Alexander Siegenfeld

🌟 𝐄𝐱𝐜𝐢𝐭𝐢𝐧𝐠 𝐍𝐞𝐰𝐬!  I'm delighted to announce that our paper, "Deep Exploration of Cross-Lingual Zero-Shot Generalization in Instruction Tuning," co-first authored with Changho Lee, has been accepted for publication at ACL 2024 (findings track)! https://lnkd.in/gMFiUevc ✅ 𝐊𝐞𝐲 𝐈𝐧𝐬𝐢𝐠𝐡𝐭𝐬: Our study introduces the innovative KORANI meta-dataset, which allows for direct comparisons between cross-lingual and monolingual instruction tuning. We found that cross-lingual tuning matches the performance of monolingual approaches. This emphasizes the importance of task relevance across languages rather than strict linguistic consistency. Notably, our findings suggest that cross-lingual instruction tuning could be a practical alternative to monolingual tuning in low-resource language scenarios. 📢 𝐅𝐨𝐫 𝐭𝐡𝐞 𝐊𝐨𝐫𝐞𝐚𝐧 𝐋𝐚𝐧𝐠𝐮𝐚𝐠𝐞 𝐂𝐨𝐦𝐦𝐮𝐧𝐢𝐭𝐲: The KORANI meta-dataset comprises 51 diverse benchmarks, complete with training and test sets. This rich resource is shared openly for researchers and practitioners interested in assessing and enhancing the performance of Korean language models from various aspects. We highly encourage its use to explore new frontiers in language model performance! #𝐀𝐂𝐋𝟐𝟎𝟐𝟒 #𝐋𝐆𝐀𝐈𝐑𝐞𝐬𝐞𝐚𝐫𝐜𝐡

69

Stuff we've been up to in the last several months that reviewers also liked: With Kamron Zaidi (EngSci 2T4), we trained a neural network to identify moves in chess that people view as brilliant. (Brilliant move ≠ good move. An obvious checkmate is good but not brilliant). The system predicts moves to be identified as brilliant based on the shape of the game tree: long thin paths are probably more brilliant. Also, moves are predicted to be perceived as brilliant if a strong system rates them higher but a weak system rates them lower. https://lnkd.in/g-jCdzR6 (International Conference on Computational Creativity 2024) With Amy Saranchuk (EngSci 2T4), we identify an intriguing property of the self-supervised pre-training process when using the rotation task: the performance of the system seems to have a periodic response to the rotation angle (makes sense) with the specific response showing very different and sometimes contradictory patterns of response to the rotation angle (mysterious). (At the Conference on Robots and Vision Workshops 2024). https://lnkd.in/gMaUMiyw With Juho Kim (EngSci 2T4), we leverage large-scale analysis using LLMs to show that Southern US honour culture https://lnkd.in/gWhB8buP) leaves its traces in internet flame wars on reddit: redditors geolocated to the US South are statistically more likely to retaliate (verbally) in response to personal attacks on the internet. (International Conference on Computational Social Science 2024) With Yiping (Sandra) Wang and Cameron Smith (EngSci 2T4), we propose synthetic datasets for exploring how computer vision networks classify images when colour is an important cue. (Usually the shape is much more important) (At the Conference on Robots and Vision Workshops 2024). Preliminary preprint: https://lnkd.in/gpgnYQUj With Jason Zuo (EngSci 2T4), we explore how Transformers can be trained on tasks that require position information without positional encodings. Intruiguingly, we find that correlations of the activations in the Transformer layers indicate that even untrained Transformers contain position information https://lnkd.in/gaKkMigS (At the Meeting of the North American Chapter Association of Computational Linguistics Student Research Workshop 2024) With Tracy Qian (EngSci 2T4) and building on work with Max Piasevoli , we show that the critic component of a WGAN can be used for model selection when fitting hierarchical models https://lnkd.in/gQVjS33K (At the Statistical Society of Canada Annual Meeting 2024) With Kexin Chen (MIE M.Eng '24), we've trained interpretable LLMs on reddit data to identify common patterns in posts of people with suspected Lyme disease and reported mental health symptoms (at the AI in Medicine 2024 conference)

55

1 Comment

I’d like to share a review/perspective paper that I co-authored with my Ph.D. advisor, Yan Wang and Jiayu Chen (from Tsinghua University, Beijing). In this paper, we discuss: 1) the key components of scenario planning and how AI can contribute to each of them; 2) the challenges involved and potential solutions; and 3) future research directions for advancing the field. The article was published in Engineering and can be accessed here: https://lnkd.in/gKGsXgc2 I am also looking for potential collaborations with other researchers interested in this direction: AI + resilience planning + scenario construction.

14

Greetings Network, I'm excited to share that I won the recent Apart Research X Goodfire Research hackathon focused on reprogramming Large Language Models (LLMs)! My project introduced a novel framework for aligning LLMs using Feature Steering (FS) automated through Reinforcement Learning (RL), achieving a remarkable 3X performance improvement in a Llama 3 8B model playing Tic Tac Toe – all without the need for backpropagation in the LLM itself. Learn more about the project here: https://lnkd.in/e-RzbzFg. Detailed Summary: Traditional alignment methods like Reinforcement Learning from Human Feedback (RLHF), while effective, can sometimes lead to undesirable model behaviours. Additionally, the sheer size and complexity of LLMs make them difficult to interpret and understand at a high level. Sparse Autoencoders (SAEs) offer a solution by extracting human-understandable concepts or features from these models. However, training SAEs is computationally expensive, and any subsequent model tuning renders the discovered features useless. This is where Feature Steering comes in. It allows us to directly influence a model's output by adjusting the importance of specific features. For example, we could amplify a feature related to "The Golden Gate Bridge" and make the model focus on it extensively. While powerful, feature steering is typically done manually which limits its potential. My framework automates this process using RL. The RL agent intelligently learns to adjust feature strengths based on a defined reward function, making the process dynamic and adaptable. Additionally, the framework is agnostic to the underlying LLM with an SAE, RL algorithm, and task, as long as a valid reward function is available. In the hackathon, I used a Proximal Policy Optimization (PPO) model to optimize feature strengths in a Llama 3 8B model to improve its Tic Tac Toe gameplay against an optimal opponent. The results were encouraging: a 3X improvement in draw rate compared to the baseline. The RL agent also demonstrated intelligent and interpretable behaviour, suppressing features associated with illegal moves and strengthening those linked to optimal strategies. From these initial results, I'm eager to explore the framework's effectiveness across different scenarios, including RLHF. I'm incredibly grateful to Apart Research and Goodfire for this incredible opportunity and for providing an API that makes high-level Feature Steering possible (something that required extensive knowledge before). I also want to thank the judges and keynote speakers for their encouraging comments and insightful feedback that greatly influenced my project. Moreover, it was amazing to read the submissions of other participants who used Feature Steering in innovative ways.

34

4 Comments

Can we leverage the strengths of both Transformer and Mamba to excel in tasks requiring both long-term context and efficient sequential processing? The Bamba-9B model does exactly that! Bamba-9B is an inference-efficient hybrid Mamba2 model. It not only achieves competitive performance against state-of-the-art transformer models such as LLama3.1-8B but also delivers 2.5x throughput improvements and 2x lower latency! Bamba is the result of a fantastic collaboration between IBM, UIUC, CMU, and Princeton. And the best part? It's publicly available for everyone to explore and experiment with! - Tech blog: https://lnkd.in/gSgNZnhf - GitHub: https://lnkd.in/gf6M_M8E Raghu Ganti, Mudhakar Srivatsa, Davis Wertheimer

24

🚀 Attention all software engineers! ACM has launched the first issue of Proceedings of the ACM on Software Engineering (PACMSE), a Gold Open Access journal dedicated to advancing software engineering research. PACMSE covers a wide range of topics, from design to deployment, aiming to improve software quality. Each issue of the journal comprises a curated selection of papers, which will be announced through publicized calls for papers. For more detail on the submission process, refer to intro in the inaugural issue: https://bit.ly/3YwVBti  📚 #SoftwareEngineering #OpenAccess #ACM

59

1 Comment

In this new paper we explore the use of classical programming techniques in implementing the quantum lattice Boltzmann method in the Intel Quantum SDK. https://lnkd.in/gpDSFe6b

14

I will be teaching an ICPSR Summer Workshop, "Sensitivity Analysis: Quantifying the Robustness of Inferences to Alternative Factors or Data", Hybrid (virtual, FtF in Ann Arbor, MI, with asynchronous access to materials). https://lnkd.in/etB9AR8h Most social scientists need to consider the unobserved conditions that could change an inference. This workshop is about the statistical tools (R, Stata or the konfound-it app https://lnkd.in/ep_z-DJS ) to do just that, quantifying the strength of evidence to inform policy, practice, or general social science. The workshop is a full 20 instruction hours of instruction including opportunities to adapt to specific models such as mediation, moderation, diff in diff, logistic regression, hazard/cox regression, multilevel models, differential attrition from treatment groups, etc. Go here to register: https://lnkd.in/etB9AR8h Group discounts available as well. https://lnkd.in/eWRv8G7q

10

In a recent paper, we (me, Georg Link, PhD, and Kevin M. Lumbard, PhD) explore how Open Source Software projects' health and sustainability depend on the human activity invested by project maintainers and contributors. Below is a brief summary, longer version in blog post and paper :) * The activity invested is labeled Maintenance Labor and can accordingly originate either from the maintainers (Maintainer Labor) or the contributors (Contributor Labor). The latter includes the full spectra between long-term to episodic contributors. * The Maintenance Labor is subtractable, meaning its availability is limited to the human activity invested, and can be depleted, e.g., due to abandonment or limitations among current maintainers. * The Labor can be compared to a common-pool resource, as also excluding others from benefitting of the maintainers' output is difficult. * To enable a long-term and viably maintained OSS project, the availability of Maintenance Labor needs to be sustained from both sources (maintainers and contributors). * Human infrastructure can help sustain this availability and need, especially considering human factors, e.g., enabling a sustainable work-life balance, managing social pressure, promoting inclusiveness, and setting up guard rails for managing toxicity. * There is a dependency between the two categories of Maintenance Labor where a limited amount of Contributor Labor creates a burden and dependency on the maintainer, while too high amounts may create too large of overhead for maintainers, resulting in increased stress and pressure. * Attraction of new contributors is considered pivotal for a healthy project but needs to be balanced with an onboarding process that can empower the newcomers to make better contributions, requiring less attention from the maintainers. * Non-code contributions, such as helping with support, community management, and marketing, are highlighted as important contributions. * Finally, Resource Funding is seen as a means of enabling maintainers and contributors to create a healthy work-life balance while also improving the project's health. * Implications: - Policy needs to recognize the sensitivity of the Maintenance Labor and its role in sustaining the digital commons and infrastructure that underpins society today. - Initaitives and human infrastructure are needed to support maintainers through funding and reduce overhead in managing volatile communities while maintaining a healthy work-life balance and in-project working environment. - Users of the OSS, especially industry and governments, must consider their role and responsibility in sustaining the OSS and the human activity that goes into its long-term maintenance. For further reading, see https://lnkd.in/dpZwrtyY

46

6 Comments

New in the AFP: The Category of Sets An axiomatization of category SET on top of HOL, following the presentation given by Halvorson, including Cantor's diagonalization theorem and mathematical induction. https://lnkd.in/g6WTQ-MX

13

After seeing one of the authors at this recent and captivating conference: https://lnkd.in/gUrZcaiR I wanted to post a few excerpts from one of the most important AI papers from last year. “Rebalancing AI techno-optimism draws on the “productivity bandwagon”: a deep-rooted belief that technological change— including automation—drives higher productivity, which raises net wages and generates shared prosperity. Such optimism is at odds with the historical record and seems particularly inappropriate for the current path of “just let AI happen,” which focuses primarily on automation (replacing people). We must recognize that there is no singular, inevitable path of development for new technology… […] Many new technologies, such as industrial robots, expand the set of tasks performed by machines and algorithms, displacing workers. Automation raises average productivity but does not increase, and in fact may reduce, worker marginal productivity. Over the past four decades, automation has raised productivity and multiplied corporate profits, but it has not led to shared prosperity in industrial countries. Replacing workers with machines is not the only way to improve economic efficiency—and history has proved this, some innovations boost how much individuals contribute to production. For example, new software tools that aid car mechanics and enable greater precision can increase worker marginal productivity. This is completely different from installing industrial robots with the goal of replacing people. […] The productivity bandwagon is not activated, however, when the productivity gains from automation are small—what we call “so-so automation.” For example, self-checkout kiosks in grocery stores bring limited productivity benefits because they merely shift the work of scanning items from employees to customers. When stores introduce self-checkout kiosks, fewer cashiers are employed, but there is no major productivity boost to stimulate the creation of new jobs… Even nontrivial productivity gains from automation can be offset when they are not accompanied by new tasks. For example, in the American Midwest, the rapid adoption of robots has contributed to mass layoffs and ultimately prolonged regional decline. The situation is similarly troubling for workers when new technologies focus on surveillance. Increased monitoring of workers may lead to some small improvements in productivity, but its main function is to extract more effort from workers…” In short they say “AI offers an opportunity to complement worker skill and expertise if we direct its development accordingly.”

16

2 Comments

Check out the latest paper by Moritz Zanger published at #ICLR-2024. The paper, titled "Diverse Projection Ensembles for Distributional Reinforcement Learning," explores the use of projection ensembles to enhance diversity and exploration in distributional reinforcement learning. By combining different projections and representations, the proposed method significantly improves performance on a variety of tasks, particularly in directed exploration problems. For more details, please follow the link: https://lnkd.in/gYrQda7g Authors: Moritz Zanger, Wendelin Boehmer, Matthijs Spaan #Research #AI #EuRobotics #EuResearch #machinelearning #artificialintelligence #h2020 #FET #EUproject

7

Here is a link to a new paper by Larry Suter and me illustrating how cluster analysis can be used to study the semantic structure of mathematics standards and item specifications. We also explore what it means to measure measurement. https://lnkd.in/evG5JWDM

17

5 Comments

Glad to see this paper appearing in the Stevens News. Nowadays the improved LLMs know to reject many problems that are very wrong, but still, it's quite easy to find adversarially helpful examples. We still recommend LLM users (especially when using LLMs to explain unknown problems) fact-check the explanations.

43

Ran this research through an academic review pilot we are designing in the background. Bottom line - not a high quality research report due to methodological and data gaps. Excerpt from the report (1-10 scoring scale): 3. Methodological Evaluation [Score: 6] Research Design Appropriateness: The experimental design aligns with the methodologies used by leading AI organizations, making it appropriate for the research objectives. However, the slight improvements to the agent scaffolding may introduce variables that are not fully accounted for. Methodology Clarity and Replicability: The methodology provides a general overview but lacks detailed descriptions of the experimental setup, such as specific configurations, safeguards, and the exact nature of improvements made to the agent scaffolding. This limits replicability. Sampling Techniques and Sample Size: The study uses two AI models and conducts ten trials for each. While this offers preliminary insights, the sample size is relatively small for drawing broad conclusions. Control Measures and Variable Handling: There is insufficient information on control measures implemented to isolate variables and ensure that observed behaviors are solely due to the AI systems' capabilities rather than external factors. Statistical Analysis Validity: The analysis primarily reports success rates without deeper statistical examination, such as hypothesis testing or confidence intervals, limiting the robustness of the conclusions. 4. Data Integrity and Analysis [Score: 5] Data Quality and Transparency: The paper mentions observations and outcomes but does not provide raw data or detailed logs. The lack of data transparency hinders verification and independent analysis. Analysis Rigor: The analysis is descriptive and lacks quantitative depth. There is minimal discussion of data processing methods or how conclusions were derived from observations. Results Presentation and Visualization: While figures and diagrams are included, they may not sufficiently illuminate the data due to potential lack of clarity or detail. More effective visualization could enhance understanding. Outlier Handling and Justification: There is no discussion of outliers or anomalies in the trials, nor how they were handled or analyzed, which could be significant in understanding the variability of results. Data Accessibility: Raw data and experimental logs are not made available, limiting the ability of other researchers to validate or build upon the findings. ........... 9. Quality Indicators Checklist ✓ Clear research questions and hypotheses ✗ Appropriate controls and variable management (Not fully detailed) ✗ Statistical power analysis conducted ✗ Methods are replicable (Due to insufficient methodological detail) ✗ Data transparency ensured ✓ Limitations are discussed ✗ Conflicts of interest disclosed (Not disclosed) ✓ Ethical considerations addressed ---------

3

5 Comments

Great learning materials on Equivariant Neural Networks put together by Wenhan Gao, along with toy code to play with. If you are interested in Group CNNs, Steerable CNNs, Spherical Tensors, Cartesian Tensors, Irreps, Tensor Product, Spherical Harmonics, Group Averaging, Frame Averaging, etc, can check this for an easy start: https://lnkd.in/enfwyenj Other topics that Wenhan is working on are here: https://lnkd.in/ebmjMwm4 #Equivariance, #Symmetries, #AI4Science, #SciML, #GeometricDeepLearning

34