Month: February 2024

The alignment of Large Language Models (LLMs) with human preferences has become a crucial area of research. As these models gain complexity and capability, ensuring their actions and outputs align with human values and intentions is paramount. The conventional route to this alignment has involved sophisticated reinforcement learning techniques, with Proximal Policy Optimization (PPO) leading the charge. While effective, this method comes with its own challenges, including high computational demands and the need for delicate hyperparameter adjustments. These challenges raise the question: Is there a more efficient yet equally effective way to achieve the same goal?

A research team from Cohere For AI and Cohere performed an exploration to address this question, turning their focus to a less computationally intensive approach that does not compromise performance. They revisited the foundations of reinforcement learning in the context of human feedback, specifically evaluating the efficiency of REINFORCE-style optimization variants against the traditional PPO and recent “RL-free” methods like DPO and RAFT. Their investigation revealed that simpler methods could match or even surpass the performance of their more complex counterparts in aligning LLMs with human preferences.

The methodology employed dissected the RL component of RLHF, stripping away the complexities associated with PPO to highlight the efficacy of simpler, more straightforward approaches. Through their analysis, they identified that the core principles driving the development of PPO, principally its focus on minimizing variance and maximizing stability in updates, may not be as critical in the context of RLHF as previously thought.

Their empirical analysis, utilizing datasets from Google Vizier, demonstrated a notable performance improvement when employing REINFORCE and its multi-sample extension, REINFORCE Leave-One-Out (RLOO), over traditional methods. Their findings showed an over 20% increase in performance, marking a significant leap forward in the efficiency and effectiveness of LLM alignment with human preferences.

This research challenges the prevailing norms regarding the necessity of complex reinforcement learning methods for LLM alignment and opens the door to more accessible and potentially more effective alternatives. The key insights from this study underscore the potential of simpler reinforcement learning variants in achieving high-quality LLM alignment at a lower computational cost.

In conclusion, Cohere’s research suggests some key insights, including:

Simplifying the RL component of RLHF can lead to improved alignment of LLMs with human preferences without sacrificing computational efficiency.

Traditional, complex methods such as PPO might not be indispensable in RLHF settings, paving the way for simpler, more efficient alternatives.

REINFORCE and its multi-sample extension, RLOO, emerge as promising candidates, offering a blend of performance and computational efficiency that challenges the status quo.

This work marks a pivotal shift in the approach to LLM alignment, suggesting that simplicity, rather than complexity, might be the key to more effective and efficient alignment of artificial intelligence with human values and preferences.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Can Machine Learning Models Be Fine-Tuned More Efficiently? This AI Paper from Cohere for AI Reveals How REINFORCE Beats PPO in Reinforcement Learning from Human Feedback appeared first on MarkTechPost.

The challenges in developing instruction-following agents in grounded environments include sample efficiency and generalizability. These agents must learn effectively from a few demonstrations while performing successfully in new environments with novel instructions post-training. Techniques like reinforcement learning and imitation learning are commonly used but often demand numerous trials or costly expert demonstrations due to their reliance on trial and error or expert guidance.

In language-grounded instruction following, agents receive instructions and partial observations in the environment, taking actions accordingly. Reinforcement learning involves receiving rewards, while imitation learning mimics expert actions. Behavioral cloning collects offline expert data to train the policy, different from online imitation learning, aiding in long-horizon tasks in grounded environments. Recent studies demonstrate that large language models (LLMs), when pretrained, display sample-efficient learning via prompting and in-context learning across textual and grounded tasks, including robotic control. Nonetheless, existing methods for instruction following grounded scenarios depend on LLMs online during inference, posing impracticality and high costs.

Researchers from Microsoft Research and the University of Waterloo have proposed Language Feedback Models (LFMs) for policy improvement in instruction. LFMs leverage LLMs to provide feedback on agent behavior in grounded environments, aiding in identifying desirable actions. By distilling this feedback into a compact LFM, the technique enables sample-efficient and cost-effective policy improvement without continuous reliance on LLMs. LFMs generalize to new environments and offer interpretable feedback for human validation of imitation data.

The proposed method introduces LFMs to enhance policy learning in the following instruction. LFMs leverage LLMs to identify productive behavior from a base policy, facilitating batched imitation learning for policy improvement. By distilling world knowledge from LLMs into compact LFMs, the approach achieves sample-efficient and generalizable policy enhancement without needing continuous online interactions with expensive LLMs during deployment. Instead of using the LLM at each step, we modify the procedure to collect LLM feedback in batches over long horizons for a cost-effective language feedback model.

They have used GPT-4 LLM for action prediction and feedback for experimentation and fine-tuned the 770M FLANT5 to obtain policy and feedback models. Utilizing LLMs, LFMs identify productive behavior, enhancing policies without continual LLM interactions. LFMs outperform direct LLM usage, generalize to new environments, and provide interpretable feedback. They offer a cost-effective means for policy improvement and foster user trust. Overall, LFMs significantly improve policy performance, demonstrating their efficacy in grounded instruction following.

In conclusion, Researchers from Microsoft Research and the University of Waterloo have proposed Language Feedback Models. LFM excels in identifying desirable behavior for imitation learning across various benchmarks. They surpass baseline methods and LLM-based expert imitation learning without continual LLM usage. LFMs generalize well, offering significant policy adaptation gains in new environments. Additionally, they provide detailed, human-interpretable feedback, fostering trust in imitation data. Future research could explore leveraging detailed LFMs for RL reward modeling and creating trustworthy policies with human verification.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Can Machine Learning Teach Robots to Understand Us Better? This Microsoft Research Introduces Language Feedback Models for Advanced Imitation Learning appeared first on MarkTechPost.

In machine learning, finding the perfect settings for a model to work at its best can be like looking for a needle in a haystack. This process, known as hyperparameter optimization, involves tweaking the settings that govern how the model learns. It’s crucial because the right combination can significantly improve a model’s accuracy and efficiency. However, this process can be time-consuming and complex, requiring extensive trial and error.

Traditionally, researchers and developers have resorted to manual tuning or using grid search and random search methods to find the best hyperparameters. These methods do work to some extent but could be more efficient. Manual tuning is labor-intensive and subjective, while grid and random searches can be like shooting in the dark – they might hit the target but often waste time and resources.

Meet Optuna: a software framework designed to automate and accelerate the hyperparameter optimization process. This framework employs a unique approach, allowing users to define their search space dynamically using Python code. It supports exploring various machine learning models and their configurations to identify the most effective settings.

This framework stands out due to its several vital features. It’s lightweight and flexible, meaning it can be used across different platforms and for various tasks with minimal setup. Its Pythonic search spaces allow for familiar syntax, making the definition of complex search spaces straightforward. The framework incorporates efficient optimization algorithms that can sample hyperparameters and prune less promising trials, enhancing the speed of the optimization process. Additionally, it supports easy parallelization, enabling the scaling of studies to numerous workers without significant changes to the code. Moreover, its quick visualization capabilities allow users to inspect optimization histories quickly, aiding in the analysis and decision-making process.

In conclusion, this software framework provides a powerful tool for those involved in machine learning projects, simplifying the once daunting task of hyperparameter optimization. Automating the search for the optimal model settings saves valuable time and resources and opens up new possibilities for improving model performance. Its design, which emphasizes efficiency, flexibility, and user-friendliness, makes it an option for both beginners and experienced practitioners in machine learning. As the demand for more sophisticated and accurate models grows, such tools will undoubtedly become indispensable in using the full potential of machine learning technologies.
The post Meet Optuna: An Automatic Hyperparameter Optimization Software Framework Designed for Machine Learning appeared first on MarkTechPost.

Deep learning has revolutionized view synthesis in computer vision, offering diverse approaches like NeRF and end-to-end style architectures. Traditionally, 3D modeling methods like voxels, point clouds, or meshes were employed. NeRF-based techniques implicitly represent 3D scenes using MLPs. Recent advancements focus on image-to-image approaches, generating novel views from collections of scene images. These methods often require costly re-training per scene, precise pose information, or help with variable input views at test time. Despite their strengths, each approach has limitations, underscoring the ongoing challenges in this field.

Researchers from the Department of Computer Science and the  Neuroscience and Biomedical Engineering at Aalto University, Finland, System 2 AI, and Finnish Center for Artificial Intelligence FCAI. have developed. ViewFusion is an advanced generative method for view synthesis. It employs diffusion denoising and pixel-weighting to combine informative input views, addressing previous limitations. ViewFusion is trainable across diverse scenes, adapts to varying input views, and generates high-quality results even in challenging conditions. Though it doesn’t create a 3D scene embedding and has slower inference, it outperforms existing methods on the NMR dataset.

View synthesis has explored approaches, from NeRFs to end-to-end architectures and diffusion probabilistic models. NeRFs optimize a continuous volumetric scene function but struggle with generalization and require significant retraining for different objects. End-to-end methods like Equivariant Neural Renderer and Scene Representation Transformers offer promising results but lack variability in output and often require explicit pose information. Diffusion probabilistic models leverage stochastic processes for high-quality outputs, but pre-trained backbone reliance and limited flexibility pose challenges. Despite their strengths, existing methods have drawbacks like inflexibility and dependence on specific data structures.

ViewFusion is an end-to-end generative approach to view synthesis that applies a diffusion denoising step to input views and combines noise gradients with a pixel-weighting mask. The model employs a composable diffusion probabilistic framework to generate views from an unordered collection of input views and a target viewing direction. The approach is evaluated using commonly used metrics such as PSNR, SSIM, and LPIPS and compared to state-of-the-art methods for novel view synthesis. The proposed approach resolves the limitations of previous methods by being trainable and generalizing across multiple scenes and object classes, adaptively taking in a variable number of pose-free views, and generating plausible views even in severely undetermined conditions. 

ViewFusion’s approach to view synthesis achieves top-tier performance in key metrics like PSNR, SSIM, and LPIPS. Evaluated on the diverse NMR dataset, it consistently matches or surpasses current state-of-the-art methods. ViewFusion excels in handling various scenarios, even in challenging, underdetermined conditions. Its adaptability shines through its capability to seamlessly incorporate varying numbers of pose-free views during training and inference stages, consistently delivering high-quality results regardless of input view count. Leveraging its generative nature, ViewFusion produces realistic views comparable to or surpassing existing state-of-the-art techniques.

In conclusion, ViewFusion is a groundbreaking solution for view synthesis, boasting state-of-the-art performance across metrics like PSNR, SSIM, and LPIPS. Its adaptability and flexibility surpass previous methods by seamlessly accommodating various pose-free views and generating high-quality outputs, even in challenging, underdetermined scenarios. By introducing a weighting scheme and leveraging composable diffusion models, ViewFusion sets a new standard in the field. Beyond its immediate application, the generative nature of ViewFusion holds promise for addressing broader problems, marking it as a significant contribution with potential applications beyond novel view synthesis.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter. 

Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Researchers from Aalto University ViewFusion: Revolutionizing View Synthesis with Adaptive Diffusion Denoising and Pixel-Weighting Techniques appeared first on MarkTechPost.

Underwater image processing combined with machine learning offers significant potential for enhancing the capabilities of underwater robots across various marine exploration tasks. Image segmentation, a key aspect of machine vision, is crucial for identifying and isolating objects of interest within underwater images. Traditional segmentation methods, such as threshold-based and morphology-based algorithms, have been employed but need help accurately delineating objects in the complex underwater environment where image degradation is common.

Researchers increasingly use deep learning techniques for underwater image segmentation to address these challenges. Deep learning methods, including semantic and instance segmentation, provide more precise analysis by enabling pixel-level and object-level segmentation. Recent advancements, such as FCN-DenseNet and Mask R-CNN, promise to improve segmentation accuracy and speed. However, further research is needed to overcome challenges like limited dataset availability and image quality degradation, ensuring robust performance in underwater exploration scenarios.

To deal with the challenges posed by limited underwater image datasets and image quality degradation, a research team from China recently published a new paper proposing innovative solutions.

The proposed method is based on the following steps: Firstly, they expanded the size of the underwater image dataset by employing techniques such as image rotation, flipping, and a Generative Adversarial Network (GAN) to generate additional images. Secondly, they applied an underwater image enhancement algorithm to preprocess the dataset, addressing issues related to image quality degradation. Thirdly, the researchers reconstructed the deep learning network by removing the last layer of the feature map with the largest receptive field in the Feature Pyramid Network (FPN) and replacing the original backbone network with a lightweight feature extraction network.

Using image transformations and a ConSinGan network, they enhanced the initial images from the Underwater Robot Picking Contest (URPC2020) to create an underwater image dataset, for instance, segmentation. This network uses three convolutional layers to expand the dataset by producing higher-resolution images after several training cycles. They also labeled target positions and categories using a Mask R-CNN network for image annotation, building a fully labeled dataset in Visual Object Classes (VOC) format. Creating new datasets increases their diversity and unpredictability, which is important for developing strong segmentation models that can adapt to various undersea conditions.

The experimental study assessed the effectiveness of the proposed approach in enhancing underwater image quality and refining instance segmentation accuracy. Quantitative metrics, including information entropy, root mean square contrast, average gradient, and underwater color image quality evaluation, were utilized to evaluate image enhancement algorithms, where the combination algorithm, notably WAC, exhibited superior performance. Validation experiments confirmed the efficacy of data augmentation techniques in refining segmentation accuracy and underscored the effectiveness of image preprocessing algorithms, with WAC surpassing alternative methods. Modifications to the Mask R-CNN network, particularly the Feature Pyramid Network (FPN), improved segmentation accuracy and processing speed. Integrating image preprocessing with network enhancements further bolstered recognition and segmentation accuracy, validating the approach’s efficacy in underwater image analysis and segmentation tasks.

In summary, integrating underwater image processing with machine learning holds promise for enhancing underwater robot capabilities in marine exploration. Deep learning techniques, including semantic and instance segmentation, offer precise analysis despite the challenges of the underwater environment. Recent advancements like FCN-DenseNet and Mask R-CNN show potential for improving segmentation accuracy. A recent study proposed a comprehensive approach involving dataset expansion, image enhancement algorithms, and network modifications, demonstrating effectiveness in enhancing image quality and refining segmentation accuracy. This approach has significant implications for underwater image analysis and segmentation tasks.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Enhancing Underwater Image Segmentation with Deep Learning: A Novel Approach to Dataset Expansion and Preprocessing Techniques appeared first on MarkTechPost.

The issue of bias in LLMs is a critical concern as these models, integral to advancements across sectors like healthcare, education, and finance, inherently reflect the biases in their training data, predominantly sourced from the internet. The potential for these biases to perpetuate and amplify societal inequalities necessitates a rigorous examination and mitigation strategy, highlighting a technical challenge and a moral imperative to ensure fairness and equity in AI applications.

Central to this discourse is the nuanced problem of geographic bias. This form of bias manifests through systematic errors in predictions about specific locations, leading to misrepresentations across cultural, socioeconomic, and political spectrums. Despite the extensive efforts to address biases concerning gender, race, and religion, the geographic dimension has remained relatively underexplored. This oversight underscores an urgent need for methodologies capable of detecting and correcting geographic disparities to foster AI technologies that are just and representative of global diversities.

A recent Stanford University study pioneers a novel approach to quantifying geographic bias in LLMs. The researchers propose a biased score that ingeniously combines mean absolute deviation and Spearman’s rank correlation coefficients, offering a robust metric to assess the presence and extent of geographic biases. This methodology stands out for its ability to systematically evaluate biases across various models, shedding light on the differential treatment of regions based on socioeconomic statuses and other geographically relevant criteria.

Delving deeper into the methodology reveals a sophisticated analysis framework. The researchers employed a series of carefully designed prompts aligned with ground truth data to evaluate LLMs’ ability to make zero-shot geospatial predictions. This innovative approach not only confirmed LLMs’ capability to process and predict geospatial data accurately but also exposed pronounced biases, particularly against regions with lower socioeconomic conditions. These biases manifest vividly in predictions related to subjective topics such as attractiveness and morality, where areas like Africa and parts of Asia were systematically undervalued.

The examination across different LLMs showcased significant monotonic correlations between the models’ predictions and socioeconomic indicators, such as infant survival rates. This correlation highlights a predisposition within these models to favor more affluent regions, thereby marginalizing lower socioeconomic areas. Such findings question the fairness and accuracy of LLMs and emphasize the broader societal implications of deploying AI technologies without adequate safeguards against biases.

This research underscores a pressing call to action for the AI community. The study stresses the importance of incorporating geographic equity into model development and evaluation by unveiling a previously overlooked aspect of AI fairness. Ensuring that AI technologies benefit humanity equitably necessitates a commitment to identifying and mitigating all forms of bias, including geographic disparities. Pursuing models that are not only intelligent but also fair and inclusive becomes paramount. The path forward involves technological advancements and collective ethical responsibility to harness AI in ways that respect and uplift all global communities, bridging divides rather than deepening them.

This comprehensive exploration into geographic bias in LLMs advances our understanding of AI fairness and sets a precedent for future research and development efforts. It serves as a reminder of the complexities inherent in building technologies that are truly beneficial for all, advocating for a more inclusive approach to AI that acknowledges and addresses the rich tapestry of human diversity. 

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Charting New Frontiers: Stanford University’s Pioneering Study on Geographic Bias in AI appeared first on MarkTechPost.

In an era where digital information proliferates, the capability of artificial intelligence (AI) to digest and understand extensive texts is more critical than ever. Despite their language prowess, traditional Large Language Models (LLMs) falter when faced with long documents, primarily due to inherent constraints on processing lengthy inputs. This limitation hampers their utility in scenarios where comprehension of vast texts is essential, underscoring a pressing need for innovative solutions that mirror human cognitive flexibility in dealing with extensive information.

The quest to transcend these boundaries led researchers from Google DeepMind and Google Research to pioneer ReadAgent. This groundbreaking system draws inspiration from human reading strategies to significantly enhance AI’s text comprehension capabilities. Unlike conventional approaches that either expand the context window LLMs can perceive or rely on external data retrieval systems to patch gaps in understanding, ReadAgent introduces a more nuanced, human-like method to navigate through lengthy documents efficiently.

At the heart of ReadAgent’s design is a clever emulation of human reading behaviors, specifically the practice of summarizing and recalling. This method involves a three-step process:

Segmenting the text into manageable parts

Condensing these segments into concise, gist-like summaries

Dynamically remembering detailed information from these summaries as necessary

This innovative approach allows the AI to grasp a document’s overarching narrative or argument, despite its length, by focusing on the core information and strategically revisiting details when needed.

The methodology behind ReadAgent is both simple and ingenious. Initially, the system segments a long text into episodes based on natural pause points, akin to chapters or sections in human reading. These segments are then compressed into ‘gist memories,’ which capture the essence of the text in a fraction of the original size. When specific information is required to address a query or task, ReadAgent revisits the relevant detailed segments, leveraging these gist memories as a roadmap to the original text. This process not only mimics human strategies for dealing with long texts but also significantly extends the effective context length that LLMs can handle, effectively overcoming one of the major limitations of current AI models.

The efficacy of ReadAgent is underscored by its performance across several long-document comprehension tasks. In experiments, ReadAgent demonstrated a substantial improvement over existing methods, extending the effective context length by up to 20 times. Specifically, on the NarrativeQA Gutenberg test set, ReadAgent improved the LLM rating by 12.97% and ROUGE-L by 31.98% over the best retrieval baseline, showcasing its superior ability to understand and process lengthy documents. This remarkable performance highlights not only the potential of AI to assimilate human-like reading and comprehension strategies and the practical applicability of such approaches in enhancing AI’s understanding of complex texts.

Developed by the innovative minds at Google DeepMind and Google Research, ReadAgent represents a significant leap forward in AI’s text comprehension capabilities. Embodying human reading strategies broadens AI’s applicability across domains requiring deep text understanding and paves the way for more sophisticated, cognitive-like AI systems. This advancement showcases the potential of human-inspired AI development and sets a new benchmark for AI’s role in navigating the ever-expanding digital information landscape.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Meet Google Deepmind’s ReadAgent: Bridging the Gap Between AI and Human-Like Reading of Vast Documents! appeared first on MarkTechPost.

Large language models (LLMs) have witnessed significant advancements, aiming to enhance their capabilities for interpreting and processing extensive textual data. LLMs like GPT-3 have revolutionized our interactions with AI, offering insights and analyses across various domains, from writing assistance to complex data interpretation. However, a key limitation has been their context window size, the amount of text they can consider in a single instance. LLMs could process up to a few thousand tokens, constraining their ability to understand and generate responses for longer documents.

Researchers from Microsoft Research have developed LongRoPE, a novel approach that significantly extends the context window of pre-trained LLMs to an impressive 2 million tokens. This breakthrough was achieved through three innovative strategies: identifying and leveraging non-uniformities in positional interpolation, introducing a progressive extension strategy, and readjusting LongRoPE to recover performance in shorter context windows. These innovations allow LLMs to perform well even when processing longer texts than initially designed.

LongRoPE utilizes an evolutionary search algorithm to optimize positional interpolation, enabling it to extend the context window of LLMs by up to 8 times without fine-tuning for extra-long texts. This is particularly beneficial because it overcomes the challenges of training on long texts, which are scarce and computationally expensive to process. The method has been extensively tested across various LLMs and tasks, demonstrating its effectiveness in maintaining low perplexity and high accuracy even in extended contexts.

The performance of LongRoPE retains the original model’s accuracy within the conventional short context window and significantly reduces perplexity in extended contexts up to 2 million tokens. This capability opens new avenues for LLM applications, enabling them to process and analyze long documents or books in their entirety without losing coherence or accuracy. For instance, LongRoPE’s application in LLaMA2 and Mistral models has shown superior performance in standard benchmarks and specific tasks like passkey retrieval from extensive texts, highlighting its potential to revolutionize leveraging LLMs for complex text analysis and generation tasks.

In conclusion, LongRoPE represents a significant leap forward in the field of LLMs, addressing a critical limitation in context window size. Enabling LLMs to process and understand texts of up to 2 million tokens paves the way for more sophisticated and nuanced AI applications. This innovation not only enhances the capabilities of existing models but also sets a new benchmark for future developments in large language models.

Key highlights of the conducted research in the following points:

LongRoPE’s innovative approach extends LLM context windows to 2 million tokens, a significant advancement in AI.

The evolutionary search algorithm optimizes positional interpolation, overcoming the traditional limitations of LLMs.

Extensive testing demonstrates LongRoPE’s ability to maintain accuracy and reduce perplexity in extended contexts.

This breakthrough opens new possibilities for complex text analysis and generation, enhancing LLM applications.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Breaking Barriers in Language Understanding: How Microsoft AI’s LongRoPE Extends Large Language Models to a 2048k Token Context Window appeared first on MarkTechPost.

The integration of artificial intelligence in mathematical reasoning marks a pivotal advancement in our quest to understand and utilize the very language of the universe. Mathematics, a discipline that stretches from the rudimentary principles of arithmetic to the complexities of algebra and calculus, serves as the bedrock for innovation across various fields, including science, engineering, and technology. The challenge, however, has always been to move beyond mere computation to achieve a level of reasoning and proof akin to human capability.

Significant advancements have been made in the field of large language models (LLMs) to confront this challenge head-on. Through their extensive training on diverse datasets, these models have demonstrated an ability to compute, reason, infer, and even prove mathematical theorems. This evolution from computation to reasoning represents a significant leap forward, offering new tools for solving some of mathematics’ most enduring problems.

InternLM-Math, a state-of-the-art model developed by Shanghai AI Laboratory in collaboration with prestigious academic institutions such as Tsinghua University, Fudan University, and the University of Southern California, is at the forefront of this evolution. InternLM-Math, an offspring of the foundational InternLM2 model, represents a paradigm shift in mathematical reasoning. It incorporates a suite of advanced features, including chain-of-thought reasoning, reward modeling, formal reasoning, and data augmentation, all within a unified sequence-to-sequence (seq2seq) framework. This comprehensive approach has positioned InternLM-Math as a frontrunner in the field, capable of tackling a wide range of mathematical tasks with unprecedented accuracy and depth.

The methodology behind InternLM-Math is as innovative as it is effective. The team has significantly enhanced the model’s reasoning capabilities by continuing the pre-training of InternLM2, focusing on mathematical data. Including chain-of-thought reasoning, in particular, allows InternLM-Math to approach problems step-by-step, mirroring the human thought process. Coding integration further bolsters this through the reasoning interleaved with the coding (RICO) technique, enabling the model to solve complex problems and generate proofs more naturally and intuitively.

The performance of InternLM-Math speaks volumes about its capabilities. On various benchmarks, including GSM8K, MATH, and MiniF2F, InternLM-Math has consistently outperformed existing models. Notably, it scored 30.3 on the MiniF2F test set without any fine-tuning, a testament to its robust pre-training and innovative methodology. Furthermore, the model’s ability to use LEAN for solving and proving mathematical statements showcases its versatility and potential as a tool for both research and education.

The implications of InternLM-Math’s achievements are far-reaching. By providing a model capable of verifiable reasoning and proof, Shanghai AI Laboratory has not only advanced the field of artificial intelligence. Still, it has also opened new avenues for exploration in mathematics. InternLM-Math’s ability to synthesize new problems, verify solutions, and even improve itself through data augmentation positions it as a pivotal tool in the ongoing quest to deepen our understanding of mathematics.

In summary, InternLM-Math represents a significant milestone in achieving human-like reasoning in mathematics through artificial intelligence. Its development by Shanghai AI Laboratory and academic collaborators marks an important step forward in our ability to solve, reason, and prove mathematical concepts, promising a future where AI-driven tools augment our understanding and exploration of the mathematical world.

Check out the Paper and Github. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Unlocking the Future of Mathematics with AI: Meet InternLM-Math, the Groundbreaking Language Model for Advanced Math Reasoning and Problem-Solving appeared first on MarkTechPost.

The progress and development of artificial intelligence (AI) heavily rely on human evaluation, guidance, and expertise. In computer vision, convolutional networks acquire a semantic understanding of images through extensive labeling provided by experts, such as delineating object boundaries in datasets like COCO or categorizing images in ImageNet. 

Similarly, in robotics, reinforcement learning often relies on human-defined reward functions to guide machines toward optimal performance. In Natural Language Processing (NLP), recurrent neural networks and Transformers can learn the intricacies of language from vast amounts of unsupervised text generated by humans. This symbiotic relationship highlights how AI models advance by leveraging human intelligence, tapping into the depth and breadth of human expertise to enhance their capabilities and understanding.

Researchers from Huawei introduced the concept of ” superalignment ” to address the challenge of effectively leveraging human expertise to supervise superhuman AI models. Superalignment aims to align superhuman models to maximize their learning from human input. A seminal concept in this area is Weak-to-Strong Generalization (WSG), which explores using weaker models to supervise stronger ones. 

WSG research has shown that stronger models can surpass their weaker counterparts in performance through simple supervision, even with incomplete or flawed labels. This approach has demonstrated effectiveness in natural language processing and reinforcement learning.

Researchers extend their idea to “vision superalignment,” specifically examining the application of Weak-to-Strong Generalization (WSG) within the context of vision foundation models. Multiple scenarios in computer vision, including few-shot learning, transfer learning, noisy label learning, and traditional knowledge distillation settings, were meticulously designed and examined. 

Their approach’s effectiveness stems from its capacity to blend direct learning from the weak model with the strong model’s inherent capability to comprehend and interpret visual data. By leveraging the guidance provided by the weak model while capitalizing on the advanced capabilities of the strong model, this method enables the strong model to transcend the constraints of the weak model, thereby enhancing its predictions.

However, to deal with the problems of weak models not providing precise guidance and strong models sometimes giving incorrect labels, one needs a smarter method than just mixing these labels. Since it’s hard to know how accurate each label is, in the future, researchers plan to use confidence as a measure to pick the most likely correct label. This way, by considering confidence levels, one can choose the best labels more effectively, making the model’s predictions more accurate and reliable overall.

Check out the Paper and Github. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Huawei Researchers Introduce a Novel and Adaptively Adjustable Loss Function for Weak-to-Strong Supervision appeared first on MarkTechPost.

In artificial intelligence, integrating multimodal inputs for video reasoning stands as a frontier, challenging yet ripe with potential. Researchers increasingly focus on leveraging diverse data types – from visual frames and audio snippets to more complex 3D point clouds – to enrich AI’s understanding and interpretation of the world. This endeavor aims to mimic human sensory integration and surpass it in depth and breadth, enabling machines to make sense of complex environments and scenarios with unprecedented clarity.

At the heart of this challenge is the problem of efficiently and effectively fusing these varied modalities. Traditional approaches have often fallen short, either by needing to be more flexible in accommodating new data types or necessitating prohibitive computational resources. Thus, the quest is for a solution that not only embraces the diversity of sensory data but does so with agility and scalability.

Current methodologies in multimodal learning have shown promise but are hampered by their computational intensity and inflexibility. These systems typically require substantial parameter updates or dedicated modules for each new modality, making the integration of new data types cumbersome and resource-intensive. Such limitations hinder the adaptability and scalability of AI systems in dealing with the richness of real-world inputs.

A groundbreaking framework proposed by UNC-Chapel Hill researchers was designed to revolutionize how AI systems handle multimodal inputs for video reasoning. This innovative approach introduces a modular, efficient system for fusing different modalities, such as optical flow, 3D point clouds, and audio, without requiring extensive parameter updates or bespoke modules for each data type. At its core, CREMA utilizes a query transformer architecture that integrates diverse sensory data, paving the way for a more nuanced and comprehensive AI understanding of complex scenarios.

CREMA’s methodology is notable for its efficiency and adaptability. Employing a set of parameter-efficient modules allows the framework to project diverse modality features into a common embedding space, facilitating seamless integration without overhauling the underlying model architecture. This approach conserves computational resources and ensures the model’s future-proofing, ready to accommodate new modalities as they become relevant.

CREMA’s performance has been rigorously validated across various benchmarks, demonstrating superior or equivalent results compared to existing multimodal learning models with a fraction of the trainable parameters. This efficiency does not come at the cost of effectiveness; CREMA adeptly balances the inclusion of new modalities, ensuring that each contributes meaningfully to the reasoning process without overwhelming the system with redundant or irrelevant information.

In conclusion, CREMA represents a significant leap forward in multimodal video reasoning. Its innovative fusion of diverse data types into a coherent, efficient framework not only addresses the challenges of flexibility and computational efficiency but also sets a new standard for future developments in the field. The implications of this research are profound, promising to enhance AI’s ability to interpret and interact with the world more nuanced and intelligently.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post CREMA by UNC-Chapel Hill: A Modular AI Framework for Efficient Multimodal Video Reasoning appeared first on MarkTechPost.

With all the changes happening in the world of Facebook advertising, staying ahead requires more than just a basic understanding of ad placements and targeting. 

As technology evolves and privacy concerns continue to intensify, advertisers tasked with maximizing ROI must continuously refine their approaches. 

This means going beyond the basics and focusing on things like advanced segmentation and targeting, AI and machine learning for optimization, creative testing, and more. 

The thing is, we know how much potential there is in the Facebook ecosystem, it’s why we are an official Meta partner. When done right, advertisers can see really exceptional results.

So let’s dive in and look at how we can maximize performance and take our Facebook ad strategies to the next level. 

Convert Website Visitors into Real Contacts!

Identify who is visiting your site with name, email and more. Get 500 contacts for free!

Please enable JavaScript in your browser to complete this form.Website / URL *Grade my website

Advanced Segmentation and Targeting 

We are big fans of segmentation here at Customers.ai. 

After all, detailed audience segmentation allows advertisers to create and deliver highly relevant content to specific audiences. By looking at demographics, interests, behaviors, and more, you can significantly increase engagement rates and improve overall campaign effectiveness. 

Seems like a good thing right?

There are several ways you can add advanced segmentation strategies to your ad campaigns. Let’s look at three: Custom Audiences, Lookalike Audiences & Advantage+ Audiences.

Custom Audiences

Custom audiences are created from existing data on your customers (think email lists or users who have previously interacted with your content).

While Facebook is pushing advertisers to Advantage+, we think custom audiences are way more valuable. After all, they are YOUR audience. 

And with the right data, you can create fantastic segments with highly targeted creative. This includes:

Website Traffic: Create custom audiences from users who have visited specific pages of your website. For example, with the Customers.ai Website Visitor ID X-Ray pixel, you can identify over 20% of site visitors, even if they don’t give you their information. That data can then be used to retarget individuals on Facebook. With Restore, you can get an even higher match rate (check out this post).

Engage Past Customers: No one wants a one and done buyer. Target users who have previously made a purchase with ads for related products or exclusive offers. Is it time for a refill? Is there an update on the product they purchased? Get creative and make your segments work for you.

Segment by Intent: We know that not every buyer is ready to make a purchase. That’s ok! That’s what ads are for. The ads for visitors who hit the shopping cart should be much different than the ads for visitors who hit the T-Shirts page. By being able to track which pages your users visit, you can create really specific segments based on intent. 

Lucky for you, our tools make all of this not just possible, but easy. Including our customer journey mapping feature which shows you all the pages your buyer visited.

And our Facebook integration which allows you to send audiences directly to your Facebook campaigns:

Lookalike & Advantage+ Audiences 

Lookalike audiences allow you to reach new people whose interests and behaviors are similar to those of your existing customers while Advantage+ audiences use Meta’s advanced AI to build your campaign audience.

The thing with both of these audiences is that they are only as good as the data you populate them with. 

So whether you are still using lookalike audiences or have made the switch to Advantage+, you want to keep the following in mind:

Start with a High-Quality Source Audience: You want to reach those most likely to buy right? That means your source audience must be people who…you guessed it, bought! Use your best-performing customer segments, such as those with high engagement or conversion rates, as the basis for creating lookalike audiences or informing your Advantage+ audiences.

Specify Audience Size and Similarity: In the case of lookalike audiences, size matters. While Facebook generally recommends a source audience between 1,000-5,000 people, smaller audiences can actually perform better. For Advantage+ audiences, more data is usually better. The more data you can “train” Facebook’s AI with, the better.

Update and Refine: Like anything in marketing, it’s important to continuously update and refine your source audiences to ensure your segments remain relevant and effective. 

Whatever kind of audience you decide on (maybe all of them!), segmentation not only elevates the efficiency of your ad spend but also drives superior conversion rates by aligning your content with the distinct needs and interests of each audience segment.

See Who Is On Your Site Right Now!

Turn anonymous visitors into genuine contacts.

Try it Free, No Credit Card Required

Get The X-Ray Pixel

Leveraging AI and Machine Learning for Optimization

AI and machine learning are revolutionizing how Facebook ad campaigns are optimized. 

We’re talking automated bidding processes, campaign data analysis, and real-time campaign optimization. AI algorithms can predict outcomes based on historical performance and real-time data, adjusting bids in milliseconds to maximize ad visibility and engagement. 

These dynamic capabilities ensure that advertisers not only reach their target audience more effectively but also optimize their ad spend, reducing CPA and increasing ROI. 

Here are a few examples of how AI is being used (and can be used) in Facebook advertising:

Dynamic Creative Optimization: AI can be used to automatically test different combinations of ad elements such as images, videos, headlines, descriptions, and CTAs. By analyzing the performance data in real-time, the system identifies the most effective combination for each segment of the audience. Even if you aren’t doing dynamic ad creation, AI tools can be really helpful for creative. 

Predictive Analytics for Audience Targeting: Facebook uses AI to analyze data on things like user behavior, preferences, and engagement patterns. This analysis helps predict which users are most likely to take a specific action (think making a purchase, clicking on a link, or engaging with content) and ensures ads are shown to these particular people. 

Audience Expansion: We already touched on how AI is driving audience expansion above but it’s worth noting just how big of an impact this is now having in ads. With iOS 14 and now the loss of cookies, Facebook audiences have shrunk dramatically. AI is helping Facebook (and you) to build back those audiences and ensure targeting capabilities don’t completely fade away. 

AI is here to stay and as and advertiser, you can use it to your advantage or you can let your competitors use it to their advantage. 

Creative Testing and Iteration

Testing isn’t exactly an “advanced” technique in the general sense. However, the process of creative testing can be. 

Testing allows you to understand which elements of your ads resonate most and it allows you to see results quicker. 

Let’s look at a few creative testing strategies:

A/B Testing: At its core, A/B testing is simple – compare two versions of an ad and see which performs better. The key here is variable isolation; change one element at a time—be it the image, headline, ad copy, or CTA—while keeping all other variables constant. By doing it this way, you can get clear insights into which specific changes improve ad performance.

Multivariate Testing: For a more comprehensive analysis, multivariate testing lets you test multiple variations of several elements at once. While it can be a bit more challenging, this approach can show how different elements interact with one another and their combined effect on ad performance. The key to multivariate testing is a large audience and statistically significant results.

Sequential Testing: Sequential testing involves giving your audience a series of ad variations over time. While more time-consuming than A/B testing, this strategy can be particularly useful for understanding how changes in creative elements impact ad fatigue and engagement over longer campaigns. The goal – figuring out the optimal frequency and timing for refreshing ad creatives.

In the analysis phase, having the right tools in place is key. Whether it’s Facebook’s analytics tools or third-party platforms, you will need something to understand ad performance and test results. 

By prioritizing creative testing and iteration, advertisers can significantly enhance the effectiveness of their Facebook ads. This cycle of analysis and iteration ensures that ad strategies remain dynamic, targeted, and continually optimized.

Making Facebook Ads Work for You  

As we said earlier, there is so much potential in Facebook Ads. I think sometimes we forget there is actually a whole Meta ecosystem of Facebook, Instagram, WhatsApp, and Messenger, giving us multiple places to reach our audience.

But we won’t be successful if we rely on old strategies. 

In order to make Facebook ads work for you, you must go beyond the basics. We need advanced segmentation and targeting, we need to capitalize on AI technology, and we certainly need to remember to test, test, test.

And if you really want to go above and beyond, you need Customers.ai. 

Our Restore product can skyrocket your Facebook reach and allow you to target website visitors you had no idea existed. 

Want to learn more? Try it free or contact our sales team for more information.

Convert Website Visitors into Real Contacts!

Identify who is visiting your site with name, email and more. Get 500 contacts for free!

Please enable JavaScript in your browser to complete this form.Website / URL *Grade my website

Important Next Steps

See what targeted outbound marketing is all about. Capture and engage your first 500 website visitor leads with Customers.ai X-Ray website visitor identification for free.

Talk and learn about sales outreach automation with other growth enthusiasts. Join Customers.ai Island, our Facebook group of 40K marketers and entrepreneurs who are ready to support you.

Advance your marketing performance with Sales Outreach School, a free tutorial and training area for sales pros and marketers.

The post Maximizing ROI with Advanced Facebook Ad Strategies appeared first on Customers.ai.

In the vast world of artificial intelligence, developers face a common challenge – ensuring the reliability and quality of outputs generated by large language models (LLMs). The outputs, like generated text or code, must be accurate, structured, and aligned with specified requirements. These outputs may contain biases, bugs, or other usability issues without proper validation.

While developers often rely on LLMs to generate various outputs, there is a need for a tool that can add a layer of assurance, validating and correcting the results. Existing solutions are limited, often requiring manual intervention or lacking a comprehensive approach to ensure both structure and type guarantees in the generated content. This gap in the existing tools prompted the development of Guardrails, an open-source Python package designed to address these challenges.

Guardrails introduces the concept of a “rail spec,” a human-readable file format (.rail) that allows users to define the expected structure and types of LLM outputs. This spec also includes quality criteria, such as checking for biases in generated text or bugs in code. The tool utilizes validators to enforce these criteria and takes corrective actions, such as reasking the LLM when validation fails.

One of Guardrails‘ notable features is its compatibility with various LLMs, including popular ones like OpenAI’s GPT and Anthropic’s Claude, as well as any language model available on Hugging Face. This flexibility allows developers to integrate Guardrails seamlessly into their existing workflows.

To showcase its capabilities, Guardrails offers Pydantic-style validation, ensuring that the outputs conform to the specified structure and predefined variable types. The tool goes beyond simple structuring, allowing developers to set up corrective actions when the output fails to meet the specified criteria. For example, if a generated pet name exceeds the defined length, Guardrails triggers a reask to the LLM, prompting it to generate a new, valid name.

Guardrails also supports streaming, enabling users to receive validations in real-time without waiting for the entire process to complete. This enhancement enhances efficiency and provides a dynamic way to interact with the LLM during the generation process.

In conclusion, Guardrails addresses a crucial aspect of AI development by providing a reliable solution to validate and correct the outputs of LLMs. Its rail spec, Pydantic-style validation, and corrective actions make it a valuable tool for developers striving to enhance AI-generated content’s accuracy, relevance, and quality. With Guardrails, developers can navigate the challenges of ensuring reliable AI outputs with greater confidence and efficiency.
The post Meet Guardrails: An Open-Source Python Package for Specifying Structure and Type, Validating and Correcting the Outputs of Large Language Models (LLMs) appeared first on MarkTechPost.

Graph-based machine learning is undergoing a significant transformation, largely propelled by the introduction of Graph Neural Networks (GNNs). These networks have been pivotal in harnessing the complexity of graph-structured data, offering innovative solutions across various domains. Despite their initial success, traditional GNNs face critical challenges, particularly those relying on local message-passing mechanisms. They need help managing long-range dependencies within graphs and often encounter the issue of over-squashing, where information from distant nodes is compressed excessively as it passes through the network layers.

Graph Mamba Networks (GMNs) by researchers from Cornell University emerge as a groundbreaking solution to these challenges. By integrating the principles of State Space Models (SSMs), widely celebrated for their efficiency and effectiveness across different data modalities, GMNs offer a novel approach to graph learning. This innovative framework is designed to overcome the limitations of both traditional GNNs and their more recent advancements, such as Graph Transformers, which, despite their promise, grapple with scalability due to their quadratic computational requirements.

At the heart of GMNs lies a meticulously crafted architecture that embraces neighborhood tokenization, token ordering, and a bidirectional selective SSM encoder, among other features. This structure enhances the network’s ability to capture and model long-range dependencies effectively and addresses the computational and structural constraints that have hampered previous models. GMNs adopt a selective approach to SSM application on graph data, enabling more nuanced and efficient handling of the inherent complexities of graph-structured information.

The introduction of GMNs into the landscape of graph-based machine learning is not without empirical validation. Rigorous testing across a spectrum of benchmarks reveals that GMNs excel in tasks requiring modeling long-range interactions within graphs. This exceptional performance is not just a testament to the architectural ingenuity of GMNs but also highlights the strategic leverage of SSMs’ strengths in a graph-learning context. GMNs distinguish themselves through their computational efficiency, setting a new standard in the field.

GMNs stand out as a beacon of progress. They signify a major leap in our capacity to learn from graph-structured data and open up a myriad of possibilities for exploration and application. From analyzing complex social networks to deciphering the intricate molecular structures that define life, GMNs offer a robust and efficient framework for understanding how data connects and interacts.

In conclusion, the advent of Graph Mamba Networks marks a pivotal moment in graph-based machine learning:

GMNs adeptly incorporate state space models to address the limitations of traditional GNNs and Graph Transformers, paving the way for more efficient graph learning.

The unique architecture of GMNs, featuring neighborhood tokenization and a bidirectional selective SSM encoder, enables the nuanced handling of graph-structured data.

Demonstrated through extensive benchmarks, GMNs excel in capturing long-range dependencies within graphs, showcasing superior performance and remarkable computational efficiency.

GMNs open new avenues for research and application across various domains by enhancing our ability to model and understand graph-structured data.

Check out the Paper. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post Cornell Researchers Introduce Graph Mamba Networks (GMNs): A General Framework for a New Class of Graph Neural Networks Based on Selective State Space Models appeared first on MarkTechPost.

In the fast-paced world of technology, where innovation often outpaces human interaction, LAION and its collaborators at the ELLIS Institute Tübingen, Collabora, and the Tübingen AI Center are taking a giant leap towards revolutionizing how we converse with artificial intelligence. Their brainchild, BUD-E (Buddy for Understanding and Digital Empathy), seeks to break down the barriers of stilted, mechanical responses that have long hindered our immersive experiences with AI voice assistants.

The journey began with a mission to create a baseline voice assistant that not only responded in real time but also embraced natural voices, empathy, and emotional intelligence. The team recognized the shortcomings of existing models, focusing on reducing latency and enhancing the overall conversational quality. The result? A carefully evaluated model boasts response times as low as 300 to 500 ms, setting the stage for a more seamless and responsive interaction.

However, the developers acknowledge that the road to a truly empathic and natural voice assistant is still in progress. Their open-source initiative invites contributions from a global community, emphasizing the need to tackle immediate problems and work towards a shared vision.

One key area of focus is the reduction of latency and system requirements. The team aims to achieve response times below 300 ms through sophisticated quantization techniques and fine-tuning streaming models, even with larger models. This dedication to real-time interaction lays the groundwork for an AI companion that mirrors the fluidity of human conversation.

The quest for naturalness extends to speech and responses. Leveraging a dataset of natural human dialogues, the developers are fine-tuning BUD-E to respond similarly to humans, incorporating interruptions, affirmations, and thinking pauses. The goal is to create an AI voice assistant that not only understands language but also mirrors the nuances of human expression.

BUD-E’s memory is another remarkable feature in development. With tools like Retrieval Augmented Generation (RAG) and Conversation Memory, the model aims to keep track of conversations over extended periods, unlocking a new level of context familiarity.

The developers are not stopping there. BUD-E is envisioned to be a multi-modal assistant, incorporating visual input through a lightweight vision encoder. The incorporation of webcam images to evaluate user emotions adds a layer of emotional intelligence, bringing the AI voice assistant closer to understanding and responding to human feelings.

Building a user-friendly interface is also a priority. The team plans to implement LLamaFile for easy cross-platform installation and deployment, introducing an animated avatar akin to Meta’s Audio2Photoreal. A chat-based interface capturing conversations in writing and providing ways to capture user feedback aims to make the interaction intuitive and enjoyable.

Furthermore, BUD-E is not limited by language or the number of speakers. The developers are extending streaming Speech-to-Text to more languages, including low-resource ones, and plan to accommodate multi-speaker environments seamlessly.

In conclusion, the development of BUD-E represents a collective effort to create AI voice assistants that engage in natural, intuitive, and empathetic conversations. The future of conversational AI looks promising as BUD-E stands as a beacon, lighting the way for the next era of human-technology interaction.

Check out the Code and Blog. All credit for this research goes to the researchers of this project. Also, don’t forget to follow us on Twitter and Google News. Join our 37k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group.

If you like our work, you will love our newsletter..

Don’t Forget to join our Telegram Channel
The post LAION Presents BUD-E: An Open-Source Voice Assistant that Runs on a Gaming Laptop with Low Latency without Requiring an Internet Connection appeared first on MarkTechPost.