Rakesh Shivanna

Kiran Vodrahalli, Santiago Ontanon, Nilesh Tripuraneni et al. · deepmind

We introduce Michelangelo: a minimal, synthetic, and unleaked long-context reasoning evaluation for large language models which is also easy to automatically score. This evaluation is derived via a novel, unifying framework for evaluations over arbitrarily long contexts which measure the model's ability to do more than retrieve a single piece of information from its context. The central idea of the Latent Structure Queries framework (LSQ) is to construct tasks which require a model to ``chisel away'' the irrelevant information in the context, revealing a latent structure in the context. To verify a model's understanding of this latent structure, we query the model for details of the structure. Using LSQ, we produce three diagnostic long-context evaluations across code and natural-language domains intended to provide a stronger signal of long-context language model capabilities. We perform evaluations on several state-of-the-art models and demonstrate both that a) the proposed evaluations are high-signal and b) that there is significant room for improvement in synthesizing long-context information.

Rohan Anil, Sandra Gadanho, Da Huang et al. · deepmind

For industrial-scale advertising systems, prediction of ad click-through rate (CTR) is a central problem. Ad clicks constitute a significant class of user engagements and are often used as the primary signal for the usefulness of ads to users. Additionally, in cost-per-click advertising systems where advertisers are charged per click, click rate expectations feed directly into value estimation. Accordingly, CTR model development is a significant investment for most Internet advertising companies. Engineering for such problems requires many machine learning (ML) techniques suited to online learning that go well beyond traditional accuracy improvements, especially concerning efficiency, reproducibility, calibration, credit attribution. We present a case study of practical techniques deployed in Google's search ads CTR model. This paper provides an industry case study highlighting important areas of current ML research and illustrating how impactful new ML methods are evaluated and made useful in a large-scale industrial setting.

Ankit Singh Rawat, Veeranjaneyulu Sadhanala, Afshin Rostamizadeh et al. · deepmind

A primary challenge in large language model (LLM) development is their onerous pre-training cost. Typically, such pre-training involves optimizing a self-supervised objective (such as next-token prediction) over a large corpus. This paper explores a promising paradigm to improve LLM pre-training efficiency and quality by suitably leveraging a small language model (SLM). In particular, this paradigm relies on an SLM to both (1) provide soft labels as additional training supervision, and (2) select a small subset of valuable ("informative" and "hard") training examples. Put together, this enables an effective transfer of the SLM's predictive distribution to the LLM, while prioritizing specific regions of the training data distribution. Empirically, this leads to reduced LLM training time compared to standard training, while improving the overall quality. Theoretically, we develop a statistical framework to systematically study the utility of SLMs in enabling efficient training of high-quality LLMs. In particular, our framework characterizes how the SLM's seemingly low-quality supervision can enhance the training of a much more capable LLM. Furthermore, it also highlights the need for an adaptive utilization of such supervision, by striking a balance between the bias and variance introduced by the SLM-provided soft labels. We corroborate our theoretical framework by improving the pre-training of an LLM with 2.8B parameters by utilizing a smaller LM with 1.5B parameters on the Pile dataset.

Kiran Vodrahalli, Rakesh Shivanna, Maheswaran Sathiamoorthy et al.

We propose a novel low-rank initialization framework for training low-rank deep neural networks -- networks where the weight parameters are re-parameterized by products of two low-rank matrices. The most successful prior existing approach, spectral initialization, draws a sample from the initialization distribution for the full-rank setting and then optimally approximates the full-rank initialization parameters in the Frobenius norm with a pair of low-rank initialization matrices via singular value decomposition. Our method is inspired by the insight that approximating the function corresponding to each layer is more important than approximating the parameter values. We provably demonstrate that there is a significant gap between these two approaches for ReLU networks, particularly as the desired rank of the approximating weights decreases, or as the dimension of the inputs to the layer increases (the latter point holds when the network width is super-linear in dimension). Along the way, we provide the first provably efficient algorithm for solving the ReLU low-rank approximation problem for fixed parameter rank $r$ -- previously, it was unknown that the problem was computationally tractable to solve even for rank $1$. We also provide a practical algorithm to solve this problem which is no more expensive than the existing spectral initialization approach, and validate our theory by training ResNet and EfficientNet models (He et al., 2016; Tan & Le, 2019) on ImageNet (Russakovsky et al., 2015).

Jiaxi Tang, Rakesh Shivanna, Zhe Zhao et al.

Knowledge Distillation (KD) is a model-agnostic technique to improve model quality while having a fixed capacity budget. It is a commonly used technique for model compression, where a larger capacity teacher model with better quality is used to train a more compact student model with better inference efficiency. Through distillation, one hopes to benefit from student's compactness, without sacrificing too much on model quality. Despite the large success of knowledge distillation, better understanding of how it benefits student model's training dynamics remains under-explored. In this paper, we categorize teacher's knowledge into three hierarchical levels and study its effects on knowledge distillation: (1) knowledge of the `universe', where KD brings a regularization effect through label smoothing; (2) domain knowledge, where teacher injects class relationships prior to student's logit layer geometry; and (3) instance specific knowledge, where teacher rescales student model's per-instance gradients based on its measurement on the event difficulty. Using systematic analyses and extensive empirical studies on both synthetic and real-world datasets, we confirm that the aforementioned three factors play a major role in knowledge distillation. Furthermore, based on our findings, we diagnose some of the failure cases of applying KD from recent studies.

Aadirupa Saha, Rakesh Shivanna, Chiranjib Bhattacharyya

We consider the problem of optimal recovery of true ranking of $n$ items from a randomly chosen subset of their pairwise preferences. It is well known that without any further assumption, one requires a sample size of $Ω(n^2)$ for the purpose. We analyze the problem with an additional structure of relational graph $G([n],E)$ over the $n$ items added with an assumption of \emph{locality}: Neighboring items are similar in their rankings. Noting the preferential nature of the data, we choose to embed not the graph, but, its \emph{strong product} to capture the pairwise node relationships. Furthermore, unlike existing literature that uses Laplacian embedding for graph based learning problems, we use a richer class of graph embeddings---\emph{orthonormal representations}---that includes (normalized) Laplacian as its special case. Our proposed algorithm, {\it Pref-Rank}, predicts the underlying ranking using an SVM based approach over the chosen embedding of the product graph, and is the first to provide \emph{statistical consistency} on two ranking losses: \emph{Kendall's tau} and \emph{Spearman's footrule}, with a required sample complexity of $O(n^2 χ(\bar{G}))^{\frac{2}{3}}$ pairs, $χ(\bar{G})$ being the \emph{chromatic number} of the complement graph $\bar{G}$. Clearly, our sample complexity is smaller for dense graphs, with $χ(\bar G)$ characterizing the degree of node connectivity, which is also intuitive due to the locality assumption e.g. $O(n^\frac{4}{3})$ for union of $k$-cliques, or $O(n^\frac{5}{3})$ for random and power law graphs etc.---a quantity much smaller than the fundamental limit of $Ω(n^2)$ for large $n$. This, for the first time, relates ranking complexity to structural properties of the graph. We also report experimental evaluations on different synthetic and real datasets, where our algorithm is shown to outperform the state-of-the-art methods.