Method prevents an AI model from being overconfident about wrong answers

MIT News  July 31, 2024 Recent studies have found that common interventions such as instruction tuning often result in poorly calibrated large language models (LLMs). Although calibration is well-explored in traditional applications, calibrating LLMs is uniquely challenging. The challenges stem as much from the severe computational requirements of LLMs as from their versatility, which allows them to be applied to diverse tasks. To address these challenges, researchers at MIT proposed THERMOMETER, a calibration approach tailored to LLMs. For calibrating the LLMTHERMOMETER learned an auxiliary model, using the data given from multiple tasks. According to the researchers it was computationally efficient, […]

Doing more but learning less: Addressing the risks of AI in research

Phys.org   March 8, 2024 Scientists are enthusiastically imagining ways in which artificial intelligence (AI) tools might improve research. A team of researchers in the US (Yale University, Princeton University) developed a taxonomy of scientists’ visions for AI, observing that their appeal comes from promises to improve productivity and objectivity by overcoming human shortcomings. But proposed AI solutions can also exploit our cognitive limitations, making us vulnerable to illusions of understanding in which we believe we understand more about the world than we actually do. Such illusions obscure the scientific community’s ability to see the formation of scientific monocultures, in which […]

AI just got 100-fold more energy efficient

Science Daily  October 12, 2023 Although support vector machine algorithms for electrocardiogram classification show high classification accuracy, hardware implementations for edge applications are impractical due to the complexity and substantial power consumption needed for kernel optimization when using conventional CMOS circuits. A team of researchers in the US (Northwestern University, University of Southern California) has shown that reconfigurable mixed-kernel transistors based on dual-gated van der Waals heterojunctions could generate fully tunable individual and mixed Gaussian and sigmoid functions for analogue support vector machine kernel applications. The heterojunction-generated kernels can be used for arrhythmia detection from electrocardiogram signals with high classification […]

AI models struggle to identify nonsense, says study

Phys.org   September 14, 2023 Neural network language models appear to be increasingly aligned with how humans process and generate language, but identifying their weaknesses through adversarial examples is challenging due to the discrete nature of language and the complexity of human language perception. An international team of researchers (USA – Columbia University, Israel) turned the models against each other by generating controversial sentence pairs where two language models disagreed about which sentence is more likely to occur. Considering nine language models (including n-gram, recurrent neural networks, and transformers), they created hundreds of controversial sentence pairs through synthetic optimization or by […]

Ensuring AI works with the right dose of curiosity

MIT News  November 10, 2022 To address the challenge of exploration, incentivizing the agent to visit novel states using an exploration bonus can lead to excellent results on hard exploration tasks but can suffer from intrinsic reward bias and underperform when compared to an agent trained using only task rewards. An international team of researchers (USA – MIT, Finland) has proposed a principled constrained policy optimization procedure that automatically tunes the importance of the intrinsic reward: it suppresses the intrinsic reward when exploration is unnecessary and increases it when exploration is required. According to the researchers this resulted in superior […]

Researchers Say It’ll Be Impossible to Control a Super-Intelligent AI

Science Alert  September 18, 2022 Considering recent advances in machine intelligence, several scientists, philosophers, and technologists have revived the discussion about the potentially catastrophic risks entailed by such an entity. An international team of researchers (Spain, Germany, USA – UC San Diego, Chile) traced the origins and development of the neo-fear of superintelligence, and some of the major proposals for its containment. They argue that total containment is, in principle, impossible, due to fundamental limits inherent to computing itself. Turing’s halting program centers on knowing whether a computer program will reach a conclusion and answer (so it halts), or simply […]

An AI Just Independently Discovered Alternate Physics

Science Alert  July 29, 2022 Despite the prevalence of computing power and artificial intelligence, the process of identifying the hidden state variables themselves has resisted automation. Most data-driven methods for modelling physical phenomena still rely on the assumption that the relevant state variables are already known. A longstanding question is whether it is possible to identify state variables from only high-dimensional observational data. Researchers at Columbia University proposed a principle for determining how many state variables an observed system is likely to have, and what these variables might be. They demonstrated the effectiveness of this approach using video recordings of […]

Scientists Have Created an AI That Can Think Like a Human Baby

Science Alert  July 11, 2022 Current artificial intelligence systems pale in their understanding of intuitive physics, in comparison to even very young children. An international team of researchers (USA – Princeton University, UK) addressed this gap between humans and machines by drawing on the field of developmental psychology. They introduced and open-sourced a machine-learning dataset designed to evaluate conceptual understanding of intuitive physics, adopting the violation-of-expectation (VoE) paradigm from developmental psychology. Then they built a deep-learning system that learns intuitive physics directly from visual data, inspired by studies of visual cognition in children. They demonstrated that their model could learn […]

AI Improves Robotic Performance in DARPA’s Machine Common Sense Program

DARPA News  June 22, 2022 A team of researchers in the US(UC Berkeley, Oregon State University, University of Utah, University of Washington) working on DARPA’s Machine Common Sense (MCS) program demonstrated a series of improvements to robotic system performance over the course of multiple experiments. Just as infants must learn from experience, MCS seeks to construct computational models that mimic the core domains of child cognition for objects (intuitive physics), agents (intentional actors), and places (spatial navigation). Using only simulated training, recent MCS experiments demonstrated advancements in systems’ abilities – ranging from understanding how to grasp objects and adapting to […]

The promise and pitfalls of artificial intelligence explored at TEDxMIT event

MIT News  January 11, 2022 MIT scientists discuss the future of AI with applications across many sectors, as a tool that can be both beneficial and harmful. They said some new capabilities that could be enabled by AI: an automated personal assistant that could monitor your sleep phases and wake you at the optimal time, as well as on-body sensors that monitor everything from your posture to your digestive system. Intelligent assistance can help empower and augment our lives. But these intriguing possibilities should only be pursued if we can simultaneously resolve the challenges that these technologies bring…read more.