2022 Breakthrough Technologies

Artificial Intelligence for Protein Folding

Our human body requires proteins for nearly everything we do, and the way a protein folds determines its activity. However, determining proteins' structure can take months.

Solution: #AlphaFold2 . AlphaFold2 is an open-source #AI system developed by #DeepMind that predicts a protein's 3D structure from its amino acid sequence. Knowing this data should make it more efficient in designing drugs for a wide range of diseases.

Synthetic Data for Artificial Intelligence

#MachineLearning models require large sets of training, validation and testing datasets. These datasets are generally messy, reflect real-world biases, susceptible to privacy concerns, and/or are resources intensive to acquire.

Solution: #SyntheticData for AI. Though not perfect, a move in a viable and desirable direction.

AI-Generated Art

Art creation is a hobby and a career for some. For some, the art collection is a passion.

Meet #StableDiffusion - a new open-source AI art generator that's built on a #DeepLearning text-to-image model. The tool, along with other popular image-generation AI models, allows anyone to create impressive images based on text prompts.

CICERO

#CICERO is the first AI by #Meta to achieve human-level performance in the strategy game?#Diplomacy . It uses an interface between #StrategicReasoning and #NaturalLanguage . At each point in the game, CICERO looks at the game board and its conversation history, and models how the other players are likely to act. It then uses this plan to control a language model that can generate free-form dialogue, informing other players of its plans and proposing reasonable actions for the other players that coordinate well with them. In doing such, CICERO converses simultaneously with six other players, sending 100s of messages over the course of the game - forming alliances and building relationships with other players, communicating with empathy and building rapport with the players, while tying back to the strategic goal. The strategy informs communication and communication informs the strategy.

In the development of this?#IntelligentAgent , the team started with a 2.7 billion parameters?#BART -like language model pre-trained on text from the internet and fine-tuned it on over 40,000 human Diplomacy games on?#webDiplomacy .net. The team also developed techniques to automatically annotate messages in the training data with corresponding planned moves in the game so that at inference time dialogue generation can be controlled to discuss specific desired actions for the agent and its conversation partners. Why control? Controlling generation in this manner allows CICERO to ground its conversations in a set of plans that it develops and revises over time to better negotiate. This helps the agent coordinate with and persuade other players more effectively.

While there are still CICERO optimization opportunities, what's exciting is the business and practical opportunities this breakthrough has the potential to unlock. Imagine the ability to connect with an adaptive intelligent agent, via a chat (#Chatbot ) or an?#Avatar , in the physical world or the?#metaverse , and affectively converse - converse with an agent who empathizes with you, converses for extended periods, builds an affective relationship, and efficiently helps you achieve your objective consistently during each interaction. Compare this to the status quo of the chatbots, avatars and other natural language agents that have very limited and transactional capabilities e.g.?https://lnkd.in/g9JNQifC . Imagine the value this?#breakthrough ?can unlock for the customers and consumers, complementing call center agents!

Imitation Learning

Online videos are a vast and untapped source of training data. However, it requires extensive resources to annotate the video frames by hand, not to mention the efficacy - defects of manual annotation.

Solution: Video Pre-Training (#VPT ) approach. The team at #OpenAI leveraged crowd-workers to play Minecraft, and recorded their keyboard and mouse clicks alongside the video from their screens. This gave the researchers 2,000 hours of annotated Minecraft play, which they used to train a model. They then used this model to generate action labels for 70,000 hours of unlabelled video taken from the internet and then train the Minecraft bot on this larger dataset.