Mostafa Dehghani
I'm a Research Scientist at Google Brain, where I work on machine learning, in particular, deep learning. My areas of interest include self-supervised learning, generative models, training giant models, and sequence modeling.
Before Google, I was doing a PhD at the University of Amsterdam. My PhD research was focused on improving the process of learning with imperfect supervision. I explored ideas around using injecting inductive biases into algorithms, incorporating prior knowledge, and meta-learning the properties of the data using the data itself, in order to help learning algorithms to better learn from noisy or/and limited data.
You can know more about me here: mostafadehghani.com.
Research Areas
Authored Publications
Sort By
PaLI-X: On Scaling up a Multilingual Vision and Language Model
Josip Djolonga
Piotr Padlewski
Basil Mustafa
Carlos Riquelme
Sebastian Goodman
Yi Tay
Siamak Shakeri
Daniel Salz
Michael Tschannen
Hexiang (Frank) Hu
Mandar Joshi
Filip Pavetić
Gang Li
Lucas Beyer
Anurag Arnab
Yuanzhong Xu
Keran Rong
Alexander Kolesnikov
Xiaohua Zhai
Neil Houlsby
Computer Vision and Pattern Recognition Conference (CVPR) (2024)
Preview abstract
We explore the boundaries of scaling up a multilingual vision and language model, both in terms of size of the components and the breadth of its training task mixture. Our model achieves new levels of performance on a wide-range of varied and complex tasks, including multiple image-based captioning and question-answering tasks, image-based document understanding and few-shot (in-context) learning, as well as object detection, video question answering, and video captioning. Our model advances the state-of-the-art on most vision-and-language benchmarks considered (20+ of them). Finally, we observe emerging capabilities, such as complex counting and multilingual object detection, tasks that are not explicitly in the training mix.
View details
UL2: Unifying Language Learning Paradigms
Yi Tay
Xavier Garcia
Jason Wei
Hyung Won Chung
Steven Zheng
Neil Houlsby
ICLR (2023)
Preview abstract
Existing pre-trained models are generally geared towards a particular class of
problems. To date, there seems to be still no consensus on what the right architecture and pre-training setup should be. This paper presents a unified framework for
pre-training models that are universally effective across datasets and setups. We
begin by disentangling architectural archetypes with pre-training objectives – two
concepts that are commonly conflated. Next, we present a generalized and unified perspective for self-supervision in NLP and show how different pre-training
objectives can be cast as one another and how interpolating between different
objectives can be effective. We then propose Mixture-of-Denoisers (MoD), a pretraining objective that combines diverse pre-training paradigms together. We furthermore introduce a notion of mode switching, wherein downstream fine-tuning
is associated with specific pre-training schemes. We conduct extensive ablative
experiments to compare multiple pre-training objectives and find that our method
pushes the Pareto-frontier by outperforming T5 and/or GPT-like models across
multiple diverse setups. Finally, by scaling our model up to 20B parameters, we
achieve SOTA performance on 50 well-established supervised NLP tasks ranging from language generation (with automated and human evaluation), language
understanding, text classification, question answering, commonsense reasoning,
long text reasoning, structured knowledge grounding and information retrieval.
Our model also achieve strong results at in-context learning, outperforming 175B
GPT-3 on zero-shot SuperGLUE and tripling the performance of T5-XXL on oneshot summarization. Finally, we show that UL2 20B works well with chain-ofthought prompting and reasoning tasks, making it an appealing choice for research
into reasoning at a small to medium scale of 20B parameters. We publicly release
Flax-based T5X model checkpoints for the 20B model.
View details
DSI++: Updating Transformer Memory with New Documents
Yi Tay
Jinfeng Rao
Emma Strubell
Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing
Preview abstract
Differentiable Search Indices (DSIs) encode a corpus of documents in model parameters and use the same model to answer user queries directly. Despite the strong performance of DSI models, deploying them in situations where the corpus changes over time is computationally expensive because reindexing the corpus requires re-training the model. In this work, we introduce DSI++, a continual learning challenge for DSI to incrementally index new documents while being able to answer queries related to both previously and newly indexed documents.
Across different model scales and document identifier representations, we show that continual indexing of new documents leads to considerable forgetting of previously indexed documents. We also hypothesize and verify that the model experiences forgetting events during training, leading to unstable learning. To mitigate these issues, we investigate two approaches. The first focuses on modifying the training dynamics. Flatter minima implicitly alleviate forgetting, so we optimize for flatter loss basins and show that the model stably memorizes more documents (+12%). Next, we introduce a generative memory to sample pseudo-queries for documents and supplement them during continual indexing to prevent forgetting for the retrieval task. Extensive experiments on novel continual indexing benchmarks based on Natural Questions (NQ) and MS MARCO demonstrate that our proposed solution mitigates forgetting significantly. Concretely, it improves the average Hits@10 by +21.1% over competitive baselines for NQ and requires 6 times fewer model updates compared to re-training the DSI model for incrementally indexing five corpora in a sequence.
View details
Dual PatchNorm
Neil Houlsby
Transactions on Machine Learning Research (2023) (to appear)
Preview abstract
We discover that just placing two LayerNorms: before and after the patch embedding layer leads to improvements over well-tuned ViT models. In particular, this outperforms exhaustive search for alternative LayerNorm placement strategies in the transformer block itself.
View details
How (not) to ensemble LVLMs for VQA
Lisa Alazraki
Lluis Castrejon
Fantine Huot
"I Can't Believe It's Not Better: Failure Modes in the Age of Foundation Models" at NeurIPS 2023 Workshops
Preview abstract
This paper studies ensembling in the era of Large Vision-Language Models (LVLMs). Ensembling is a classical method to combine different models to get increased performance. In the recent work on Encyclopedic-VQA the authors examine a wide variety of models to solve their task: from vanilla LVLMs, to models including the caption as extra context, to models augmented with Lens-based retrieval of Wikipedia pages. Intuitively these models are highly complementary which should make them ideal for ensembling. Indeed, an oracle experiment shows potential gains from 48.8% accuracy (the best single model) all the way up to 67% (best possible ensemble). So it is a trivial exercise to create an ensemble with substantial real gains. Or is it?
View details
Scaling Vision Transformers to 22 Billion Parameters
Josip Djolonga
Basil Mustafa
Piotr Padlewski
Justin Gilmer
Mathilde Caron
Rodolphe Jenatton
Lucas Beyer
Michael Tschannen
Anurag Arnab
Carlos Riquelme
Gamaleldin Elsayed
Fisher Yu
Avital Oliver
Fantine Huot
Mark Collier
Vighnesh Birodkar
Yi Tay
Alexander Kolesnikov
Filip Pavetić
Thomas Kipf
Xiaohua Zhai
Neil Houlsby
Arxiv (2023)
Preview abstract
The scaling of Transformers has driven breakthrough capabilities for language models.
At present, the largest large language models (LLMs) contain upwards of 100B parameters.
Vision Transformers (ViT) have introduced the same architecture to image and video modeling, but these have not yet been successfully scaled to nearly the same degree; the largest dense ViT contains 4B parameters. We present a recipe for highly efficient training of a 22B-parameter ViT and perform a wide variety of experiments on the resulting model. When evaluated on downstream tasks (often with a lightweight linear model on frozen features) ViT22B demonstrates increasing performance with scale. We further observe other interesting benefits of scale, including an improved tradeoff between bias and performance, an improved alignment to human visual perception in terms of shape/texture bias, and improved robustness. ViT22B demonstrates the potential for "LLM-like'' scaling in vision, and provides key steps towards getting there.
View details
Confident Adaptive Language Modeling
Adam Fisch
Yi Tay
NeurIPS 2022
Preview abstract
Recent advances in Transformer-based large language models (LLMs) achieved significant performance improvements across many tasks.
These gains come with a drastic increase in the models' size, leading to slow and costly use at inference time.
In practice, however, the series of generations made by LLMs is composed of varying levels of difficulty. While certain predictions truly benefit from the models' full capacity, other continuations are more trivial and can be solved with reduced compute.
In this work, we introduce Confident Adaptive Language Modeling (CALM), a method for dynamically allocating different amounts of compute per example and per generation timestep.
Early exit decoding involves several challenges that we address here, such as: (1) what confidence measure to use; (2) connecting sequence-level constraints to local per-token exit decisions; and (3) attending back to missing hidden representations due to early exits in previous tokens.
Through theoretical analysis and empirical experiments on three diverse generation tasks, we demonstrate the efficacy of our method in reliably reducing compute while maintaining high performance.
View details
Preview abstract
Recent developments in large-scale machine learning have created a tempting picture suggesting that by scaling up data, model size and training time properly, one can obtain a model that can be used successfully in few-shot settings in all downstream tasks. In this work, we investigate this premise empirically and provide a strong case against it. In particular, we consider image recognition task with large scale models (Vision Transformers) trained on the largest scale of available data (JFT). We show that as we improve the performance of upstream task either by scaling up or hyper-parameter and architectural choices, the performance of many downstream tasks eventually plateau. We showcase an even more extreme scenario where performance on upstream and downstream contradict each other, i.e., in order to have a better downstream performance, we need to hurt upstream accuracy. We delve deeper into understanding the reasons that give rise to these phenomena by designing interventions and investigating different components of the models which gives us crude yet useful insights into the mechanisms behind these observations.
View details
Preview abstract
Vision Transformer (ViT) is emerging as the state-of-the-art architecture for image recognition. While recent studies suggest that ViTs are more robust than their convolutional counterparts, our experiments find that ViTs trained on ImageNet are overly reliant on local textures and fail to make adequate use of shape information. ViTs thus have difficulties generalizing to out-of-distribution, real-world data. To address this deficiency, we present a simple and effective architecture modification to ViT's input layer by adding discrete tokens produced by a vector-quantized encoder. Different from the standard continuous pixel tokens, discrete tokens are invariant under small perturbations and contain less information individually, which promote ViTs to learn global information that is invariant. Experimental results demonstrate that adding discrete representation on four architecture variants strengthens ViT robustness by up to 12% across seven ImageNet robustness benchmarks while maintaining the performance on ImageNet.
View details
Transformer Memory as a Differentiable Search Index
Yi Tay
Jianmo Ni
Harsh Mehta
Zhe Zhao
NeurIPS 2022
Preview abstract
In this paper, we demonstrate that information retrieval can be accomplished with a single Transformer, in which all information about the corpus is encoded in the parameters of the model. To this end, we introduce the Differentiable Search Index (DSI), a new paradigm that learns a text-to-text model that maps string queries directly to relevant docids; in other words, a DSI model answers queries directly using only its parameters, dramatically simplifying the whole retrieval process. We study variations in how documents and their identifiers are represented, variations in training procedures, and the interplay between models and corpus sizes. Experiments demonstrate that given appropriate design choices, DSI significantly outperforms strong baselines such as dual encoder models. Moreover, DSI demonstrates strong generalization capabilities, outperforming a BM25 baseline in a zero-shot setup.
View details