Although instruction tuning is widely used to adjust behavior in Large Language Models (LLMs), extensive empirical evidence and research indicates that it is primarily a process where the model fits to specific task formats, rather than acquiring new knowledge or capabilities. We propose that this limitation stems from biased features learned during instruction tuning, which differ from ideal task-specfic features, leading to learn less underlying semantics in downstream tasks. However, ideal features are unknown and incalculable, constraining past work to rely on prior knowledge to assist reasoning or training, which limits LLMs' capabilities to the developers' abilities, rather than data-driven scalable learning. In our paper, through our novel data synthesis method, DELIA (Diversity-Enhanced Learning for Instruction Adaptation), we leverage the buffering effect of extensive diverse data in LLMs training to transform biased features in instruction tuning into approximations of ideal features, without explicit prior ideal features. Experiments show DELIA's better performance compared to common instruction tuning and other baselines. It outperforms common instruction tuning by 17.07%-33.41% on Icelandic-English translation bleurt score (WMT-21 dataset, gemma-7b-it) and improves accuracy by 36.1% on formatted text generation (Llama2-7b-chat). Notably, among knowledge injection methods we've known, DELIA uniquely align the internal representations of new special tokens with their prior semantics.

Current trends to pre-train capable Large Language Models (LLMs) mostly focus on scaling of model and dataset size. However, the quality of pre-training data is an important factor for training powerful LLMs, yet it is a nebulous concept that has not been fully characterized. Therefore, we use the recently proposed Task2Vec diversity coefficient to ground and understand formal aspects of data quality, to go beyond scale alone. Specifically, we measure the diversity coefficient of publicly available pre-training datasets to demonstrate that their formal diversity is high when compared to theoretical lower and upper bounds. In addition, to build confidence in the diversity coefficient, we conduct interpretability experiments and find that the coefficient aligns with intuitive properties of diversity, e.g., it increases as the number of latent concepts increases. We conclude the diversity coefficient is reliable, show it's high for publicly available LLM datasets, and conjecture it can be used to build useful diverse datasets for LLMs.

The gaming industry is experiencing a rapid evolution in the field of artificial intelligence (AI), which has traditionally focused on improving computer-controlled opponents. However, the latest example of AI's expanding role in gaming is the creation of intelligent computer-controlled characters. Sony AI, the company's artificial intelligence research division, has partnered with PlayStation developers to create game AI agents that can be players' in-game opponents or collaboration partners. By using reinforcement learning, an area of machine learning where an AI teaches itself to act through trial and error, the characters will mimic human players and, to some extent, think. As open-world games become more complex and ambitious, developers must build systems capable of generating intelligent, reactive, creative characters and emergent side quests.

In 2012, Authors Björn Bloching, Lars Luck, and Thomas Ramge published In Data We Trust: How Customer Data is Revolutionising Our Economy. The book goes into detail about how a lot of companies have all the information they need at their fingertips. Companies no longer need to make decisions based on their gut feeling and the market, they can use streams of data to give them a better understanding of what the future looks like and what their next move should be. As the world of data, in particular, Artificial Intelligence continues to grow - more and more people are skeptical. Some may say that the use of data and autonomous features have improved our day-to-day lives.

What is the state of machine learning in 2022? Running a business that is closely tied to the progress of state-of-the-art machine learning means I’m trying to stay up to date with what is going on. In this post, I go through what I consider to be the most interesting breakthroughs and share my thoughts on what that means. We cover embeddings, attention, transformers and multi-modal models.

What are the most important ethical considerations for artificial intelligence (AI) in health care? The World Health Organization tried to answer this question in its recent report "Ethics and Governance of Artificial Intelligence for Health." It offers recommendations on how to design safe, transparent, and equitable AI products and applications that can help providers make informed medical decisions and help patients achieve positive outcomes. All of these are laudable. But as someone deeply involved in applying AI to health care, I found one element grating: Highlighting inclusivity and equality as things to be "encouraged" is not the way forward, especially with something as important as health care. Inclusivity and equality must be built into the DNA of a product from day one, and it does not happen by simply checking a politically correct box.

Blockchain and Artificial Intelligence are two of the hottest technology trends right now. Even though the two technologies have highly different developing parties and applications, researchers have been discussing and exploring their combination [6]. PwC predicts that by 2030 AI will add up to $15.7 trillion to the world economy, and as a result, global GDP will rise by 14%. According to Gartner's prediction, business value added by blockchain technology will increase to $3.1 trillion by the same year. By definition, a blockchain is a distributed, decentralized, immutable ledger used to store encrypted data.

Elik co-founded BigPanda with a vision for enabling companies to pursue fully autonomous IT operations. For those of us in the tech space, you've likely heard of AIOps, or artificial intelligence for IT operations, which "involves using AI and ML technologies along with big data, data integration, and automation technologies to help make IT operations smarter and more predictive." The research firm Gartner recently defined two different high-level categories of AIOps: domain-centric and domain-agnostic. Domain-centric tools focus on homogenous, first-party data sets and introduce AI capabilities to solve specific use cases, such as network and application diagnostics. Domain-agnostic AIOps platforms combine diverse data sets and data types and synthesize them into insight or action.

In May, as part of an otherwise unremarkable corporate strategy meeting, Sony CEO Kenichiro Yoshida made an interesting announcement. The company's artificial intelligence research division, Sony AI, would be collaborating with PlayStation developers to create intelligent computer-controlled characters. "By leveraging reinforcement learning," he wrote, "we are developing game AI agents that can be a player's in-game opponent or collaboration partner." Reinforcement learning is an area of machine learning in which an AI effectively teaches itself how to act through trial and error. In short, these characters will mimic human players.

Health care is at a junction, a point where artificial intelligence tools are being introduced to all areas of the space. This introduction comes with great expectations: AI has the potential to greatly improve existing technologies, sharpen personalized medicines, and, with an influx of big data, benefit historically underserved populations. But in order to do those things, the health care community must ensure that AI tools are trustworthy, and that they don't end up perpetuating biases that exist in the current system. Researchers at the MIT Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic), an initiative to support AI research in health care, call for creating a robust infrastructure that can aid scientists and clinicians in pursuing this mission. The Jameel Clinic recently hosted the AI for Health Care Equity Conference to assess current state-of-the-art work in this space, including new machine learning techniques that support fairness, personalization, and inclusiveness; identify key areas of impact in health care delivery; and discuss regulatory and policy implications.