Andrea Gesmundo

Andrea Gesmundo, Jeff Dean

Multitask learning assumes that models capable of learning from multiple tasks can achieve better quality and efficiency via knowledge transfer, a key feature of human learning. Though, state of the art ML models rely on high customization for each task and leverage size and data scale rather than scaling the number of tasks. Also, continual learning, that adds the temporal aspect to multitask, is often focused to the study of common pitfalls such as catastrophic forgetting instead of being studied at a large scale as a critical component to build the next generation artificial intelligence.We propose an evolutionary method capable of generating large scale multitask models that support the dynamic addition of new tasks. The generated multitask models are sparsely activated and integrates a task-based routing that guarantees bounded compute cost and fewer added parameters per task as the model expands.The proposed method relies on a knowledge compartmentalization technique to achieve immunity against catastrophic forgetting and other common pitfalls such as gradient interference and negative transfer. We demonstrate empirically that the proposed method can jointly solve and achieve competitive results on 69public image classification tasks, for example improving the state of the art on a competitive benchmark such as cifar10 by achieving a 15% relative error reduction compared to the best model trained on public data.

Andrea Gesmundo, Jeff Dean

Most uses of machine learning today involve training a model from scratch for a particular task, or sometimes starting with a model pretrained on a related task and then fine-tuning on a downstream task. Both approaches offer limited knowledge transfer between different tasks, time-consuming human-driven customization to individual tasks and high computational costs especially when starting from randomly initialized models. We propose a method that uses the layers of a pretrained deep neural network as building blocks to construct an ML system that can jointly solve an arbitrary number of tasks. The resulting system can leverage cross tasks knowledge transfer, while being immune from common drawbacks of multitask approaches such as catastrophic forgetting, gradients interference and negative transfer. We define an evolutionary approach designed to jointly select the prior knowledge relevant for each task, choose the subset of the model parameters to train and dynamically auto-tune its hyperparameters. Furthermore, a novel scale control method is employed to achieve quality/size trade-offs that outperform common fine-tuning techniques. Compared with standard fine-tuning on a benchmark of 10 diverse image classification tasks, the proposed model improves the average accuracy by 2.39% while using 47% less parameters per task.

Andrea Gesmundo

The traditional Machine Learning (ML) methodology requires to fragment the development and experimental process into disconnected iterations whose feedback is used to guide design or tuning choices. This methodology has multiple efficiency and scalability disadvantages, such as leading to spend significant resources into the creation of multiple trial models that do not contribute to the final solution.The presented work is based on the intuition that defining ML models as modular and extensible artefacts allows to introduce a novel ML development methodology enabling the integration of multiple design and evaluation iterations into the continuous enrichment of a single unbounded intelligent system. We define a novel method for the generation of dynamic multitask ML models as a sequence of extensions and generalizations. We first analyze the capabilities of the proposed method by using the standard ML empirical evaluation methodology. Finally, we propose a novel continuous development methodology that allows to dynamically extend a pre-existing multitask large-scale ML system while analyzing the properties of the proposed method extensions. This results in the generation of an ML model capable of jointly solving 124 image classification tasks achieving state of the art quality with improved size and compute cost.

Andrea Gesmundo, Kaitlin Maile

Training state-of-the-art neural networks requires a high cost in terms of compute and time. Model scale is recognized to be a critical factor to achieve and improve the state-of-the-art. Increasing the scale of a neural network normally requires restarting from scratch by randomly initializing all the parameters of the model, as this implies a change of architecture's parameters that does not allow for a straightforward transfer of knowledge from smaller size models. In this work, we propose six composable transformations to incrementally increase the size of transformer-based neural networks while preserving functionality, allowing to expand the capacity of the model as needed. We provide proof of exact function preservation under minimal initialization constraints for each transformation. The proposed methods may enable efficient training pipelines for larger and more powerful models by progressively expanding the architecture throughout training.

Andrea Gesmundo

The traditional ML development methodology does not enable a large number of contributors, each with distinct objectives, to work collectively on the creation and extension of a shared intelligent system. Enabling such a collaborative methodology can accelerate the rate of innovation, increase ML technologies accessibility and enable the emergence of novel capabilities. We believe that this novel methodology for ML development can be demonstrated through a modularized representation of ML models and the definition of novel abstractions allowing to implement and execute diverse methods for the asynchronous use and extension of modular intelligent systems. We present a multiagent framework for the collaborative and asynchronous extension of dynamic large-scale multitask systems.

Andrea Gesmundo

A standard ML model is commonly generated by a single method that specifies aspects such as architecture, initialization, training data and hyperparameters configuration. The presented work introduces a novel methodology allowing to define multiple methods as distinct agents. Agents can collaborate and compete to generate and improve ML models for a given tasks. The proposed methodology is demonstrated with the generation and extension of a dynamic modular multitask ML system solving more than one hundred image classification tasks. Diverse agents can compete to produce the best performing model for a task by reusing the modules introduced to the system by competing agents. The presented work focuses on the study of agents capable of: 1) reusing the modules generated by concurrent agents, 2) activating in parallel multiple modules in a frozen state by connecting them with trainable modules, 3) condition the activation mixture on each data sample by using a trainable router module. We demonstrate that this simple per-sample parallel routing method can boost the quality of the combined solutions by training a fraction of the activated parameters.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Adam Roberts, Hyung Won Chung, Anselm Levskaya et al.

Recent neural network-based language models have benefited greatly from scaling up the size of training datasets and the number of parameters in the models themselves. Scaling can be complicated due to various factors including the need to distribute computation on supercomputer clusters (e.g., TPUs), prevent bottlenecks when infeeding data, and ensure reproducible results. In this work, we present two software libraries that ease these issues: $\texttt{t5x}$ simplifies the process of building and training large language models at scale while maintaining ease of use, and $\texttt{seqio}$ provides a task-based API for simple creation of fast and reproducible training data and evaluation pipelines. These open-source libraries have been used to train models with hundreds of billions of parameters on datasets with multiple terabytes of training data. Along with the libraries, we release configurations and instructions for T5-like encoder-decoder models as well as GPT-like decoder-only architectures. $\texttt{t5x}$ and $\texttt{seqio}$ are open source and available at https://github.com/google-research/t5x and https://github.com/google/seqio, respectively.

Mark Collier, Efi Kokiopoulou, Andrea Gesmundo et al.

The core challenge with continual learning is catastrophic forgetting, the phenomenon that when neural networks are trained on a sequence of tasks they rapidly forget previously learned tasks. It has been observed that catastrophic forgetting is most severe when tasks are dissimilar to each other. We propose the use of sparse routing networks for continual learning. For each input, these network architectures activate a different path through a network of experts. Routing networks have been shown to learn to route similar tasks to overlapping sets of experts and dissimilar tasks to disjoint sets of experts. In the continual learning context this behaviour is desirable as it minimizes interference between dissimilar tasks while allowing positive transfer between related tasks. In practice, we find it is necessary to develop a new training method for routing networks, which we call co-training which avoids poorly initialized experts when new tasks are presented. When combined with a small episodic memory replay buffer, sparse routing networks with co-training outperform densely connected networks on the MNIST-Permutations and MNIST-Rotations benchmarks.

Alina Dubatovka, Efi Kokiopoulou, Luciano Sbaiz et al.

Neural architecture search has recently attracted lots of research efforts as it promises to automate the manual design of neural networks. However, it requires a large amount of computing resources and in order to alleviate this, a performance prediction network has been recently proposed that enables efficient architecture search by forecasting the performance of candidate architectures, instead of relying on actual model training. The performance predictor is task-aware taking as input not only the candidate architecture but also task meta-features and it has been designed to collectively learn from several tasks. In this work, we introduce a pairwise ranking loss for training a network able to rank candidate architectures for a new unseen task conditioning on its task meta-features. We present experimental results, showing that the ranking network is more effective in architecture search than the previously proposed performance predictor.

Krzysztof Maziarz, Efi Kokiopoulou, Andrea Gesmundo et al.

This paper proposes a novel learning method for multi-task applications. Multi-task neural networks can learn to transfer knowledge across different tasks by using parameter sharing. However, sharing parameters between unrelated tasks can hurt performance. To address this issue, we propose a framework to learn fine-grained patterns of parameter sharing. Assuming that the network is composed of several components across layers, our framework uses learned binary variables to allocate components to tasks in order to encourage more parameter sharing between related tasks, and discourage parameter sharing otherwise. The binary allocation variables are learned jointly with the model parameters by standard back-propagation thanks to the Gumbel-Softmax reparametrization method. When applied to the Omniglot benchmark, the proposed method achieves a 17% relative reduction of the error rate compared to state-of-the-art.

Iulia M. Comsa, Krzysztof Potempa, Luca Versari et al.

The timing of individual neuronal spikes is essential for biological brains to make fast responses to sensory stimuli. However, conventional artificial neural networks lack the intrinsic temporal coding ability present in biological networks. We propose a spiking neural network model that encodes information in the relative timing of individual neuron spikes. In classification tasks, the output of the network is indicated by the first neuron to spike in the output layer. This temporal coding scheme allows the supervised training of the network with backpropagation, using locally exact derivatives of the postsynaptic spike times with respect to presynaptic spike times. The network operates using a biologically-plausible alpha synaptic transfer function. Additionally, we use trainable synchronisation pulses that provide bias, add flexibility during training and exploit the decay part of the alpha function. We show that such networks can be trained successfully on noisy Boolean logic tasks and on the MNIST dataset encoded in time. The results show that the spiking neural network outperforms comparable spiking models on MNIST and achieves similar quality to fully connected conventional networks with the same architecture. We also find that the spiking network spontaneously discovers two operating regimes, mirroring the accuracy-speed trade-off observed in human decision-making: a slow regime, where a decision is taken after all hidden neurons have spiked and the accuracy is very high, and a fast regime, where a decision is taken very fast but the accuracy is lower. These results demonstrate the computational power of spiking networks with biological characteristics that encode information in the timing of individual neurons. By studying temporal coding in spiking networks, we aim to create building blocks towards energy-efficient and more complex biologically-inspired neural architectures.

Zalán Borsos, Andrey Khorlin, Andrea Gesmundo

Recent advances in Neural Architecture Search (NAS) have produced state-of-the-art architectures on several tasks. NAS shifts the efforts of human experts from developing novel architectures directly to designing architecture search spaces and methods to explore them efficiently. The search space definition captures prior knowledge about the properties of the architectures and it is crucial for the complexity and the performance of the search algorithm. However, different search space definitions require restarting the learning process from scratch. We propose a novel agent based on the Transformer that supports joint training and efficient transfer of prior knowledge between multiple search spaces and tasks.

Efi Kokiopoulou, Anja Hauth, Luciano Sbaiz et al.

Neural architecture search has been shown to hold great promise towards the automation of deep learning. However in spite of its potential, neural architecture search remains quite costly. To this point, we propose a novel gradient-based framework for efficient architecture search by sharing information across several tasks. We start by training many model architectures on several related (training) tasks. When a new unseen task is presented, the framework performs architecture inference in order to quickly identify a good candidate architecture, before any model is trained on the new task. At the core of our framework lies a deep value network that can predict the performance of input architectures on a task by utilizing task meta-features and the previous model training experiments performed on related tasks. We adopt a continuous parametrization of the model architecture which allows for efficient gradient-based optimization. Given a new task, an effective architecture is quickly identified by maximizing the estimated performance with respect to the model architecture parameters with simple gradient ascent. It is key to point out that our goal is to achieve reasonable performance at the lowest cost. We provide experimental results showing the effectiveness of the framework despite its high computational efficiency.

Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski et al.

Fine-tuning large pre-trained models is an effective transfer mechanism in NLP. However, in the presence of many downstream tasks, fine-tuning is parameter inefficient: an entire new model is required for every task. As an alternative, we propose transfer with adapter modules. Adapter modules yield a compact and extensible model; they add only a few trainable parameters per task, and new tasks can be added without revisiting previous ones. The parameters of the original network remain fixed, yielding a high degree of parameter sharing. To demonstrate adapter's effectiveness, we transfer the recently proposed BERT Transformer model to 26 diverse text classification tasks, including the GLUE benchmark. Adapters attain near state-of-the-art performance, whilst adding only a few parameters per task. On GLUE, we attain within 0.4% of the performance of full fine-tuning, adding only 3.6% parameters per task. By contrast, fine-tuning trains 100% of the parameters per task.

Stanisław Jastrzębski, Quentin de Laroussilhe, Mingxing Tan et al.

Neural Architecture Search (NAS) enabled the discovery of state-of-the-art architectures in many domains. However, the success of NAS depends on the definition of the search space. Current search spaces are defined as a static sequence of decisions and a set of available actions for each decision. Each possible sequence of actions defines an architecture. We propose a more expressive class of search space: directed graphs. In our formalism, each decision is a vertex and each action is an edge. This allows us to model iterative and branching architecture design decisions. We demonstrate in simulation, and on image classification experiments, basic iterative and branching search structures, and show that the graph representation improves sample efficiency.

Krzysztof Maziarz, Mingxing Tan, Andrey Khorlin et al.

Neural Architecture Search has shown potential to automate the design of neural networks. Deep Reinforcement Learning based agents can learn complex architectural patterns, as well as explore a vast and compositional search space. On the other hand, evolutionary algorithms offer higher sample efficiency, which is critical for such a resource intensive application. In order to capture the best of both worlds, we propose a class of Evolutionary-Neural hybrid agents (Evo-NAS). We show that the Evo-NAS agent outperforms both neural and evolutionary agents when applied to architecture search for a suite of text and image classification benchmarks. On a high-complexity architecture search space for image classification, the Evo-NAS agent surpasses the accuracy achieved by commonly used agents with only 1/3 of the search cost.

Catherine Wong, Neil Houlsby, Yifeng Lu et al.

We reduce the computational cost of Neural AutoML with transfer learning. AutoML relieves human effort by automating the design of ML algorithms. Neural AutoML has become popular for the design of deep learning architectures, however, this method has a high computation cost. To address this we propose Transfer Neural AutoML that uses knowledge from prior tasks to speed up network design. We extend RL-based architecture search methods to support parallel training on multiple tasks and then transfer the search strategy to new tasks. On language and image classification tasks, Transfer Neural AutoML reduces convergence time over single-task training by over an order of magnitude on many tasks.

Christian Buck, Jannis Bulian, Massimiliano Ciaramita et al.

We analyze the language learned by an agent trained with reinforcement learning as a component of the ActiveQA system [Buck et al., 2017]. In ActiveQA, question answering is framed as a reinforcement learning task in which an agent sits between the user and a black box question-answering system. The agent learns to reformulate the user's questions to elicit the optimal answers. It probes the system with many versions of a question that are generated via a sequence-to-sequence question reformulation model, then aggregates the returned evidence to find the best answer. This process is an instance of \emph{machine-machine} communication. The question reformulation model must adapt its language to increase the quality of the answers returned, matching the language of the question answering system. We find that the agent does not learn transformations that align with semantic intuitions but discovers through learning classical information retrieval techniques such as tf-idf re-weighting and stemming.

Catherine Wong, Andrea Gesmundo

Deep learning models require extensive architecture design exploration and hyperparameter optimization to perform well on a given task. The exploration of the model design space is often made by a human expert, and optimized using a combination of grid search and search heuristics over a large space of possible choices. Neural Architecture Search (NAS) is a Reinforcement Learning approach that has been proposed to automate architecture design. NAS has been successfully applied to generate Neural Networks that rival the best human-designed architectures. However, NAS requires sampling, constructing, and training hundreds to thousands of models to achieve well-performing architectures. This procedure needs to be executed from scratch for each new task. The application of NAS to a wide set of tasks currently lacks a way to transfer generalizable knowledge across tasks. In this paper, we present the Multitask Neural Model Search (MNMS) controller. Our goal is to learn a generalizable framework that can condition model construction on successful model searches for previously seen tasks, thus significantly speeding up the search for new tasks. We demonstrate that MNMS can conduct an automated architecture search for multiple tasks simultaneously while still learning well-performing, specialized models for each task. We then show that pre-trained MNMS controllers can transfer learning to new tasks. By leveraging knowledge from previous searches, we find that pre-trained MNMS models start from a better location in the search space and reduce search time on unseen tasks, while still discovering models that outperform published human-designed models.

Christian Buck, Jannis Bulian, Massimiliano Ciaramita et al.

We frame Question Answering (QA) as a Reinforcement Learning task, an approach that we call Active Question Answering. We propose an agent that sits between the user and a black box QA system and learns to reformulate questions to elicit the best possible answers. The agent probes the system with, potentially many, natural language reformulations of an initial question and aggregates the returned evidence to yield the best answer. The reformulation system is trained end-to-end to maximize answer quality using policy gradient. We evaluate on SearchQA, a dataset of complex questions extracted from Jeopardy!. The agent outperforms a state-of-the-art base model, playing the role of the environment, and other benchmarks. We also analyze the language that the agent has learned while interacting with the question answering system. We find that successful question reformulations look quite different from natural language paraphrases. The agent is able to discover non-trivial reformulation strategies that resemble classic information retrieval techniques such as term re-weighting (tf-idf) and stemming.