implications for the design of crowdsourced prediction platforms: for example, they suggest that the performance of the crowd should be more comprehensively characterized by using both accuracy and risk (as is standard in financial and statistical forecasting), in contrast to prior work where risk of prediction has been overlooked. [paper]
Over the past 15 years, 73 percent of deadly terrorist attacks on U.S. soil were carried out by far-right extremists. While the emergence and persistence of these far-right groups is a complex process, influenced and enabled by structural forces and institutions, the organizations themselves often manifest as online social networks. These social networks serve as radicalizing and organizing spaces for extremist groups. In this work, we analyze leaked data from the Iron March website, a neo-nazi online forum active from 2011 through 2017. We provide a descriptive analysis of how the social network of users changes over time, and how the contents of their discourse appears to shift. The work presented here offers a step towards understanding the organizing behaviors of online extremist communities. This study provides evidence that extremist groups may start out ideologically centered, but that their growth may be understood as a process of social bonding.
- Leveraging Communication Topologies Between Learning Agents in Deep Reinforcement LearningAutonomous Agents and Multi-Agent Systems (AAMAS) 2020
- Far-Right Forum Network and LDA analysisInternational Conference on Computational Social Science 2020
Polarization, Inaccurate Meta-perceptions, and the Erosion of Democratic NormsInternational Conference on Computational Social Science 2020
Symbolic Relation Networks for Reinforcement LearningReinforcement Learning Relational Representation Learning workshop, Neural Information Processing Systems 2018
How to Organize Your Deep Reinforcement Learning Agents: The Importance of Communication TopologyDeep Reinforcement Learning Symposium, Neural Information Processing Systems 2018
Inferring Mobile App Preference via Multi-view Geometric Information FusionINFORMS Conference on Information Systems and Technology (CIST) 2018
Social Bayesian Learning in the Wisdom of the CrowdInternational Conference on Computational Social Science 2017
Cognitive Limitations in Financial NetworksNetwork Science 2017
Harnessing Social Learning to Improve Crowd-Sourced PredictionInternational Conference on Computational Social Science 2016
Efficiency in Prediction Markets, Evidence from SciCast.orgCollective Intelligence 2015