AI Technology & Industry Review
By combining State-Space Models (SSMs) for efficient long-range dependency modeling with dense local attention for coherence, and using training strategies like diffusion forcing and frame local attention, researchers from Adobe Research successfully overcome the long-standing challenge of long-term memory in video generation.
A newly released 14-page technical paper from the team behind DeepSeek-V3, with DeepSeek CEO Wenfeng Liang as a co-author, sheds light on the “Scaling Challenges and Reflections on Hardware for AI Architectures.”
DeepSeek AI releases DeepSeek-Prover-V2, an open-source LLM for Lean 4 theorem proving. It uses recursive proof search with DeepSeek-V3 for training data and reinforcement learning, achieving top results on MiniF2F.
Kwai AI’s SRPO framework slashes LLM RL post-training steps by 90% while matching DeepSeek-R1 performance in math and code. This two-stage RL approach with history resampling overcomes GRPO limitations.
DeepSeek AI, a prominent player in the large language model arena, has recently published a research paper detailing a new technique aimed at enhancing the scalability of general reward models (GRMs) during the inference phase.
Meta AI’s recent research introduces the BLT architecture, eliminating tokenizers for improved multimodal processing, and the Large Concept Model (LCM), which operates on semantic “concepts” instead of tokens for more human-like reasoning and better cross-lingual generalization. These innovations challenge the traditional “next-token prediction” paradigm in LLMs.
Nvidia will launch Jetson Thor for humanoid robots in H1 2025, entering a growing market where Google is also active. The robotics sector is projected for substantial growth. Nvidia offers integrated hardware and software solutions. Simultaneously, China’s rapidly developing domestic humanoid robot market presents emerging competition.
In a recent technical report, OpenAI introduces Sora, a groundbreaking text-to-video model. Sora stands out for its ability to generate videos and images spanning a wide range of durations, aspect ratios, and resolutions, producing up to a minute of high-definition video content.
A Google DeepMind research team introduces a groundbreaking family of multimodal models Gemini, which showcase exceptional proficiency across image, audio, video, and text comprehension, pushing the boundaries of large-scale language modeling, image interpretation, audio processing, and video understanding.
In a new paper Llama 2: Open Foundation and Fine-Tuned Chat Model, a Meta AI research team presents and releases Llama 2 and Llama 2-Chat, the former one is a family of pretrained and fine-tuned LLMs and the later one is a fine-tuned version of Llama 2 that is optimized for dialogue, paving the way to develop more responsible LLMs.
In the new paper Language Is Not All You Need: Aligning Perception with Language Models, a Microsoft research team presents KOSMOS-1, a multimodal large language model (MLLM) that can perceive general modalities, learn in context, and follow instructions.
Turing Award winner and deep learning pioneer Geoffrey Hinton, one of the original proponents of backpropagation, has argued in recent years that backpropagation does not explain how the brain works. In his NeurIPS 2022 keynote speech, Hinton proposes a new approach to neural network learning: the Forward-Forward algorithm.
In the new paper Locating and Editing Factual Associations in GPT, a research team from MIT CSAIL, Northeastern University and Technion IIT examines how information flows during knowledge recall in large autoregressive transformers and introduces Rank-One Model Editing (ROME), a simple, zero-shot principled model editor capable of locating and editing factual associations in such models.
While today’s deep neural networks (DNNs) are driving AI’s deep-learning revolution, determining a DNN’s appropriate complexity remains challenging. If aContinue Reading
In the new paper Image as a Foreign Language: BEiT Pretraining for All Vision and Vision-Language Tasks, a Microsoft research team presents BEiT-3, a general-purpose state-of-the-art multimodal foundation model for both vision and vision-language tasks that advances the big convergence of backbone architectures, pretraining tasks, and model scaling.
In the new paper YOLOv7: Trainable Bag-Of-Freebies Sets New State-Of-The-Art for Real-Time Object Detectors, an Academia Sinica research team releases YOLOv7. This latest YOLO version introduces novel “extend” and “compound scaling” methods that effectively utilize parameters and computation; and surpasses all known real-time object detectors in speed and accuracy.
In the new paper An Improved One millisecond Mobile Backbone, an Apple research team presents MobileOne, a novel mobile backbone that cuts inference time to under one millisecond on an iPhone12 and reaches 75.9 percent top-1 accuracy on ImageNet.
The BAAI Conference 2022 kicked off on May 31 in Beijing and ran through June 2. AI experts, industry leaders, young talents and international delegates joined the virtual gathering and live stream for three busy days of high-level keynotes, tech talks, parallel forums and networking.
In the new paper A Modern Self-Referential Weight Matrix That Learns to Modify Itself, a research team from The Swiss AI Lab, IDSIA, University of Lugano (USI) & SUPSI, and King Abdullah University of Science and Technology (KAUST) presents a scalable self-referential weight matrix (SRWM) that leverages outer products and the delta update rule to update and improve itself.
A DeepMind research team argues that the mathematical description of symmetries in group theory is an important foundation that determines the structure of the universe, constrains the nature of natural tasks, and consequently shapes both biological and artificial intelligence. The study proposes symmetry transformations as a fundamental principle for defining what makes good representations.
A DeepMind research team proposes ReLICv2, which demonstrates for the first time that representations learned without labels can consistently outperform a strong, supervised baseline on ImageNet and even achieve comparable results to state-of-the-art self-supervised vision transformers (ViTs).
In the new paper A Neural Network Solves and Generates Mathematics Problems by Program Synthesis: Calculus, Differential Equations, Linear Algebra, and More, a research team from MIT, Columbia University, Harvard University and University of Waterloo proposes a neural network that can solve university-level mathematics problems via program synthesis.
In the new paper On the Integration of Self-Attention and Convolution, a research team from Tsinghua University, Huawei Technologies Ltd. and the Beijing Academy of Artificial Intelligence proposes ACmix, a mixed model that leverages the benefits of both self-attention and convolution for computer vision representation tasks while achieving minimum computational overhead compared to its pure convolution or self-attention counterparts.
A research team from Google Research, University of Cambridge and Alan Turing Institute proposes PolyViT, a single transformer model capable of processing multiple modalities and datasets. PolyViT is parameter-efficient and learns representations that generalize across multiple domains.
In a new paper, a Facebook AI team advances autoencoding methods to the computer vision field and shows that masked autoencoders (MAE) are scalable self-supervised learners.
A research team from the University of Southern California and Google proposes TOME, a “mention memory” approach to factual knowledge extraction for NLU tasks. A transformer model with attention over a semi-parametric representation of the entire Wikipedia text corpus, TOME can extract information without supervision and achieves strong performance on multiple open-domain question answering benchmarks.
A research team from DeepMind proposes a bootstrapped meta-learning algorithm that overcomes the meta-optimization problem and myopic meta objectives, and enables the meta-learner to teach itself.
A team from Tsinghua University and Microsoft Research Asia proposes Fastformer, an efficient Transformer variant based on additive attention that achieves effective context modelling with linear complexity.
A Google Research team explores the design space of Transformer models in an effort to enable deep learning architectures to solve compositional tasks. The proposed approach provides models with inductive biases via design decisions that significantly impact compositional generalization, and achieves state-of-the-art results on the COGS and PCFG composition benchmarks.
A research team from MIT and Google Quantum AI presents a quantum algorithm for training classical neural networks in logarithmic time and provides numerical evidence of its efficiency on the standard MNIST image dataset.
In a new paper published in the prestigious scientific journal Nature, DeepMind presents AlphaFold2, a redesigned neural-network system based on last year’s AlphaFold that can predict protein structures with atomic-level accuracy.
A research team from Facebook AI and UC Berkeley finds a solution for vision transformers’ optimization instability problem by simply using a standard, lightweight convolutional stem for ViT models. The approach dramatically increases optimizer stability and improves peak performance without sacrificing computation efficiency.
A research team from MIT, Allen Institute for AI and Microsoft Research open-sources Python Programming Puzzles (P3), a novel programming challenge suite that captures the essence of puzzles and can be used to teach and evaluate an AI’s programming proficiency.
A research team from McGill University, Université de Montréal, DeepMind and Mila presents an end-to-end, model-based deep reinforcement learning (RL) agent that dynamically attends to relevant parts of its environments to facilitate out-of-distribution (OOD) and systematic generalization.
A research team from Google Research combines the benefits of implicit differentiation and autodiff and proposes a unified, efficient and modular approach for implicit differentiation of optimization problems.
A team from Cornell University and NTT Research proposes Physical Neural Networks (PNNs), a universal framework that leverages a backpropagation algorithm to train arbitrary, real physical systems to execute deep neural networks.
A research team from ETH Zürich presents an overview of priors for (deep) Gaussian processes, variational autoencoders and Bayesian neural networks. The researchers propose that well-chosen priors can achieve theoretical and empirical properties such as uncertainty estimation, model selection and optimal decision support; and provide guidance on how to choose them.
A research team from Google shows that replacing transformers’ self-attention sublayers with Fourier Transform achieves 92 percent of BERT accuracy on the GLUE benchmark with training times seven times faster on GPUs and twice as fast on TPUs.
Twitter Chief Scientist Michael Bronstein, Joan Bruna from New York University, Taco Cohen from Qualcomm AI and Petar Veličković from DeepMind publish a paper that aims to geometrically unify the typical architectures of CNNs, GNNs, LSTMs, Transformers, etc. from the perspective of symmetry and invariance to build an “Erlangen Programme” for deep neural networks.
IBM and ETH Zurich researchers make progress in reconciling neurophysiological insights with machine intelligence, proposing a novel biologically inspired optimizer for artificial (ANNs) and spiking neural networks (SNNs) that incorporates synaptic integration principles from biology. GRAPES (Group Responsibility for Adjusting the Propagation of Error Signals) leads to improvements in the training time convergence, accuracy and scalability of ANNs and SNNs.