In this paper, we investigate the problem of "generation supervision" in large language models, and present a novel bicameral architecture to separate supervision signals from their core capability, helpfulness. Doppelg\"anger, a new module parallel to the underlying language model, supervises the generation of each token, and learns to concurrently predict the supervision score(s) of the sequences up to and including each token. In this work, we present the theoretical findings, and leave the report on experimental results to a forthcoming publication.

Using computerized verifiable formal languages like Lean 4 to prove mathematical theorems has a significant impact on mathematical formalization. Lean 4 offers prominent potential for advancing mathematical reasoning. However, existing efforts are limited to mathematical formalization languages in substantial online corpora and are dedicated to keeping pace with rapidly evolving languages. To bridge the gap between the traditional and computerized proof, my approach to formalizing theorem proving involves generating formal steps and complete proofs using Large Language Models (LLMs) based on Natural Language (NL) proofs. The method is to introduce the basic structure and tactics in general, determine how AI can assist the mathematical formalization process to improve its performance, and give examples of solving problems in Lean 4 comparing to NL, mainly in IMO, and a sample theorem proving in abstract algebra.

We introduce Instruct-SkillMix, an automated approach for creating diverse, high quality SFT data. The Instruct-SkillMix pipeline involves two stages, each leveraging an existing powerful LLM: (1) Skill extraction: uses the LLM to extract core "skills" for instruction-following, either from existing datasets, or by directly prompting the model; (2) Data generation: uses the powerful LLM to generate (instruction, response) data that exhibit a randomly chosen pair of these skills. Here, the use of random skill combinations promotes diversity and difficulty. Vanilla SFT (i.e., no PPO, DPO, or RL methods) on data generated from Instruct-SkillMix leads to strong gains on instruction following benchmarks such as AlpacaEval 2.0, MT-Bench, and WildBench. With just $4$K examples, LLaMA-3-8B-Base achieves 42.76% length-controlled win rate on AlpacaEval 2.0. To our knowledge, this achieves state-of-the-art performance among all models that have only undergone SFT (no RL methods) and competes with proprietary models such as Claude 3 Opus and LLaMA-3.1-405B-Instruct. Ablation studies also suggest plausible reasons for why creating open instruction-tuning datasets via naive crowd-sourcing has proved difficult. Introducing low quality answers ("shirkers") in $20\%$ of Instruct-SkillMix examples causes performance to plummet, sometimes catastrophically. The Instruct-SkillMix pipeline is flexible and is adaptable to other settings.

Object grasping is a crucial technology enabling robots to perceive and interact with the environment sufficiently. However, in practical applications, researchers are faced with missing or noisy ground truth while training the convolutional neural network, which decreases the accuracy of the model. Therefore, different loss functions are proposed to deal with these problems to improve the accuracy of the neural network. For missing ground truth, a new predicted category probability method is defined for unlabeled samples, which works effectively in conjunction with the pseudo-labeling method. Furthermore, for noisy ground truth, a symmetric loss function is introduced to resist the corruption of label noises. The proposed loss functions are powerful, robust, and easy to use. Experimental results based on the typical grasping neural network show that our method can improve performance by 2 to 13 percent.

Recent research has seen significant interest in methods for concept removal and targeted forgetting in diffusion models. In this paper, we conduct a comprehensive white-box analysis to expose significant vulnerabilities in existing diffusion model unlearning methods. We show that the objective functions used for unlearning in the existing methods lead to decoupling of the targeted concepts (meant to be forgotten) for the corresponding prompts. This is concealment and not actual unlearning, which was the original goal. The ineffectiveness of current methods stems primarily from their narrow focus on reducing generation probabilities for specific prompt sets, neglecting the diverse modalities of intermediate guidance employed during the inference process. The paper presents a rigorous theoretical and empirical examination of four commonly used techniques for unlearning in diffusion models. We introduce two new evaluation metrics: Concept Retrieval Score (CRS) and Concept Confidence Score (CCS). These metrics are based on a successful adversarial attack setup that can recover forgotten concepts from unlearned diffusion models. The CRS measures the similarity between the latent representations of the unlearned and fully trained models after unlearning. It reports the extent of retrieval of the forgotten concepts with increasing amount of guidance. The CCS quantifies the confidence of the model in assigning the target concept to the manipulated data. It reports the probability of the unlearned model's generations to be aligned with the original domain knowledge with increasing amount of guidance. Evaluating existing unlearning methods with our proposed stringent metrics for diffusion models reveals significant shortcomings in their ability to truly unlearn concepts. Source Code: https://respailab.github.io/unlearning-or-concealment

Information retrieval (IR) is the task of finding relevant documents in response to a user query. Although Spanish is the second most spoken native language, current IR benchmarks lack Spanish data, hindering the development of information access tools for Spanish speakers. We introduce MessIRve, a large-scale Spanish IR dataset with around 730 thousand queries from Google's autocomplete API and relevant documents sourced from Wikipedia. MessIRve's queries reflect diverse Spanish-speaking regions, unlike other datasets that are translated from English or do not consider dialectal variations. The large size of the dataset allows it to cover a wide variety of topics, unlike smaller datasets. We provide a comprehensive description of the dataset, comparisons with existing datasets, and baseline evaluations of prominent IR models. Our contributions aim to advance Spanish IR research and improve information access for Spanish speakers.

Given an image query, visually prompted keyword localisation (VPKL) aims to find occurrences of the depicted word in a speech collection. This can be useful when transcriptions are not available for a low-resource language (e.g. if it is unwritten). Previous work showed that VPKL can be performed with a visually grounded speech model trained on paired images and unlabelled speech. But all experiments were done on English. Moreover, transcriptions were used to get positive and negative pairs for the contrastive loss. This paper introduces a few-shot learning scheme to mine pairs automatically without transcriptions. On English, this results in only a small drop in performance. We also - for the first time - consider VPKL on a real low-resource language, Yoruba. While scores are reasonable, here we see a bigger drop in performance compared to using ground truth pairs because the mining is less accurate in Yoruba.

The clustering of categorical data is a common and important task in computer science, offering profound implications across a spectrum of applications. Unlike purely numerical data, categorical data often lack inherent ordering as in nominal data, or have varying levels of order as in ordinal data, thus requiring specialized methodologies for efficient organization and analysis. This review provides a comprehensive synthesis of categorical data clustering in the past twenty-five years, starting from the introduction of K-modes. It elucidates the pivotal role of categorical data clustering in diverse fields such as health sciences, natural sciences, social sciences, education, engineering and economics. Practical comparisons are conducted for algorithms having public implementations, highlighting distinguishing clustering methodologies and revealing the performance of recent algorithms on several benchmark categorical datasets. Finally, challenges and opportunities in the field are discussed.

In the Naive Bayes classification model the class conditional densities are estimated as the products of their marginal densities along the cardinal basis directions. We study the problem of obtaining an alternative basis for this factorisation with the objective of enhancing the discriminatory power of the associated classification model. We formulate the problem as a projection pursuit to find the optimal linear projection on which to perform classification. Optimality is determined based on the multinomial likelihood within which probabilities are estimated using the Naive Bayes factorisation of the projected data. Projection pursuit offers the added benefits of dimension reduction and visualisation. We discuss an intuitive connection with class conditional independent components analysis, and show how this is realised visually in practical applications. The performance of the resulting classification models is investigated using a large collection of (162) publicly available benchmark data sets and in comparison with relevant alternatives. We find that the proposed approach substantially outperforms other popular probabilistic discriminant analysis models and is highly competitive with Support Vector Machines.

Large Language Model (LLM) pretraining traditionally relies on autoregressive language modeling on randomly sampled data blocks from web-scale datasets. We take inspiration from human learning techniques like spaced repetition to hypothesize that random data sampling for LLMs leads to high training cost and low quality models which tend to forget data. In order to effectively commit web-scale information to long-term memory, we propose the LFR (Learn, Focus, and Review) pedagogy, a new dynamic training paradigm which focuses and repeatedly reviews complex data blocks at systematic intervals based on the model's learning pace and progress. LFR records the model perplexities for different data blocks and frequently revisits blocks with higher perplexity which are more likely to be forgotten. We pretrain the GPT-2 models (124M - 1.5B) from scratch on the OpenWebText dataset using LFR. We test on downstream tasks from the language modeling, question answering, translation, and problem solving domains to achieve consistently lower perplexity and higher accuracy than the baseline OpenAI models, while obtaining a 20x pretraining speed-up.