Why Hinton Was Wrong, Causal AI & Science | Thanos Vlontzos Ep 15 | CausalBanditsPodcast.com
May 6, 2024
auto_awesome
Athanasios (Thanos) Vlontzos, a Research Scientist at Spotify's Advanced Causal Inference Lab, tackles intriguing questions about AI's future and causal modeling. He discusses why many AI predictions miss the mark and explores the evolving role of radiologists amid AI advancements. Thanos dives into challenges in medical AI, the humor of causal model pitfalls, and the essence of interdisciplinary collaboration. The conversation also highlights the connection between music and ideas, emphasizing the drive for exploration in science.
Causality should permeate every stage of the research process, not just the modeling phase, to enhance scientific accuracy and effectiveness.
The limitations of deep learning in medical roles highlight the irreplaceable value of human expertise in nuanced decision-making scenarios like radiology.
Interdisciplinary collaboration fosters creativity and innovation, as diverse teams draw on varied expertise to tackle complex challenges in causal AI.
Deep dives
The Nature of Causality in Science
Science is described as a non-linear process characterized by exploration, challenging the notion that causality exists solely within the modeling phase. The discussion emphasizes that causality must be considered throughout the entire research process, from data collection to model implementation. The guest remarks on the limitations of deep learning in replicating the nuanced decision-making inherent in medical roles like radiology, where human expertise plays an irreplaceable role. This sets the stage for a deeper contemplation of causality that transcends mere analytical modeling.
Challenges in Medical Imaging
The podcast underscores the fundamental challenges present in the medical field, specifically regarding data integrity and model reliability. The guest highlights historical biases in datasets, such as during the COVID-19 pandemic, where x-ray data was skewed due to geography, leading to ineffective models. Moreover, the complexities of imaging modalities such as MRIs and ultrasounds introduce further complications that require careful consideration of causality in modeling. The conversation posits that frameworks must not only account for data variability but also incorporate causal reasoning for more effective medical tools.
Importance of Technology Readiness Levels
The discussion touches upon technology readiness levels (TRL) as a crucial framework for assessing the maturity of technologies before their implementation in critical fields like medicine. The guest recalls historical applications of TRLs within military contexts and their relevance in the development of ML algorithms for medical imaging. Skipping stages within this framework can lead to problems in accountability and effectiveness when deploying AI in real-world scenarios. Hence, the importance of iterative checks and balances in technology development is emphasized to ensure ethical application in medicine.
Causal Discovery vs. Causal Inference
Causal discovery is presented as a more challenging and impactful pursuit than causal inference, as it aims to uncover unknown causal relationships rather than relying on established assumptions. This aspect is illustrated through examples from fields like medicine and physics, where the quest for causality often leads to groundbreaking insights. The guest highlights the appeal of causal discovery due to its exploratory nature, which aligns with the intrinsic human desire for understanding. By framing it as an essential scientific endeavor, the conversation drives home the idea that uncovering new causal relationships fuels innovation.
Interdisciplinary Collaboration in Research
The podcast emphasizes the value of interdisciplinary collaboration, particularly in the context of the Advanced Causal Inference Lab at Spotify. The guest shares that the diversity of expertise—from engineering to astrophysics—on their team fosters creativity and drives innovative solutions. This collaborative approach draws parallels to historical instances where diverse teams succeeded in their endeavors, proving that varied perspectives can yield significant advancements. The conversation reinforces the message that breaking down silos between disciplines ultimately enhances problem-solving capabilities and enriches scientific exploration.
What makes so many predictions about the future of AI wrong?
And what's possible with the current paradigm?
From medical imaging to song recommendations, the association-based paradigm of learning can be helpful, but is not sufficient to answer our most interesting questions.
Meet Athanasios (Thanos) Vlontzos who looks for inspirations everywhere around him to build causal machine learning and causal inference systems at Spotify's Advanced Causal Inference Lab.
In the episode we discuss: - Why is causal discovery a better riddle than causal inference? - Will radiologists be replaced by AI in 2024 or 2025? - What are causal AI skeptics missing? - Can causality emerge in Euclidean latent space?
Ready to dive in?
About The Guest Athanasios (Thanos) Vlontzos, PhD is a Research Scientist at Advanced Causal Inference Lab at Spotify. Previousl;y, he worked at Apple, at SETI Institute with NASA stakeholders and published in some of the best scientific journals, including Nature Machine Learning. He's specialized in causal modeling, causal inferernce, causal discovery and medical imaging.
