Hassan Sajjad from Dalhousie University presented research on exploring the latent space of AI models to assess their safety and trustworthiness. He discussed use cases where analyzing latent space helps understand the robustness-generalization tradeoff in adversarial training and evaluate language comprehension. Sajjad's work aims to build better AI models and increase trust in their capabilities by looking at model internals. Why it matters: Intrinsic evaluation of model internals will become important to improving AI safety and robustness.
The paper introduces the Unscented Autoencoder (UAE), a novel deep generative model based on the Variational Autoencoder (VAE) framework. The UAE uses the Unscented Transform (UT) for a more informative posterior representation compared to the reparameterization trick in VAEs. It replaces Kullback-Leibler (KL) divergence with the Wasserstein distribution metric and demonstrates competitive performance in Fréchet Inception Distance (FID) scores.
MBZUAI Professor Kun Zhang is developing machine learning techniques to identify hidden causal variables, which are underlying concepts driving cause-and-effect relationships. Zhang and colleagues from Carnegie Mellon University are presenting a new approach for this at ICML 2024. Their method, causal representation learning, assumes that measured variables are generated by unobserved latent variables. Why it matters: Uncovering hidden causal relationships can significantly advance understanding in various fields by revealing the underlying mechanisms driving observed phenomena.
MBZUAI researchers presented a new machine learning method at ICLR for uncovering hidden variables from observed data. The method, called "complementary gains," combines two weak assumptions to provide identifiability guarantees. This approach aims to recover true latent variables reflecting real-world processes, while solving problems efficiently. Why it matters: The research advances disentangled representation learning by finding minimal assumptions necessary for identifiability, improving the applicability of AI models to real-world data.
This article discusses approximating a high-dimensional distribution using Gaussian variational inference by minimizing Kullback-Leibler divergence. It builds upon previous research and approximates the minimizer using a Gaussian distribution with specific mean and variance. The study details approximation accuracy and applicability using efficient dimension, relevant for analyzing sampling schemes in optimization. Why it matters: This theoretical research can inform the development of more efficient and accurate AI algorithms, particularly in areas dealing with high-dimensional data such as machine learning and data analysis.
Researchers from MBZUAI presented a new algorithm at ICLR 2024 that identifies causal relationships involving both observed and latent variables. The algorithm addresses limitations of existing methods that struggle with latent variables or assume observed variables don't directly influence latent variables. The proposed algorithm can accommodate both scenarios, offering a more generalizable approach to causal discovery. Why it matters: This research advances the development of AI systems that can analyze complex data and identify causal relationships, with potential applications in fields like medicine where understanding causality is crucial for developing treatments and preventative measures.
The paper introduces UAE-3D, a multi-modal VAE for 3D molecule generation that compresses molecules into a unified latent space, maintaining near-zero reconstruction error. This approach simplifies latent diffusion modeling by eliminating the need to handle multi-modality and equivariance separately. Experiments on GEOM-Drugs and QM9 datasets show UAE-3D establishes new benchmarks in de novo and conditional 3D molecule generation, with significant improvements in efficiency and quality.
The paper introduces the Prism Hypothesis, which posits a correspondence between an encoder's feature spectrum and its functional role, with semantic encoders capturing low-frequency components and pixel encoders retaining high-frequency information. Based on this, the authors propose Unified Autoencoding (UAE), a model that harmonizes semantic structure and pixel details using a frequency-band modulator. Experiments on ImageNet and MS-COCO demonstrate that UAE effectively unifies semantic abstraction and pixel-level fidelity, achieving state-of-the-art performance.