Interview
Regina Barzilay: Deep Learning for Cancer Diagnosis and Treatment | Lex Fridman Podcast #40
- Regina Barsley is an MIT professor specializing in natural language processing (NLP), deep learning applications in chemistry, and oncology.
- Barsley cites The Emperor of All Melodies as a pivotal book that revealed the imperfect, non-linear nature of scientific discovery and the necessity of human devotion to implement ideas.
- The fiction book Americana provided Barsley with a lens to understand cultural adaptation and the social tendency to misinterpret communication barriers as disabilities.
- Barsley argues that while ideas are fundamental, "local" adoption of new scientific paradigms is often driven by personalities and their advocacy rather than the intrinsic superiority of the idea.
- She notes that statistical approaches in NLP took decades to become mainstream due to entrenched skepticism from prominent figures who resisted the paradigm shift.
- Barsley highlights that modern drug development grew from the 19th-century dye industry, where chemists observed molecular effects on cells, illustrating the serendipitous and imperfect origins of medical breakthroughs.
- Barsley contrasts computer science's pattern-matching approach (e.g., Amazon recommendations) with biology's heavy emphasis on mechanistic understanding, suggesting the latter may be too complex for deterministic models to solve entirely.
- She advocates for a "probabilistic matching" approach in medicine for early diagnostics, parallel to how statistical models work in computer science, rather than solely pursuing deep mechanistic understanding.
- Barsley was diagnosed with breast cancer at age 43 in 2014, a personal experience that fundamentally shifted her perspective on the triviality of incremental AI research improvements.
- Following her diagnosis, Barsley re-evaluated her research priorities, concluding that limited time necessitates focusing on work that alleviates real-world suffering rather than improving parsers by 2-3%.
- She estimates there are 1.7 million new cancer cases and 600,000 cancer-related deaths annually in the US, emphasizing the scale of suffering that AI must address.
- Barsley believes machine learning will enable earlier cancer prediction and faster drug discovery but remains skeptical about the speed at which regulatory bodies will adopt these technologies.
- Early detection is identified as crucial; for instance, pancreatic cancer has a few percent survival rate once detected, but early discovery through scans (even of unrelated issues) can be life-saving.
- Barsley notes that 80% of breast cancer patients are the first in their families, rendering current simplistic statistical risk models ineffective for identifying susceptible individuals.
- Machine learning can improve risk assessment by synthesizing weak signals from diverse data sources like imaging and liquid biopsies, which human eyes cannot discriminate.
- A major barrier to advancing cancer AI is the lack of publicly available, modern medical datasets; existing datasets like the Florida Dataset (film mammograms from the 90s) are obsolete.
- Data access is currently restricted by hospital policies, legal liability, and a complex IRB approval process, unlike the open-access culture of computer vision datasets like ImageNet.
- Barsley proposes a future model where patients voluntarily donate data for research, similar to organ donation, to facilitate the creation of large-scale, diverse medical datasets.
- She suggests that while technical solutions for data privacy (e.g., encoding data before processing) exist, the primary hurdle is societal trust and the lack of incentive for hospitals to share data.
- Barsley critiques the current breast cancer density assessment law (2019), noting it classifies 40-50% of women as high-risk despite weak underlying science, a standard adopted over 15 years due to patient advocacy.
- Deep learning models can already predict breast cancer risk more accurately than the 1967 density heuristic, but adoption is stalled by regulatory and anthropological barriers rather than algorithmic limitations.
- Barsley identifies drug design as a critical, technically underutilized area for ML, currently relying on human experts' domain knowledge for high-throughput screening.
- She explains that small molecules can be represented as graphs (atoms as nodes, bonds as edges), making graph generation a unique and exciting frontier for machine learning innovation.
- Early ML efforts in drug design used handcrafted features, but Barsley's team has progressed to using encoder-decoder models to modify molecular structures for improved bioactivity and reduced toxicity.
- Labs are already manufacturing molecules generated by these ML models, marking the transition from theoretical property prediction to physical drug synthesis.
- Barsley entered the NLP field in 1997 during a transition from rule-based, linguistically heavy approaches to statistical, corpus-based methods.
- She observes that the field of NLP has largely abandoned formal linguistic structures (e.g., syntactic trees) in favor of data-driven statistical methods that prioritize functional outcomes over explainability.
- Machine translation has seen massive success in recent years, evolving from "impossible" to a daily utility, validating the power of deep learning even without human-like understanding.
- Barsley maintains that while ML can achieve high performance in specific tasks, it remains brittle to distributional shifts and fails in contexts outside its training distribution, such as translating obscure Finnish recipes.
- She argues that passing a high-quality Turing test (sustaining a one-hour conversation) is currently difficult due to data limitations and a lack of compositional generalization in training systems.
- Citing the historical story of "Eliza" at MIT, Barsley suggests that the success of chat interfaces depends less on technological sophistication and more on the human psychological readiness to attribute understanding to machines.
- Barsley is optimistic about the potential for human-level or super-human-level intelligence, viewing it as an evolution of functionality across many domains rather than a singular definition of intelligence.
- She draws parallels to the book Flowers for Algernon, envisioning a future of cognitive augmentation where technology accelerates human thinking and perception.
- Barsley expresses hope for brain-computer interfaces (BCIs) like Neuralink, specifically for monitoring attention and providing real-time feedback to modify behavior and improve focus.
- She warns that direct biometric feedback loops can lead to addictive behaviors, citing her own experience with a gym app that compelled her to run while injured to maintain a status metric.
- Barsley introduced a new MIT course, "Machine Learning from Algorithms to Modeling," designed for non-majors, addressing the mathematical barriers (linear algebra, probability) that exclude many interested students.
- She advises students to find a specific domain they care about (e.g., curing cancer, self-driving cars) where data exists and strong patterns can be leveraged, rather than chasing abstract algorithms.
- Barsley reflects that personal purpose and meaning are often obscured by external validation, and she encourages individuals to define their own "mission" independent of the academic crowd.