Maelstrom networks, logical explanations, erasing concepts, and new perspectives on AI Alignment
In the News
🗞 This Week in News
Pulse on AI in 2024 - a brief article on current public views on AI from the Institute for Human-Centered Artificial Intelligence at Stanford University.
🥁 Interesting Products & Features
Custom “Gems” from Gemini now rolling out - We heard about these awhile back at Google I/O. Gems lets you customize Gemini to create your own personal AI experts on any topic you want. Only available to paid Gemini users.
Hermes Function-Calling V1 Dataset - designed to train LLM models in performing function calls and returning structured output based on natural language instructions.
Higgs Llama V2 - Improved model from Boson AI. Interestingly, much of the performance boost of Higgs v2 comes from an improved judging system, which guides the model alignment through synthetic feedback signals. They built an in-house LLM reward model, named Higgs Judger, to evaluate model outputs.
📄 Interesting Papers
Spiking Diffusion Models - Diffusion models have achieved great success in image synthesis through iterative noise estimation using deep neural networks. However, the slow inference, high memory consumption, and computation intensity of the noise estimation model hinder the efficient adoption of diffusion models. The authors propose to overcome these challenges by introducing spiking diffusion models - with a Temporal-wise Spiking Mechanism (TSM) that allows spiking neural networks to capture more temporal features from a bio-plasticity perspective. Shows substantial energy efficiency, consuming ∼30% of the energy required by the ANN model, while still delivering superior generative outcomes. Authors from various institutions, including Hong Kong University and North Carolina State University.
The Mamba in the Llama: Distilling and Accelerating Hybrid Models - demonstrates that it is feasible to distill large Transformers into linear RNNs by reusing the linear projection weights from attention layers. The resulting hybrid model, which incorporates a quarter of the attention layers, achieves performance comparable to the original Transformer in chat benchmarks and outperforms open-source hybrid Mamba models trained from scratch with trillions of tokens in both chat benchmarks and general benchmarks. Authors from Cornell University and Together AI.
STEREO: Towards Adversarially Robust Concept Erasing from Text-to-Image Generation Models - Large-scale diffusion models for text-to-image generation are susceptible to adversarial attacks that can regenerate harmful concepts despite erasure efforts. This paper introduces a new robust concept erasing strategy - STEREO - designed to prevent this regeneration while preserving the model's ability to generate benign content. Authors from Mohamed bin Zayed University of Artificial Intelligence.
Towards Symbolic XAI – Explanation Through Human Understandable Logical Relationships Between Features - Attributes relevance to symbolic queries expressing logical relationships between input features, thereby capturing the abstract reasoning behind a model’s predictions. The methodology is built upon a simple yet general multi-order decomposition of model predictions. This decomposition can be specified using higher-order propagation-based relevance methods, such as GNN-LRP, or perturbation-based explanation methods commonly used in XAI. They demonstrate Symbolic XAI in the natural language processing (NLP), vision, and quantum chemistry. The Symbolic XAI framework provides an understanding of the model’s decision-making process that is both flexible for customization by the user and human-readable through logical formulas. Authors from Berlin Institute for the Foundations of Learning and Data.
Text2SQL is Not Enough: Unifying AI and Databases with TAG - proposes Table-Augmented Generation (TAG), a unified and general-purpose paradigm for answering natural language questions over databases. The TAG model represents a wide range of interactions between the LM and database that have been previously unexplored. Authors from UC Berkeley and Stanford.
Maelstrom Networks - Proposal for a new type of neural network that combines the strength of recurrent networks, with the pattern matching capability of feed-forward neural networks. They leave the recurrent component - the Maelstrom - unlearned, and offload the learning to a powerful feed-forward network. This allows the network to leverage the strength of feed-forward training without unrolling the network, and allows for the memory to be implemented in new neuromorphic hardware. It endows a neural network with a sequential memory that takes advantage of the inductive bias that data is organized causally in the temporal domain, and imbues the network with a state that represents the agent’s “self”, moving through the environment. Authors from University of Maryland, College Park.
Beyond Preferences in AI Alignment - critiques current approach to AI Alignment and argues that these limitations motivate a reframing of the targets of AI alignment: Instead of alignment with the preferences of a human user, developer, or humanity-writ-large, AI systems should be aligned with normative standards appropriate to their social roles, such as the role of a general-purpose assistant. Authors from MIT, UC Berkeley, University College London, and University of Cambridge.
🧠 Sources of Inspiration
rerankers python library - a simple API for all popular rerankers that makes it easy to add rerankers to your RAG application
NVIDIA Inception Accelerator now accepting applications (free)!
Cover image from STEREO: Towards Adversarially Robust Concept Erasing from Text-to-Image Generation Models.