Introduction:
Raia Hadsell is the Director of Research on Robotics at DeepMind for around eight years. She emphasizes the application of modern AI methods to solve diverse problems, from scientific to historical, and their future impact.

Alan Turing’s Vision:
Hadsell refers to Turing’s article titled “Intelligent Machinery” from 1948 which was unpublished. Turing speculated on building an intelligent machine using the analogy of an infant cortex, emphasizing the importance of training such a machine with data and experience. He believed that crucial components for this intelligent machine would be memory, sensory inputs, and feedback mechanisms such as rewards or punishments.

This early perspective closely resembles our modern understanding of AGI, even after 70 years.

Neuroscience’s Role in AI:
Turing believed understanding human intelligence is vital for constructing intelligent machines. Over recent decades, significant progress has been made in understanding intelligence across species. Mapping of the human genome provides insights and has more to offer.

Study of grid cells and spatial navigation unveils how humans comprehend their location and navigate. The visual cortex has been decoded to better understand human visual perception, with the study greatly benefiting from advances in computer vision. There’s a synergetic relationship between neuroscience and AI.

Computational Evolution:
Turing felt the lack of advanced computation during his time. Moore’s Law, as postulated by Gordon Moore in 1975, predicted that the number of transistors and resistors on a chip would double every 24 months. This prediction has largely held true for almost 50 years.

As a result of Moore’s Law, we possess vast distributed compute systems that don’t just serve entertainment but enable large-scale AI simulations and training.

AI Methods Today:
The methodologies for constructing and training intelligent machines have evolved considerably since Turing’s era.

Resurgence of Neural Networks and Backpropagation:
Emphasis on gradient-based optimization techniques. First appearance in the 1980s, went out of fashion, then returned due to scaling with computational power and Moore’s law.

Importance of Bayesian Approaches:
Use of Bayesian probability theory to understand what AI models can and cannot do. Allows for the assessment of model uncertainty, enhancing human interaction and providing better answers.

Advancements in Reinforcement Learning:
Existing for decades, but recent coupling with neural networks has led to new solutions. Utilizing positive and negative feedback to train neural networks. Capable of solving complex problems and even achieving superhuman performance in some cases.

Return of Larger Neural Networks:
Over the last 10 years, an observable trend towards much bigger neural networks. Example: From Raia Hadsell’s first encounter with a 60,000-parameter network (Limit 5) to the 70 billion-parameter model (Chinchilla) released by DeepMind in 2022. Six-order of magnitude increase in parameters since Hadsell’s PhD in the early 2000s, reflecting a substantial growth in complexity and capabilities.

Evolution of Key Models:
Highlighting the progress from small-scale models to significant breakthroughs:

Limit 5: Used to identify handwritten digits, containing 60,000 parameters.

AlexNet: Identified 1,000 object types in images, with 62 million parameters.

Chinchilla: Language generation model by DeepMind with 70 billion parameters, reflecting the rapid and monumental increase in scale.

These points underline the dynamic evolution, resurgence, and unprecedented growth in the field of neural networks, emphasizing the significance of understanding model limitations, reinforcement learning, and the continual trend toward larger, more complex models.

Introduction to the Challenge:
Turing’s Perspective: Exciting era for understanding what an intelligent machine can be. Aesthetic to Relevant Spectrum: Different examples ranging from aesthetic to relevant areas. Focus on Ancient Text Restoration: A case study using AI for understanding and restoring ancient texts.

Collaboration & Objective:
Collaborators: DeepMind, Google teams, Oxford, Athens University, University of Venice.

Core Problem: Understanding inscribed texts from the past, known as epigraphy.

Purpose: Inscriptions provide firsthand evidence for language, society, culture of ancient civilizations.

Challenges in Ancient Text Restoration:
Damaged Text: Many inscriptions are damaged, requiring text restoration.

Geographical Attribution: Finding the original location of stones that may have been moved or stolen.

Chronological Attribution: Determining the actual time of creation, as radio carbon dating doesn’t work on stone.

The Model: ISECA:
Input: Complete inscriptions fed as a stream of characters.

Processing: Neural network predicts missing characters, geographic attribution, and chronology.

Multitask System: The three problems are intertwined, supporting each other.

Interpretability: Multiple possible words and heat maps showing context importance for answers.

Results & Effectiveness:
Ithaca vs. Previous Models: Outperforms previous works and even human experts.

Human + Model Collaboration: Combining human experts with the model yields better results than either alone.

Future Applications: Applicable to any discipline dealing with ancient texts, in any language.

Overall:
AI as Tools: Advocating the use of AI as tools for various professionals, such as doctors, mathematicians, scientists.

General Method: Can be adapted for different types of artifacts, including potential forgeries or newly discovered items.

Future Impact: Anticipating continued impact on the field of ancient epigraphy.

Nuclear Fusion’s Importance:
Nuclear fusion as a potential energy source. Aston’s 1920s observation: Four hydrogen atoms have more mass than a helium atom. Inference: Fusion inside suns emits energy; hydrogen atoms fuse into helium, emitting a neutron and energy.

History of Fusion Devices:
1950s: Conceptualization of the Tokamak.

Tokamak: Magnetic confinement device to control nuclear fusion experiments. Today, it’s the most viable and practical approach to nuclear fusion energy.

Plasma Fusion Basics:
Fusion requires pushing nuclei together. Tokamak’s function: Creates heat and pressure to produce plasma (fourth state of matter) for fusion. Magnetic coils both inside and outside its toroidal shape control the plasma. Stable plasma maintenance is vital for energy extraction.

Plasma Control Mechanism:
Inside view of Tokamak resembles a game of controlling plasma’s position and shape using magnetic coils. 19 control buttons and 92 observational metrics aid in controlling the plasma, much like trying to keep a balloon afloat using air jets without physical touch.

Robotic Perspective:
Approach from a robotics angle given the parallels between controlling multiple joints in robots and plasma in a Tokamak. Used an algorithm called MPO, a reinforcement learning neural network. The algorithm’s goal: Maximize rewards based on three criteria related to plasma control.

Real-life Implementation:
Extensive simulation testing due to initial access restrictions to the real Tokamak. Successfully controlled the Tokamak with a deep reinforcement learning network. Demonstrated the ability to create stable configurations known in theory but challenging to stabilize in a real reactor.

Contribution to the Field:
Collaboration with scientists at EPFL opens up new toolsets for nuclear fusion research. Small neural network was used, emphasizing the power of the algorithm and its learning method. Neural networks provided a distinct advantage over traditional control optimization methods.

Significance of Weather Forecasting:
Weather has a broad variety that impacts our lives. Forecasting serves as the foundation for understanding climate. Predicting short-term weather can help us project long-term climatic changes.

Collaboration with the Met Office:
DeepMind’s best projects often involve collaborations with domain experts. Partnership with the UK Met Office aimed to address the challenge of precipitation nowcasting.

Nowcasting:
Focuses on predicting rainfall amounts in the next 60 to 90 minutes for a specific region. Despite effective forecasting for extended periods, current models struggle with short-term predictions.

Importance extends beyond daily planning to critical activities like issuing flood warnings and air traffic control.

Challenges and Importance:
Requires high accuracy. Must account for uncertainties and provide probabilistic outcomes. Capturing rare events, like catastrophic storms, is paramount due to their potential impact.

DeepMind’s Approach:
Treated radar data as video streams, analyzing “frames” that show rain patterns. Each pixel corresponds to one square kilometer on the ground. Used the data to train models on video prediction by providing partial information and asking the model to forecast subsequent frames. Desired a model that offers multiple potential weather outcomes for the upcoming 1.5 hours.

Conditional Generative Adversarial Networks (cGAN):
Used cGANs to predict future weather based on initial conditions. cGANs involve a generator (predicts future conditions) and a discriminator (evaluates the quality of those predictions). This adversarial process improves prediction accuracy. While GANs are infamously known for creating “deep fakes”, their underlying technology can be beneficial, as in this weather forecasting application.

Practical Application:
A difficult forecasting scenario was independently chosen by the Met Office’s chief forecaster. The model’s output was compared against traditional forecasting methods and evaluated by meteorologists.

Goal: to create a tool useful for experts, not just an academic exercise.

Observations on Traditional Systems:
The conventional forecasting system used by the Met Office displayed significant discrepancies even within 30 minutes of the ground truth. Traditional methods may mispredict or exaggerate the intensity and location of weather phenomena.

This encapsulation provides a concise summary of Raia Hadsell’s segment on DeepMind’s innovative approach to weather forecasting, emphasizing the potential of AI techniques in addressing complex real-world challenges.

Connection to Science Fiction:
Raia Hadsell references Douglas Adams’ sci-fi solution to interstellar communication: the “beable fish”. This fictional creature, when placed in the ear, could translate any language instantly. Such concepts, once considered pure fantasy, now border on reality due to advances in AI.

Current State of Machine Translation:
Modern AI-driven methods can now translate between languages in real-time, nearly as quickly as one can speak. Although these tools aren’t yet on par with expert human translators, their proficiency is adequate for many practical applications.

Historical Context:
Over the past decade, there has been a significant revolution in the field of machine translation. Originally, translation systems were highly modular, involving separate processes for understanding, parsing, structural transfer, target generation, and more.

Shift in Approach:
In recent years, the trend has shifted towards training large neural networks to handle the entire translation process internally. Humans don’t translate by compartmentalizing each aspect of the process; similarly, these AI systems are designed to learn from vast amounts of data, achieving an end-to-end translation.

Introduction of WaveNet:
While initial machine translation primarily focused on text, the desire grew for AI to audibly speak translations. Early voice outputs sounded robotic. This led to the development of WaveNet. WaveNet is designed to generate raw audio. It predicts and produces audio waveforms millisecond by millisecond. The technology was groundbreaking as it could generate audio that was close to human voice quality, given the right training.

Practical Demonstration:
Hadsell showcased a real-world application by translating an English sentence about avocados into French, and then potentially Danish. The system was not only able to translate but also audibly speak the translations.

Implications:
These advancements in machine translation blur the line between science fiction and reality. While the technology is impressive, it might inadvertently discourage the younger generation from learning new languages, as mentioned with Hadsell’s son. The technology might not be flawless, but its proficiency and utility are undeniable.

Language Models as Future of Search:
Definition and Training: Language models are trained to predict the next word in a sequence. Example given is ‘Chinchilla’ with 70 billion parameters, trained on all English text on the web.

Capabilities: These large models are now capable of having meaningful conversations, accessing a vast amount of data they have been trained on.

Application in Search Engines: They will become the next step in search, allowing users to talk to an ‘expert’ in various fields. Users will be able to ask complex questions, with the model providing in-depth expertise.

Challenges: Ensuring information is verifiable and correct, and providing explanations for answers.

Impact on Society: Democratisation of knowledge, as everyone will have access to this expertise, just as they now have access to basic facts via search engines. It’s deemed important for democracy.

Robotics:
Current Limitations: Robotics is challenging and has not transformed as quickly as other AI-driven fields. Robots are mainly restricted to factories and controlled environments.

Integration into Human Spaces: Robots will start to enter human spaces, supporting us in various tasks.

Construction: They can aid in dangerous, hard, and repetitive tasks.

Agriculture: Similar impacts expected in supporting human labor.

Dealing with Waste: Vision of robots that can deal with garbage and recycling, akin to the scenario in the movie Wally.

Requirement for General Robots: Need for robots that are general enough to interact with and support humans.

DeepMind’s Approach and Responsibility:
Collaborative Mindset: DeepMind works responsibly, thinks carefully about its approach, and collaborates with other experts. Responsibility in Development: Emphasizes that responsibility is paramount in developing future tools and technologies.

Turing’s Thoughts on Training Systems:
Turing contemplated how reward or punishment feedback could be used in system training. In his 1948 article, “Intelligent Machinery”, he discusses using reward or punishment to select or suppress outputs after introducing random inputs. The concept of gradient descent optimization as we know it wasn’t part of Turing’s discussions; it was more about selecting the best outcomes after exploration.

Weather Forecasting with Neural Networks:
Neural networks can be used to forecast weather by sidestepping the traditional dynamical system challenges associated with partial differential equations. One approach combines well-understood physical systems with data learned from observations to produce better weather predictions. The more effective method simply uses data, treating it as a problem of inputting and outputting pixels, akin to predicting video frames. The “video” in this context is layers of radar information, such as precipitation data moving across the Earth.

Neural networks, with the power of substantial data, can be more accurate than traditional methods when predicting short-term weather changes. The neural network used for this purpose might consist of approximately 20 layers and possibly up to a billion parameters. For longer-term weather predictions, traditional numerical methods perform better, but for short durations (an hour or two), the neural network-based method proves superior.

Theoretical Foundations of AI Models:
There is more work needed on foundational questions about AI. AI models are not solely reliant on the number of parameters or amount of data. Fundamental areas include understanding optimization, avoiding underfitting and overfitting, and the capacity of networks. There’s never a conclusive understanding of AI, both theoretically and empirically.

Consciousness and AI:
Defining consciousness is challenging. Consciousness, from a neuroscience perspective, involves awareness of the past and predicting the future. Intelligent animals, like elephants, are recognized by their long memory and ability to predict the future to guide actions. Consciousness isn’t binary but rather a spectrum of awareness and decision-making capabilities. Intelligent machines will need to have memory and prediction abilities, echoing Alan Turing’s thoughts on memory.

Responsible Use and Governance of AI:
Deepfakes are highlighted as a potentially harmful use of technology. Technology has both good and bad uses; responsibility is critical. Robotics, as a dual-use technology, can be used for rescue or harm. Governance of AI involves multiple layers: regulation, legislation, education, and internal corporate guidelines. Large companies like DeepMind, Google, and Meta need to prioritize responsible AI use. The approach shouldn’t be to halt technology but to understand and mitigate risks.

Role of Judgment in AI:
Some AI training processes, especially on large data, don’t involve humans due to the sheer scale and complexity. However, some algorithms, like reinforcement learning, involve human interaction during training. There’s value in humans interacting with AI technologies, especially in guiding the direction of robotic training.

Human Interaction with AI Technologies:
Not all AI models require human input during training, especially when training on vast datasets. Reinforcement algorithms can benefit from human interaction during the training process. Human-guided robotic training is an emerging and beneficial area in AI research.

Data sparsity challenge:
AI systems often rely on large amounts of data for training. There are scenarios where the desired data is limited or almost non-existent.

Simulation as a solution:
Simulation can provide initial data for training AI systems. Systems can start with simulated data before incorporating real-world data.

Data augmentation:
Enhances limited real data by creating variations of it. Example: To train a model to identify a specific storm seen only twice in three years, one can adjust the storm data slightly (e.g., move or adapt it) to create more instances of it in the dataset.

Importance of sufficient data:
High performance in AI models typically necessitates ample data. While tools like simulation and data augmentation are helpful, they aren’t magic solutions; substantial data is still a prerequisite for optimal performance.