Lingfeng Zhang

Virtual Personas for Language Models via an Anthology of Backstories S Moon, M Abdulhai, M Kang, J Suh… [UC Berkeley] (2024) https://lnkd.in/dZiVjPq5 - This paper introduces "Anthology", a method for conditioning large language models (LLMs) to diverse virtual personas by generating naturalistic backstories as prefix context. - Backstories implicitly and explicitly embody details about an author, including demographics and personality traits. Using them as prefixes strongly conditions LLM responses. - The authors generate massive, diverse sets of backstories using LLMs given an open-ended prompt about describing one's life. - They present methods to estimate backstory authors' demographics via sampling, and match backstories to target populations. - In experiments on 3 surveys from Pew Research's American Trends Panel, Anthology gives 18% lower Wasserstein distance and 27% higher consistency than baselines in approximating human responses. - Anthology also better represents diverse personas, especially underrepresented groups. - The authors provide an open-source anthology of 10,000 backstories to enable behavioral studies with virtual subjects. abs: Large language models (LLMs) are trained from vast repositories of text authored by millions of distinct authors, reflecting an enormous diversity of human traits. While these models bear the potential to be used as approximations of human subjects in behavioral studies, prior efforts have been limited in steering model responses to match individual human users. In this work, we introduce "Anthology", a method for conditioning LLMs to particular virtual personas by harnessing open-ended life narratives, which we refer to as "backstories." We show that our methodology enhances the consistency and reliability of experimental outcomes while ensuring better representation of diverse sub-populations. Across three nationally representative human surveys conducted as part of Pew Research Center's American Trends Panel (ATP), we demonstrate that Anthology achieves up to 18% improvement in matching the response distributions of human respondents and 27% improvement in consistency metrics. Our code and generated backstories are available at https://lnkd.in/dZiVjPq5. #LLM #AI #computation #language

22

2 Comments

From Google DeepMind State Soup: In-Context Skill Learning, Retrieval and Mixing M Pióro, M Wołczyk, R Pascanu, J v Oswald, J Sacramento [IDEAS NCBR & Google Deepmind] (2024) https://lnkd.in/dkG5n8ip - Modern RNNs with gated-linear recurrences enable efficient training and good optimization properties. They scale linearly in sequence length for inference. - Inspired by model soups, which interpolate parameters of different models, the authors propose state soups - interpolating internal states of RNNs. - They show state vectors of different in-context learned skills cluster together under projection, enabling task retrieval with a query state. - Mixing retrieved states boosts few-shot in-context learning performance. Preliminary evidence shows mixing states from different tasks can produce interesting blended behaviors. - For ordered data, they propose A-decay mixing which leverages the linearity of gated-linear RNNs to combine independently processed chunks. It approximates full sequence processing. abs: A new breed of gated-linear recurrent neural networks has reached state-of-the-art performance on a range of sequence modeling problems. Such models naturally handle long sequences efficiently, as the cost of processing a new input is independent of sequence length. Here, we explore another advantage of these stateful sequence models, inspired by the success of model merging through parameter interpolation. Building on parallels between fine-tuning and in-context learning, we investigate whether we can treat internal states as task vectors that can be stored, retrieved, and then linearly combined, exploiting the linearity of recurrence. We study this form of fast model merging on Mamba-2.8b, a pretrained recurrent model, and present preliminary evidence that simple linear state interpolation methods suffice to improve next-token perplexity as well as downstream in-context learning task performance. #AI #machinelearning

16

From MIT A Multimodal Automated Interpretability Agent https://lnkd.in/dp5djduP This paper describes MAIA, a Multimodal Automated Interpretability Agent. MAIA is a system that uses neural models to automate neural model understanding tasks like feature interpretation and failure mode discovery. It equips a pre-trained vision-language model with a set of tools that support iterative experimentation on subcomponents of other models to explain their behavior. These include tools commonly used by human interpretability researchers: for synthesizing and editing inputs, computing maximally activating exemplars from real-world datasets, and summarizing and describing experimental results. Interpretability experiments proposed by MAIA compose these tools to describe and explain system behavior. We evaluate applications of MAIA to computer vision models. We first characterize MAIA's ability to describe (neuron-level) features in learned representations of images. Across several trained models and a novel dataset of synthetic vision neurons with paired ground-truth descriptions, MAIA produces descriptions comparable to those generated by expert human experimenters. We then show that MAIA can aid in two additional interpretability tasks: reducing sensitivity to spurious features, and automatically identifying inputs likely to be mis-classified. #AI #neuron #LLM #computervision

22

Google presents Probing the 3D Awareness of Visual Foundation Models Visual #foundation models can learn representations that #ENCODE the depth and orientation of the #visible surface but struggle with #Multiview consistency possibly because they are learning view-dependent representations repo: https://lnkd.in/eAkkA8uq abs: https://lnkd.in/e4JJVns9 #AI

13

1 Comment

Learn how an image-text multi-modality model can perform image classification, image retrieval, and image captioning #llms #llmops #image #opencv #chatgpt4 #gemini #aiml https://lnkd.in/gJFh9Ruf

1

Kolmogorov-Arnold Networks (KAN) using Chebyshev polynomials instead of B-splines. 🔥🔥🔥🔥🔥 https://lnkd.in/dRGX7gp9

12

Diffusion Model-Augmented Behavioral Cloning Shang-Fu Chen, Hsiang-Chun Wang, Ming-Hao Hsu, Chun-Mao Lai, Shao-Hua Sun National Taiwan University rev. June 2024 https://lnkd.in/eJgXnA-H This work presents an imitation learning method. First, we train a diffusion model to model expert state-action pairs sample from the expert dataset. Specifically, the diffusion mode learns to reverse the diffusion process, i.e., denoise expert state-action pairs. We propose to use how well the diffusion model can denoise a state-action pair as a density estimate to evaluate how likely this sample is drawn from the expert demonstrations. Then, we design a learning objective encouraging the learner policy to maximize this density estimate. We combine this with the BC objective. While the BC objective directly guides the policy to replicate expert actions from sampled state-action pairs, the diffusion model objective forces the policy to match expert demonstration distribution, leading to improved generalizability. Our proposed method outperforms baselines or achieves competitive performance in various continuous control domains, including navigation (Maze & AntMaze), robot arm manipulation (FetchPush & HandRotate), and locomotion (Cheetah & Walker). abs: Imitation learning addresses the challenge of learning by observing an expert's demonstrations without access to reward signals from environments. Most existing imitation learning methods that do not require interacting with environments either model the expert distribution as the conditional probability p(a|s) (e.g., behavioral cloning, BC) or the joint probability p(s, a). Despite the simplicity of modeling the conditional probability with BC, it usually struggles with generalization. While modeling the joint probability can improve generalization performance, the inference procedure is often time-consuming, and the model can suffer from manifold overfitting. This work proposes an imitation learning framework that benefits from modeling both the conditional and joint probability of the expert distribution. Our proposed Diffusion Model-Augmented Behavioral Cloning (DBC) employs a diffusion model trained to model expert behaviors and learns a policy to optimize both the BC loss (conditional) and our proposed diffusion model loss (joint). DBC outperforms baselines in various continuous control tasks in navigation, robot arm manipulation, dexterous manipulation, and locomotion. We design additional experiments to verify the limitations of modeling either the conditional probability or the joint probability of the expert distribution, as well as compare different generative models. Ablation studies justify the effectiveness of our design choices. #AI #machinelearning #robotics

22

Can LLMs Teach Themselves? Exploring Self-Improvement in Language Model Training If so, which generation of LLMs can do so? The current generation IME cannot. https://lnkd.in/gK-cevmy

4

Using generative AI, NVIDIA operations built an AI Planner agent, developed on NVIDIA Inference Microservices (NIM). The agent leverages LLM, NeMo Retriever and CuOpt NIM to reduce re-planning time from hours to just seconds. https://nvda.ws/4bIfS2T

2

What type of inference is planning? Miguel Lázaro-Gredilla, Li Yang Ku, Kevin P. Murphy, Dileep George Google Deepmind June 2024 https://lnkd.in/e78CdGm6 - This paper provides a variational framework to analyze different types of inference for solving Markov decision processes (MDPs), including marginal, MAP, MMAP and a novel "planning" inference. - It shows planning corresponds to a different weighting of entropy terms in the variational lower bound compared to other inference types. This allows transferring insights between variational inference and planning. - The framework derives a novel linear programming formulation and loopy belief propagation algorithm for approximate planning in factored MDPs with large state spaces. - It also enables closed-form "planning" inference approximations that outperform sampling-based determinization. - Key insight: the stochasticity of the dynamics determines which type of inference is most suitable for planning. Planning inference is best for stochastic environments. - For deterministic environments, MAP and MMAP give exact solutions, but have limitations under stochastic dynamics. Marginal lacks integration over trajectories. - The variational view disentangles the type of inference from additional approximations needed in practice. abs: Multiple types of inference are available for probabilistic graphical models, e.g., marginal, maximum-a-posteriori, and even marginal maximum-a-posteriori. Which one do researchers mean when they talk about "planning as inference"? There is no consistency in the literature, different types are used, and their ability to do planning is further entangled with specific approximations or additional constraints. In this work we use the variational framework to show that all commonly used types of inference correspond to different weightings of the entropy terms in the variational problem, and that planning corresponds _exactly_ to a _different_ set of weights. This means that all the tricks of variational inference are readily applicable to planning. We develop an analogue of loopy belief propagation that allows us to perform approximate planning in factored state Markov decisions processes without incurring intractability due to the exponentially large state space. The variational perspective shows that the previous types of inference for planning are only adequate in environments with low stochasticity, and allows us to characterize each type by its own merits, disentangling the type of inference from the additional approximations that its practical use requires. We validate these results empirically on synthetic MDPs and tasks posed in the International Planning Competition. #machinelearning #AI

5

New ODE (ordinary differential equation) solvers in #MATLAB from the SUNDIALS suite, provides the ability to perform parameter sensitivity analysis. Learn more here 🔗

4

Responsible and accountable AI models are essential as generative AI continues to revolutionize industries. The rise of this technology, however, presents challenges in understanding how these models make decisions. Mechanistic interpretability addresses this by going beyond traditional methods that track statistical relationships. A recent research work by a team from #Anthropic has involved visualizing which neurons activate in response to specific prompts, advancing monosemanticity—ensuring each neuron’s activation correlates to a single, clear function. This step is crucial for fostering more transparent and safer AI systems in various applications. This is a great progress toward interpreting the thinking mechanisms of large language models (LLMs). #GenerativeAI #AI #LLMs #GenerativeAI #Grainger #MachineLearning #DeepLearning #ExplainableAI #DataScience

45

2 Comments

Learn how to build an autonomous research assistant using LangGraph with a team of specialized AI agents #llms #chatgpt4 #gemini #aiml #llmops https://lnkd.in/d6TUbZNj

2

From UC Berkeley, Stanford, Carnegie Mellon, & Google Deepmind Octo An Open-Source Generalist Robot Policy https://lnkd.in/ePvtAAeg Large policies pretrained on diverse robot datasets have the potential to transform robotic learning: instead of training new policies from scratch, such generalist robot policies may be finetuned with only a little in-domain data, yet generalize broadly. However, to be widely applicable across a range of robotic learning scenarios, environments, and tasks, such policies need to handle diverse sensors and action spaces, accommodate a variety of commonly used robotic platforms, and finetune readily and efficiently to new domains. In this work, we aim to lay the groundwork for developing open-source, widely applicable, generalist policies for robotic manipulation. As a first step, we introduce Octo, a large transformer-based policy trained on 800k trajectories from the Open X-Embodiment dataset, the largest robot manipulation dataset to date. It can be instructed via language commands or goal images and can be effectively finetuned to robot setups with new sensory inputs and action spaces within a few hours on standard consumer GPUs. In experiments across 9 robotic platforms, we demonstrate that Octo serves as a versatile policy initialization that can be effectively finetuned to new observation and action spaces. We also perform detailed ablations of design decisions for the Octo model, from architecture to training data, to guide future research on building generalist robot models. #AI #robots #opensource #HuggingFace

15

In addition to Llama 3, today Meta is also publishing a new paper: Imagine Flash: Accelerating Emu Diffusion Models with Backward Distillation https://lnkd.in/e9n478jZ This work from GenAI researchers is enabling new image generation features in Meta AI on WhatsApp & web. #AI #computervision

13

📖 "Efficient Multimodal Large Language Models: A Survey" - a comprehensive review of recent progress in efficient MLLMs. Key advances include:   💡 Architecture: optimizations in vision encoder, vision-language projector, language model, vision token compression, and using efficient structures like MoE and Mamba. 💡 Training: pre-training, instruction-tuning, and fine tuning 💡 Data: pre-training data, instruction data and benchmarks   Efficient MLLM efforts aim to make the powerful models more accessible - reducing cost and latency, and enabling deployments on device for privacy protection.   Read the full paper: https://lnkd.in/gFEw_XQ8   #generativeai #MLLM #largelanguagemodels

20

- MobileCLIP-S0 achieves comparable zero-shot performance to OpenAI's ViT-B/16, with a 4.8x speed improvement and 2.8x smaller size. - MobileCLIP-S2 outperforms SigLIP's ViT-B/16 in average zero-shot performance, boasting 2.3x faster speed, 2.1x smaller size, and trained with 3x fewer seen samples. - MobileCLIP-B(LT) achieves a remarkable zero-shot ImageNet performance of 77.2%, surpassing recent benchmarks like DFN and SigLIP with similar architectures, as well as OpenAI's ViT-L/14@336. Code: https://lnkd.in/gQfUpcQT Paper: https://lnkd.in/gHGux7MS

9

Recently, I was re-introduced to the concept of catastrophic forgetting while reading about NVIDIA's ViLa model, which BTW offers pretty cool capabilities like reasoning across multiple images and learning in context and is capable of understanding videos, offering a more comprehensive understanding of visual content and its interplay with language. Overcoming catastrophic forgetting is crucial because it allows AI models to retain valuable knowledge while continuously learning and adapting, enhancing their overall performance and flexibility. Read article here: https://lnkd.in/e8gimTk4

43

1 Comment