Dr. Terry Sejnowski: How to Improve at Learning Using Neuroscience & AI

Computational neuroscientists, like Dr. Terry Cignowski, are employing mathematical models and advanced computing methods to deepen our understanding of brain function, focusing on how neural circuits correspond to brain activities.

Computational neuroscientists use mathematical models and algorithms to understand how the brain works as a system.

Terry Sejnowski and Dr. Terry Cignowski highlight how the brain operates at multiple levels, stretching from the molecular to the entire nervous system. To unravel brain functionality, they emphasize the importance of understanding the intermediate algorithmic level, which lies between the organic implementation within neural circuits and the resulting behaviors of the whole system.

Sejnowski compares algorithms to recipes that guide neural circuit function. With the advent of new methods like optical neuron recording, researchers can now observe interactions across various brain areas, gaining insight into the cognitive behaviors these complex algorithms direct.

Sejnowski also notes that past neuroscience often isolated functions to different parts of the cortex, akin to dividing it into countries with specific roles. This idea intersects with the current approach of segmenting brain functions at different levels to better understand the systemic operation.

The basal ganglia use a simple algorithm based on predicting future rewards to learn sequences of actions that achieve goals.

Sejnowski introduces the notion that our brains use specific algorithms, such as temporal difference learning, a form of reinforcement learning, for predicting and maximizing future rewards. This approach is fundamental to both fly brains and human brains and is employed by the basal ganglia to learn action sequences that lead to desired outcomes.

This trial-and-error reinforcement learning algorithm, exem ...

Exploring the neuroscience of learning, motivation, and skill acquisition reveals the complexity and adaptability of the human brain. Dr. Terry Sejnowski, alongside Andrew Huberman, delves into the ways our brains process and adapt to new information, and the implications for learning across our lifespan.

The brain has two main learning systems - a cortical cognitive system and a subcortical procedural system, which work together to build knowledge and skills.

Procedural learning through practice and repetition is essential for developing expertise, while cognitive learning provides the underlying concepts and theories.

Terry Sejnowski describes the basal ganglia as a crucial brain region for procedural learning, which includes learning sequences of actions to achieve goals, such as serving in tennis. This form of learning takes over from the cortex to produce actions that get progressively better with practice and is evidential across various fields, including finance and neuroscience. The basal ganglia interact with the cortex's action part and with the prefrontal cortex involved in thinking. The brain utilizes an algorithm that predicts the next reward and updates synapses to increase the chances of attaining a reward next time, hence building what's known as a value function. This value function is nourished through experiences, such as dining at varied restaurants, with the cortex developing knowledge of beneficial and detrimental things.

Andrew Huberman and Terry Sejnowski highlight the importance of active learning, which allows the brain to solve problems by trial and error rather than simply memorizing facts. Sejnowski emphasizes the critical role of practice, especially for rapid and expert execution, strengthening his point by referring to his physics education where homework and problem-solving embody procedural learning.

As we age, procedural learning becomes more effortful, but can be supported by techniques like the "learning how to learn" course developed by the speaker.

Maintaining physical, cognitive, and social activity helps preserve brain health and cognitive function as we get older. As people age, procedural learning becomes more challenging due to a thinning cortex, even though older memories remain robust. The "learning how to learn" course, targeted at adults in the workforce, is designed to boost learning efficiency, highlighting the necessity of enhancing learning methods beyond traditional education.

Dr. Sejnowski and colleagues have developed an online portal to teach individuals to learn better based on their learning styles. Additionally, techniques like physical exercise, as Sejnowski practices, are crucial for enhancing cognition and abetting better learning.

Sleep's role in memory consolidation is si ...

Andrew Huberman, Terry Sejnowski, and others discuss the impressive potential of artificial intelligence (AI) and large language models (LLMs) to assist in the advancement of scientific research, from computational neuroscience to uncovering medicinal breakthroughs.

Large language models like ChatGPT can be used as "idea pumps" to generate novel hypotheses and experimental designs based on existing knowledge.

AlphaGo’s victory and the applications of LLMs in research demonstrate AI’s prowess in complex problem-solving. Sejnowski describes AI's capacity to generalize knowledge to solve new problems, which is crucial for scientific advancement. Huberman talks about AI foraging for knowledge based on future scenarios, potentially leading to the generation of novel hypotheses and designs. Sejnowski's colleague at the Salk Institute, Rusty Gage, uses LLMs as "idea pumps," inputting data and asking for experimental ideas. Sejnowski and Huberman note LLMs can handle and extrapolate from existing information, perform in-context learning, and surprisingly get better at tasks without undergoing further learning.

These models excel at tasks like summarizing research papers, critiquing statistical methods, and extrapolating new scenarios from available data.

Sejnowski recounts giving an LLM a neuroscience abstract and asking it to explain the content to a 10-year-old, showcasing the model's ability to simplify complex information, albeit with the loss of finer intricacies. Huberman mentions AI that can reference sources, analyze them critically, and predict outcomes. Sejnowski mentions Google's Gemini, which uses "chain of reasoning" to improve problem-solving accuracy by following logical steps.

Integrating the strengths of human experts and AI systems can lead to breakthroughs that neither could achieve alone.

AI can uncover patterns and connections that humans may miss. Sejnowski proposes that AI models could predict future events by integrating AI's predictive capabilities with human expertise. AI can quickly synthesize existing data, potentially hypothesizing outcomes faster than possible through traditional methods. Although AI can generate data and ideas, it doesn't necessar ...