Category: FAANG

Learning meaningful representations of complex objects that can be seen through multiple (k≥3kgeq 3k≥3) views or modalities is a core task in machine learning. Existing methods use losses originally intended for paired views, and extend them to kkk views, either by instantiating 12k(k−1)tfrac12k(k-1)21​k(k−1) loss-pairs, or by using reduced embeddings, following a one vs. average-of-resttextit{one vs. average-of-rest}one vs. average-of-rest strategy. …

Read more “Contrasting Multiple Representations with the Multi-Marginal Matching Gap”

Beyond minimizing a single training loss, many deep learning estimation pipelines rely on an auxiliary objective to quantify and encourage desirable properties of the model (e.g. performance on another dataset, robustness, agreement with a prior). Although the simplest approach to incorporating an auxiliary loss is to sum it with the training loss as a regularizer, …

Read more “Careful With That Scalpel: Improving Gradient Surgery With an EMA”

It’s no exaggeration to say that chatbots are a mainstream technology. A whopping 88% of web users chatted with bots in the previous year. What’s more, 7 out of 10 find the experience positive. However, despite these promising statistics, there’s still a significant portion of users who have less-than-stellar experiences with conversational systems. The Dark …

Read more “Is Your Chatbot Sabotaging Your Business? The Tell-Tale Signs and a Rescue Plan”

Engineers have developed a new soft, flexible device that makes robots move by expanding and contracting — just like a human muscle. To demonstrate their new device, called an actuator, the researchers used it to create a cylindrical, worm-like soft robot and an artificial bicep. In experiments, the cylindrical soft robot navigated the tight, hairpin …

Read more “A new twist on artificial ‘muscles’ for safer, softer robots”

We propose utilizing n-best reranking to enhance Sequence-Level Knowledge Distillation (Kim and Rush, 2016) where we extract pseudo-labels for student model’s training data from top n-best hypotheses and leverage a diverse set of models with different inductive biases, objective functions or architectures, including some publicly-available large language models, to pick the highest-quality hypotheses as labels. …

Read more “Accurate Knowledge Distillation via N-best Reranking”