Agents capable of carrying out general tasks on a computer can improve efficiency and productivity by automating repetitive tasks and assisting in complex problem-solving. Ideally, such agents should be able to solve new computer tasks presented to them through natural language commands. However, previous approaches to this problem require large amounts of expert demonstrations and task-specific reward functions, both of which are impractical for new tasks. In this work, we show that a pre-trained large language model (LLM) agent can execute computer tasks guided by natural language using a simple prompting scheme where the agent Recursively Criticizes and Improves its output (RCI). The RCI approach significantly outperforms existing LLM methods for automating computer tasks and surpasses supervised learning (SL) and reinforcement learning (RL) approaches on the MiniWoB++ benchmark. We compare multiple LLMs and find that RCI with the InstructGPT-3+RLHF LLM is state-of-the-art on MiniWoB++, using only a handful of demonstrations per task rather than tens of thousands, and without a task-specific reward function. Furthermore, we demonstrate RCI prompting's effectiveness in enhancing LLMs' reasoning abilities on a suite of natural language reasoning tasks, outperforming chain of thought (CoT) prompting with external feedback. We find that RCI combined with CoT performs better than either separately. Our code can be found here: https://github.com/posgnu/rci-agent.
In competitive two-agent environments, deep reinforcement learning (RL) methods based on the \emph{Double Oracle (DO)} algorithm, such as \emph{Policy Space Response Oracles (PSRO)} and \emph{Anytime PSRO (APSRO)}, iteratively add RL best response policies to a population. Eventually, an optimal mixture of these population policies will approximate a Nash equilibrium. However, these methods might need to add all deterministic policies before converging. In this work, we introduce \emph{Self-Play PSRO (SP-PSRO)}, a method that adds an approximately optimal stochastic policy to the population in each iteration. Instead of adding only deterministic best responses to the opponent's least exploitable population mixture, SP-PSRO also learns an approximately optimal stochastic policy and adds it to the population as well. As a result, SP-PSRO empirically tends to converge much faster than APSRO and in many games converges in just a few iterations.
Michael James Fenton, Alexander Shmakov, Hideki Okawa et al.
Reconstructing unstable heavy particles requires sophisticated techniques to sift through the large number of possible permutations for assignment of detector objects to the underlying partons. Anapproach based on a generalized attention mechanism, symmetry preserving attention networks (SPA-NET), has been previously applied to top quark pair decays at the Large Hadron Collider which produce only hadronic jets. Here we extend the SPA-NET architecture to consider multiple input object types, such as leptons, as well as global event features, such as the missing transverse momentum. Inaddition, we provide regression and classification outputs to supplement the parton assignment. We explore the performance of the extended capability of SPA-NET in the context of semi-leptonic decays of top quark pairs as well as top quark pairs produced in association with a Higgs boson. We find significant improvements in the power of three representative studies: a search for ttH, a measurement of the top quark mass, and a search for a heavy Z' decaying to top quark pairs. We present ablation studies to provide insight on what the network has learned in each case.
Robust reinforcement learning (RL) considers the problem of learning policies that perform well in the worst case among a set of possible environment parameter values. In real-world environments, choosing the set of possible values for robust RL can be a difficult task. When that set is specified too narrowly, the agent will be left vulnerable to reasonable parameter values unaccounted for. When specified too broadly, the agent will be too cautious. In this paper, we propose Feasible Adversarial Robust RL (FARR), a novel problem formulation and objective for automatically determining the set of environment parameter values over which to be robust. FARR implicitly defines the set of feasible parameter values as those on which an agent could achieve a benchmark reward given enough training resources. By formulating this problem as a two-player zero-sum game, optimizing the FARR objective jointly produces an adversarial distribution over parameter values with feasible support and a policy robust over this feasible parameter set. We demonstrate that approximate Nash equilibria for this objective can be found using a variation of the PSRO algorithm. Furthermore, we show that an optimal agent trained with FARR is more robust to feasible adversarial parameter selection than with existing minimax, domain-randomization, and regret objectives in a parameterized gridworld and three MuJoCo control environments.
Kolby Nottingham, Yasaman Razeghi, Kyungmin Kim et al.
Large language models (LLMs) are being applied as actors for sequential decision making tasks in domains such as robotics and games, utilizing their general world knowledge and planning abilities. However, previous work does little to explore what environment state information is provided to LLM actors via language. Exhaustively describing high-dimensional states can impair performance and raise inference costs for LLM actors. Previous LLM actors avoid the issue by relying on hand-engineered, task-specific protocols to determine which features to communicate about a state and which to leave out. In this work, we propose Brief Language INputs for DEcision-making Responses (BLINDER), a method for automatically selecting concise state descriptions by learning a value function for task-conditioned state descriptions. We evaluate BLINDER on the challenging video game NetHack and a robotic manipulation task. Our method improves task success rate, reduces input size and compute costs, and generalizes between LLM actors.
Calorimeter shower simulations are often the bottleneck in simulation time for particle physics detectors. A lot of effort is currently spent on optimizing generative architectures for specific detector geometries, which generalize poorly. We develop a geometry-aware autoregressive model on a range of calorimeter geometries such that the model learns to adapt its energy deposition depending on the size and position of the cells. This is a key proof-of-concept step towards building a model that can generalize to new unseen calorimeter geometries with little to no additional training. Such a model can replace the hundreds of generative models used for calorimeter simulation in a Large Hadron Collider experiment. For the study of future detectors, such a model will dramatically reduce the large upfront investment usually needed to generate simulations.
Alexander Shmakov, Alejandro Yankelevich, Jianming Bian et al.
The complex events observed at the NOvA long-baseline neutrino oscillation experiment contain vital information for understanding the most elusive particles in the standard model. The NOvA detectors observe interactions of neutrinos from the NuMI beam at Fermilab. Associating the particles produced in these interaction events to their source particles, a process known as reconstruction, is critical for accurately measuring key parameters of the standard model. Events may contain several particles, each producing sparse high-dimensional spatial observations, and current methods are limited to evaluating individual particles. To accurately label these numerous, high-dimensional observations, we present a novel neural network architecture that combines the spatial learning enabled by convolutions with the contextual learning enabled by attention. This joint approach, TransformerCVN, simultaneously classifies each event and reconstructs every individual particle's identity. TransformerCVN classifies events with 90\% accuracy and improves the reconstruction of individual particles by 6\% over baseline methods which lack the integrated architecture of TransformerCVN. In addition, this architecture enables us to perform several interpretability studies which provide insights into the network's predictions and show that TransformerCVN discovers several fundamental principles that stem from the standard model.
Alexander Shmakov, Mohammadamin Tavakoli, Pierre Baldi et al.
A significant point-like component from the small scale (or discrete) structure in the H2 interstellar gas might be present in the Fermi-LAT data, but modeling this emission relies on observations of rare gas tracers only available in limited regions of the sky. Identifying this contribution is important to discriminate gamma-ray point sources from interstellar gas, and to better characterize extended gamma-ray sources. We design and train convolutional neural networks to predict this emission where observations of these rare tracers do not exist and discuss the impact of this component on the analysis of the Fermi-LAT data. In particular, we evaluate prospects to exploit this methodology in the characterization of the Fermi-LAT Galactic center excess through accurate modeling of point-like structures in the data to help distinguish between a point-like or smooth nature for the excess. We show that deep learning may be effectively employed to model the gamma-ray emission traced by these rare H2 proxies within statistical significance in data-rich regions, supporting prospects to employ these methods in yet unobserved regions.
Mohammadamin Tavakoli, Yin Ting T. Chiu, Alexander Shmakov et al.
Deep learning-based reaction predictors have undergone significant architectural evolution. However, their reliance on reactions from the US Patent Office results in a lack of interpretable predictions and limited generalization capability to other chemistry domains, such as radical and atmospheric chemistry. To address these challenges, we introduce a new reaction predictor system, RMechRP, that leverages contrastive learning in conjunction with mechanistic pathways, the most interpretable representation of chemical reactions. Specifically designed for radical reactions, RMechRP provides different levels of interpretation of chemical reactions. We develop and train multiple deep-learning models using RMechDB, a public database of radical reactions, to establish the first benchmark for predicting radical reactions. Our results demonstrate the effectiveness of RMechRP in providing accurate and interpretable predictions of radical reactions, and its potential for various applications in atmospheric chemistry.
Policy Space Response Oracles (PSRO) is a reinforcement learning (RL) algorithm for two-player zero-sum games that has been empirically shown to find approximate Nash equilibria in large games. Although PSRO is guaranteed to converge to an approximate Nash equilibrium and can handle continuous actions, it may take an exponential number of iterations as the number of information states (infostates) grows. We propose Extensive-Form Double Oracle (XDO), an extensive-form double oracle algorithm for two-player zero-sum games that is guaranteed to converge to an approximate Nash equilibrium linearly in the number of infostates. Unlike PSRO, which mixes best responses at the root of the game, XDO mixes best responses at every infostate. We also introduce Neural XDO (NXDO), where the best response is learned through deep RL. In tabular experiments on Leduc poker, we find that XDO achieves an approximate Nash equilibrium in a number of iterations an order of magnitude smaller than PSRO. Experiments on a modified Leduc poker game and Oshi-Zumo show that tabular XDO achieves a lower exploitability than CFR with the same amount of computation. We also find that NXDO outperforms PSRO and NFSP on a sequential multidimensional continuous-action game. NXDO is the first deep RL method that can find an approximate Nash equilibrium in high-dimensional continuous-action sequential games. Experiment code is available at https://github.com/indylab/nxdo.
Finding approximate Nash equilibria in zero-sum imperfect-information games is challenging when the number of information states is large. Policy Space Response Oracles (PSRO) is a deep reinforcement learning algorithm grounded in game theory that is guaranteed to converge to an approximate Nash equilibrium. However, PSRO requires training a reinforcement learning policy at each iteration, making it too slow for large games. We show through counterexamples and experiments that DCH and Rectified PSRO, two existing approaches to scaling up PSRO, fail to converge even in small games. We introduce Pipeline PSRO (P2SRO), the first scalable general method for finding approximate Nash equilibria in large zero-sum imperfect-information games. P2SRO is able to parallelize PSRO with convergence guarantees by maintaining a hierarchical pipeline of reinforcement learning workers, each training against the policies generated by lower levels in the hierarchy. We show that unlike existing methods, P2SRO converges to an approximate Nash equilibrium, and does so faster as the number of parallel workers increases, across a variety of imperfect information games. We also introduce an open-source environment for Barrage Stratego, a variant of Stratego with an approximate game tree complexity of $10^{50}$. P2SRO is able to achieve state-of-the-art performance on Barrage Stratego and beats all existing bots. Experiment code is available athttps://github.com/JBLanier/pipeline-psro.
Lars Hertel, Julian Collado, Peter Sadowski et al.
Sherpa is a hyperparameter optimization library for machine learning models. It is specifically designed for problems with computationally expensive, iterative function evaluations, such as the hyperparameter tuning of deep neural networks. With Sherpa, scientists can quickly optimize hyperparameters using a variety of powerful and interchangeable algorithms. Sherpa can be run on either a single machine or in parallel on a cluster. Finally, an interactive dashboard enables users to view the progress of models as they are trained, cancel trials, and explore which hyperparameter combinations are working best. Sherpa empowers machine learning practitioners by automating the more tedious aspects of model tuning. Its source code and documentation are available at https://github.com/sherpa-ai/sherpa.
Alexander Shmakov, Kevin Greif, Michael James Fenton et al.
The measurements performed by particle physics experiments must account for the imperfect response of the detectors used to observe the interactions. One approach, unfolding, statistically adjusts the experimental data for detector effects. Recently, generative machine learning models have shown promise for performing unbinned unfolding in a high number of dimensions. However, all current generative approaches are limited to unfolding a fixed set of observables, making them unable to perform full-event unfolding in the variable dimensional environment of collider data. A novel modification to the variational latent diffusion model (VLD) approach to generative unfolding is presented, which allows for unfolding of high- and variable-dimensional feature spaces. The performance of this method is evaluated in the context of semi-leptonic top quark pair production at the Large Hadron Collider.
Haoyang Li, Marko Stamenkovic, Alexander Shmakov et al.
The production of multiple Higgs bosons at the CERN LHC provides a direct way to measure the trilinear and quartic Higgs self-interaction strengths as well as potential access to beyond the standard model effects that can enhance production at large transverse momentum $p_{\mathrm{T}}$. The largest event fraction arises from the fully hadronic final state in which every Higgs boson decays to a bottom quark-antiquark pair ($b\bar{b}$). This introduces a combinatorial challenge known as the \emph{jet assignment problem}: assigning jets to sets representing Higgs boson candidates. Symmetry-preserving attention networks (SPA-Nets) have been been developed to address this challenge. However, the complexity of jet assignment increases when simultaneously considering both $H\rightarrow b\bar{b}$ reconstruction possibilities, i.e., two "resolved" small-radius jets each containing a shower initiated by a $b$-quark or one "boosted" large-radius jet containing a merged shower initiated by a $b\bar{b}$ pair. The latter improves the reconstruction efficiency at high $p_{\mathrm{T}}$. In this work, we introduce a generalization to the SPA-Net approach to simultaneously consider both boosted and resolved reconstruction possibilities and unambiguously interpret an event as "fully resolved'', "fully boosted", or in between. We report the performance of baseline methods, the original SPA-Net approach, and our generalized version on nonresonant $HH$ and $HHH$ production at the LHC. Considering both boosted and resolved topologies, our SPA-Net approach increases the Higgs boson reconstruction purity by 57--62\% and the efficiency by 23--38\% compared to the baseline method depending on the final state.
Recent advancements in Model-Based Reinforcement Learning (MBRL) have made it a powerful tool for visual control tasks. Despite improved data efficiency, it remains challenging to train MBRL agents with generalizable perception. Training in the presence of visual distractions is particularly difficult due to the high variation they introduce to representation learning. Building on DREAMER, a popular MBRL method, we propose a simple yet effective auxiliary task to facilitate representation learning in distracting environments. Under the assumption that task-relevant components of image observations are straightforward to identify with prior knowledge in a given task, we use a segmentation mask on image observations to only reconstruct task-relevant components. In doing so, we greatly reduce the complexity of representation learning by removing the need to encode task-irrelevant objects in the latent representation. Our method, Segmentation Dreamer (SD), can be used either with ground-truth masks easily accessible in simulation or by leveraging potentially imperfect segmentation foundation models. The latter is further improved by selectively applying the reconstruction loss to avoid providing misleading learning signals due to mask prediction errors. In modified DeepMind Control suite (DMC) and Meta-World tasks with added visual distractions, SD achieves significantly better sample efficiency and greater final performance than prior work. We find that SD is especially helpful in sparse reward tasks otherwise unsolvable by prior work, enabling the training of visually robust agents without the need for extensive reward engineering.
The generation of natural and high-quality speech from text is a challenging problem in the field of natural language processing. In addition to speech generation, speech editing is also a crucial task, which requires the seamless and unnoticeable integration of edited speech into synthesized speech. We propose a novel approach to speech editing by leveraging a pre-trained text-to-speech (TTS) model, such as FastSpeech 2, and incorporating a double attention block network on top of it to automatically merge the synthesized mel-spectrogram with the mel-spectrogram of the edited text. We refer to this model as AttentionStitch, as it harnesses attention to stitch audio samples together. We evaluate the proposed AttentionStitch model against state-of-the-art baselines on both single and multi-speaker datasets, namely LJSpeech and VCTK. We demonstrate its superior performance through an objective and a subjective evaluation test involving 15 human participants. AttentionStitch is capable of producing high-quality speech, even for words not seen during training, while operating automatically without the need for human intervention. Moreover, AttentionStitch is fast during both training and inference and is able to generate human-sounding edited speech.
In Inertial Confinement Fusion (ICF) process, roughly a 2mm spherical shell made of high density carbon is used as target for laser beams, which compress and heat it to energy levels needed for high fusion yield. These shells are polished meticulously to meet the standards for a fusion shot. However, the polishing of these shells involves multiple stages, with each stage taking several hours. To make sure that the polishing process is advancing in the right direction, we are able to measure the shell surface roughness. This measurement, however, is very labor-intensive, time-consuming, and requires a human operator. We propose to use machine learning models that can predict surface roughness based on the data collected from a vibration sensor that is connected to the polisher. Such models can generate surface roughness of the shells in real-time, allowing the operator to make any necessary changes to the polishing for optimal result.
Ryan J. Miller, Alexander E. Dashuta, Brayden Rudisill et al.
Accurately predicting chemical reactions is essential for driving innovation in synthetic chemistry, with broad applications in medicine, manufacturing, and agriculture. At the same time, reaction prediction is a complex problem which can be both time-consuming and resource-intensive for chemists to solve. Deep learning methods offer an appealing solution by enabling high-throughput reaction prediction. However, many existing models are trained on the US Patent Office dataset and treat reactions as overall transformations: mapping reactants directly to products with limited interpretability or mechanistic insight. To address this, we introduce PMechRP (Polar Mechanistic Reaction Predictor), a system that trains machine learning models on the PMechDB dataset, which represents reactions as polar elementary steps that capture electron flow and mechanistic detail. To further expand model coverage and improve generalization, we augment PMechDB with a diverse set of combinatorially generated reactions. We train and compare a range of machine learning models, including transformer-based, graph-based, and two-step siamese architectures. Our best-performing approach was a hybrid model, which combines a 5-ensemble of Chemformer models with a two-step Siamese framework to leverage the accuracy of transformer architectures, while filtering away "alchemical" products using the two-step network predictions. For evaluation, we use a test split of the PMechDB dataset and additionally curate a human benchmark dataset consisting of complete mechanistic pathways extracted from an organic chemistry textbook. Our hybrid model achieves a top-10 accuracy of 94.9% on the PMechDB test set and a target recovery rate of 84.9% on the pathway dataset.
We develop a general theory of synaptic neural balance and how it can emerge or be enforced in neural networks. For a given regularizer, a neuron is said to be in balance if the total cost of its input weights is equal to the total cost of its output weights. The basic example is provided by feedforward networks of ReLU units trained with $L_2$ regularizers, which exhibit balance after proper training. The theory explains this phenomenon and extends it in several directions. The first direction is the extension to bilinear and other activation functions. The second direction is the extension to more general regularizers, including all $L_p$ regularizers. The third direction is the extension to non-layered architectures, recurrent architectures, convolutional architectures, as well as architectures with mixed activation functions. Gradient descent on the error function alone does not converge in general to a balanced state, where every neuron is in balance, even when starting from a balanced state. However, gradient descent on the regularized error function ought to converge to a balanced state, and thus network balance can be used to assess learning progress. The theory is based on two local neuronal operations: scaling which is commutative, and balancing which is not commutative. Given any initial set of weights, when local balancing operations are applied to each neuron in a stochastic manner, global order always emerges through the convergence of the stochastic balancing algorithm to the same unique set of balanced weights. The reason for this is the existence of an underlying strictly convex optimization problem where the relevant variables are constrained to a linear, only architecture-dependent, manifold. Simulations show that balancing neurons prior to learning, or during learning in alternation with gradient descent steps, can improve learning speed and final performance.
Creating controlled methods to simulate neurodegeneration in artificial intelligence (AI) is crucial for applications that emulate brain function decline and cognitive disorders. We use IQ tests performed by Large Language Models (LLMs) and, more specifically, the LLaMA 2 to introduce the concept of ``neural erosion." This deliberate erosion involves ablating synapses or neurons, or adding Gaussian noise during or after training, resulting in a controlled progressive decline in the LLMs' performance. We are able to describe the neurodegeneration in the IQ tests and show that the LLM first loses its mathematical abilities and then its linguistic abilities, while further losing its ability to understand the questions. To the best of our knowledge, this is the first work that models neurodegeneration with text data, compared to other works that operate in the computer vision domain. Finally, we draw similarities between our study and cognitive decline clinical studies involving test subjects. We find that with the application of neurodegenerative methods, LLMs lose abstract thinking abilities, followed by mathematical degradation, and ultimately, a loss in linguistic ability, responding to prompts incoherently. These findings are in accordance with human studies.
Shahriar Hojjati Emmami, Ali Pilehvar Meibody, Lobat Tayebi et al.
The deliberate manipulation of ammonium persulfate, methylenebisacrylamide, dimethyleacrylamide, and polyethylene oxide concentrations resulted in the development of a hydrogel with an exceptional stretchability, capable of extending up to 260 times its original length. This study aims to elucidate the molecular architecture underlying this unique phenomenon by exploring potential reaction mechanisms, facilitated by an artificial intelligence prediction system. Artificial intelligence predictor introduces a novel approach to interlinking two polymers, involving the formation of networks interconnected with linear chains following random chain scission. This novel configuration leads to the emergence of a distinct type of hydrogel, herein referred to as a "Span Network." Additionally, Fourier-transform infrared spectroscopy (FTIR) is used to investigate functional groups that may be implicated in the proposed mechanism, with ester formation confirmed among numerous hydroxyl end groups obtained from chain scission of PEO and carboxyl groups formed on hydrogel networks.
Generation of simulated detector response to collision products is crucial to data analysis in particle physics, but computationally very expensive. One subdetector, the calorimeter, dominates the computational time due to the high granularity of its cells and complexity of the interactions. Generative models can provide more rapid sample production, but currently require significant effort to optimize performance for specific detector geometries, often requiring many models to describe the varying cell sizes and arrangements, without the ability to generalize to other geometries. We develop a $\textit{geometry-aware}$ autoregressive model, which learns how the calorimeter response varies with geometry, and is capable of generating simulated responses to unseen geometries without additional training. The geometry-aware model outperforms a baseline unaware model by over $50\%$ in several metrics such as the Wasserstein distance between the generated and the true distributions of key quantities which summarize the simulated response. A single geometry-aware model could replace the hundreds of generative models currently designed for calorimeter simulation by physicists analyzing data collected at the Large Hadron Collider. This proof-of-concept study motivates the design of a foundational model that will be a crucial tool for the study of future detectors, dramatically reducing the large upfront investment usually needed to develop generative calorimeter models.
Alexander Shmakov, Kevin Greif, Michael Fenton et al.
High-energy collisions at the Large Hadron Collider (LHC) provide valuable insights into open questions in particle physics. However, detector effects must be corrected before measurements can be compared to certain theoretical predictions or measurements from other detectors. Methods to solve this \textit{inverse problem} of mapping detector observations to theoretical quantities of the underlying collision are essential parts of many physics analyses at the LHC. We investigate and compare various generative deep learning methods to approximate this inverse mapping. We introduce a novel unified architecture, termed latent variation diffusion models, which combines the latent learning of cutting-edge generative art approaches with an end-to-end variational framework. We demonstrate the effectiveness of this approach for reconstructing global distributions of theoretical kinematic quantities, as well as for ensuring the adherence of the learned posterior distributions to known physics constraints. Our unified approach achieves a distribution-free distance to the truth of over 20 times less than non-latent state-of-the-art baseline and 3 times less than traditional latent diffusion models.
Attention plays a fundamental role in both natural and artificial intelligence systems. In deep learning, attention-based neural architectures, such as transformer architectures, are widely used to tackle problems in natural language processing and beyond. Here we investigate the fundamental building blocks of attention and their computational properties. Within the standard model of deep learning, we classify all possible fundamental building blocks of attention in terms of their source, target, and computational mechanism. We identify and study three most important mechanisms: additive activation attention, multiplicative output attention (output gating), and multiplicative synaptic attention (synaptic gating). The gating mechanisms correspond to multiplicative extensions of the standard model and are used across all current attention-based deep learning architectures. We study their functional properties and estimate the capacity of several attentional building blocks in the case of linear and polynomial threshold gates. Surprisingly, additive activation attention plays a central role in the proofs of the lower bounds. Attention mechanisms reduce the depth of certain basic circuits and leverage the power of quadratic activations without incurring their full cost.
Policy space response oracles (PSRO) is a multi-agent reinforcement learning algorithm that has achieved state-of-the-art performance in very large two-player zero-sum games. PSRO is based on the tabular double oracle (DO) method, an algorithm that is guaranteed to converge to a Nash equilibrium, but may increase exploitability from one iteration to the next. We propose anytime double oracle (ADO), a tabular double oracle algorithm for 2-player zero-sum games that is guaranteed to converge to a Nash equilibrium while decreasing exploitability from one iteration to the next. Unlike DO, in which the restricted distribution is based on the restricted game formed by each player's strategy sets, ADO finds the restricted distribution for each player that minimizes its exploitability against any policy in the full, unrestricted game. We also propose a method of finding this restricted distribution via a no-regret algorithm updated against best responses, called RM-BR DO. Finally, we propose anytime PSRO (APSRO), a version of ADO that calculates best responses via reinforcement learning. In experiments on Leduc poker and random normal form games, we show that our methods achieve far lower exploitability than DO and PSRO and decrease exploitability monotonically.
Mohammadamin Tavakoli, Alexander Shmakov, Francesco Ceccarelli et al.
It is fundamental for science and technology to be able to predict chemical reactions and their properties. To achieve such skills, it is important to develop good representations of chemical reactions, or good deep learning architectures that can learn such representations automatically from the data. There is currently no universal and widely adopted method for robustly representing chemical reactions. Most existing methods suffer from one or more drawbacks, such as: (1) lacking universality; (2) lacking robustness; (3) lacking interpretability; or (4) requiring excessive manual pre-processing. Here we exploit graph-based representations of molecular structures to develop and test a hypergraph attention neural network approach to solve at once the reaction representation and property-prediction problems, alleviating the aforementioned drawbacks. We evaluate this hypergraph representation in three experiments using three independent data sets of chemical reactions. In all experiments, the hypergraph-based approach matches or outperforms other representations and their corresponding models of chemical reactions while yielding interpretable multi-level representations.
Mohammadamin Tavakoli, Peter Sadowski, Pierre Baldi
In a physical neural system, backpropagation is faced with a number of obstacles including: the need for labeled data, the violation of the locality learning principle, the need for symmetric connections, and the lack of modularity. Tourbillon is a new architecture that addresses all these limitations. At its core, it consists of a stack of circular autoencoders followed by an output layer. The circular autoencoders are trained in self-supervised mode by recirculation algorithms and the top layer in supervised mode by stochastic gradient descent, with the option of propagating error information through the entire stack using non-symmetric connections. While the Tourbillon architecture is meant primarily to address physical constraints, and not to improve current engineering applications of deep learning, we demonstrate its viability on standard benchmark datasets including MNIST, Fashion MNIST, and CIFAR10. We show that Tourbillon can achieve comparable performance to models trained with backpropagation and outperform models that are trained with other physically plausible algorithms, such as feedback alignment.
Stephen McAleer, John Lanier, Michael Dennis et al.
Machine learning algorithms often make decisions on behalf of agents with varied and sometimes conflicting interests. In domains where agents can choose to take their own action or delegate their action to a central mediator, an open question is how mediators should take actions on behalf of delegating agents. The main existing approach uses delegating agents to punish non-delegating agents in an attempt to get all agents to delegate, which tends to be costly for all. We introduce a Pareto Mediator which aims to improve outcomes for delegating agents without making any of them worse off. Our experiments in random normal form games, a restaurant recommendation game, and a reinforcement learning sequential social dilemma show that the Pareto Mediator greatly increases social welfare. Also, even when the Pareto Mediator is based on an incorrect model of agent utility, performance gracefully degrades to the pre-intervention level, due to the individual autonomy preserved by the voluntary mediator.
Mohammadamin Tavakoli, Aaron Mood, David Van Vranken et al.
There is a lack of scalable quantitative measures of reactivity for functional groups in organic chemistry. Measuring reactivity experimentally is costly and time-consuming and does not scale to the astronomical size of chemical space. In previous quantum chemistry studies, we have introduced Methyl Cation Affinities (MCA*) and Methyl Anion Affinities (MAA*), using a solvation model, as quantitative measures of reactivity for organic functional groups over the broadest range. Although MCA* and MAA* offer good estimates of reactivity parameters, their calculation through Density Functional Theory (DFT) simulations is time-consuming. To circumvent this problem, we first use DFT to calculate MCA* and MAA* for more than 2,400 organic molecules thereby establishing a large dataset of chemical reactivity scores. We then design deep learning methods to predict the reactivity of molecular structures and train them using this curated dataset in combination with different representations of molecular structures. Using ten-fold cross-validation, we show that graph attention neural networks applied to informative input fingerprints produce the most accurate estimates of reactivity, achieving over 91% test accuracy for predicting the MCA* plus-minus 3.0 or MAA* plus-minus 3.0, over 50 orders of magnitude. Finally, we demonstrate the application of these reactivity scores to two tasks: (1) chemical reaction prediction; (2) combinatorial generation of reaction mechanisms. The curated dataset of MCA* and MAA* scores is available through the ChemDB chemoinformatics web portal at www.cdb.ics.uci.edu.
Motivated by biological considerations, we study sparse neural maps from an input layer to a target layer with sparse activity, and specifically the problem of storing $K$ input-target associations $(x,y)$, or memories, when the target vectors $y$ are sparse. We mathematically prove that $K$ undergoes a phase transition and that in general, and somewhat paradoxically, sparsity in the target layers increases the storage capacity of the map. The target vectors can be chosen arbitrarily, including in random fashion, and the memories can be both encoded and decoded by networks trained using local learning rules, including the simple Hebb rule. These results are robust under a variety of statistical assumptions on the data. The proofs rely on elegant properties of random polytopes and sub-gaussian random vector variables. Open problems and connections to capacity theories and polynomial threshold maps are discussed.
Forest Agostinelli, Shahaf S. Shperberg, Alexander Shmakov et al.
Efficiently solving problems with large action spaces using A* search remains a significant challenge. This is because, for each iteration of A* search, the number of nodes generated and the number of heuristic function applications grow linearly with the size of the action space. This burden becomes even more apparent when A* search uses a heuristic function learned by computationally expensive function approximators, such as deep neural networks. To address this issue, we introduce Q*, a search algorithm that leverages heuristics capable of receiving a state and, in a single function call, returning cost-to-go estimates for all possible transitions from that state, along with estimates of the corresponding transition costs -- without the need to apply the transitions or generate the successor states; such action-state estimation are typically known as Q-values. This significantly reduces computation time and memory usage. In addition, we prove that Q* search is guaranteed to find a shortest path given a heuristic function that does not overestimate the sum of the transition cost and cost-to-go of the state. To obtain heuristics for Q* search, we employ a deep Q-network architecture to learn a state-action heuristic function from domain interaction, without any prior knowledge. We use Q* with our learned heuristic on different domains and action spaces, showing that Q* suffers from only a small runtime overhead as the size of the action space increases. In addition, our empirical results show Q* search is up to 129 times faster and generates up to 1288 times fewer nodes than A* search.
Coccidioidomycosis is the most common systemic mycosis in dogs in the southwestern United States. With warming climates, affected areas and number of cases are expected to increase in the coming years, escalating also the chances of transmission to humans. As a result, developing methods for automating the detection of the disease is important, as this will help doctors and veterinarians more easily identify and diagnose positive cases. We apply machine learning models to provide accurate and interpretable predictions of Coccidioidomycosis. We assemble a set of radiographic images and use it to train and test state-of-the-art convolutional neural networks to detect Coccidioidomycosis. These methods are relatively inexpensive to train and very fast at inference time. We demonstrate the successful application of this approach to detect the disease with an Area Under the Curve (AUC) above 0.99 using 10-fold cross validation. We also use the classification model to identify regions of interest and localize the disease in the radiographic images, as illustrated through visual heatmaps. This proof-of-concept study establishes the feasibility of very accurate and rapid automated detection of Valley Fever in radiographic images.
In the framework of three-active-neutrino mixing, the charge parity phase, the neutrino mass ordering, and the octant of $θ_{23}$ remain unknown. The Deep Underground Neutrino Experiment (DUNE) is a next-generation long-baseline neutrino oscillation experiment, which aims to address these questions by measuring the oscillation patterns of $ν_μ/ν_e$ and $\barν_μ/\barν_e$ over a range of energies spanning the first and second oscillation maxima. DUNE far detector modules are based on liquid argon TPC (LArTPC) technology. A LArTPC offers excellent spatial resolution, high neutrino detection efficiency, and superb background rejection, while reconstruction in LArTPC is challenging. Deep learning methods, in particular, Convolutional Neural Networks (CNNs), have demonstrated success in classification problems such as particle identification in DUNE and other neutrino experiments. However, reconstruction of neutrino energy and final state particle momenta with deep learning methods is yet to be developed for a full AI-based reconstruction chain. To precisely reconstruct these kinematic characteristics of detected interactions at DUNE, we have developed and will present two CNN-based methods, 2-D and 3-D, for the reconstruction of final state particle direction and energy, as well as neutrino energy. Combining particle masses with the kinetic energy and the direction reconstructed by our work, the four-momentum of final state particles can be obtained. Our models show considerable improvements compared to the traditional methods for both scenarios.
Two-photon excitation fluorescence (2PEF) allows imaging of tissue up to about one millimeter in thickness. Typically, reducing fluorescence excitation exposure reduces the quality of the image. However, using deep learning super resolution techniques, these low-resolution images can be converted to high-resolution images. This work explores improving human tissue imaging by applying deep learning to maximize image quality while reducing fluorescence excitation exposure. We analyze two methods: a method based on U-Net, and a patch-based regression method. Both methods are evaluated on a skin dataset and an eye dataset. The eye dataset includes 1200 paired high power and low power images of retinal organoids. The skin dataset contains multiple frames of each sample of human skin. High-resolution images were formed by averaging 70 frames for each sample and low-resolution images were formed by averaging the first 7 and 15 frames for each sample. The skin dataset includes 550 images for each of the resolution levels. We track two measures of performance for the two methods: mean squared error (MSE) and structural similarity index measure (SSIM). For the eye dataset, the patches method achieves an average MSE of 27,611 compared to 146,855 for the U-Net method, and an average SSIM of 0.636 compared to 0.607 for the U-Net method. For the skin dataset, the patches method achieves an average MSE of 3.768 compared to 4.032 for the U-Net method, and an average SSIM of 0.824 compared to 0.783 for the U-Net method. Despite better performance on image quality, the patches method is worse than the U-Net method when comparing the speed of prediction, taking 303 seconds to predict one image compared to less than one second for the U-Net method.
Michael James Fenton, Alexander Shmakov, Ta-Wei Ho et al.
Top quarks, produced in large numbers at the Large Hadron Collider, have a complex detector signature and require special reconstruction techniques. The most common decay mode, the "all-jet" channel, results in a 6-jet final state which is particularly difficult to reconstruct in $pp$ collisions due to the large number of permutations possible. We present a novel approach to this class of problem, based on neural networks using a generalized attention mechanism, that we call Symmetry Preserving Attention Networks (SPA-Net). We train one such network to identify the decay products of each top quark unambiguously and without combinatorial explosion as an example of the power of this technique.This approach significantly outperforms existing state-of-the-art methods, correctly assigning all jets in $93.0%$ of $6$-jet, $87.8%$ of $7$-jet, and $82.6%$ of $\geq 8$-jet events respectively.
Reinforcement learning algorithms can show strong variation in performance between training runs with different random seeds. In this paper we explore how this affects hyperparameter optimization when the goal is to find hyperparameter settings that perform well across random seeds. In particular, we benchmark whether it is better to explore a large quantity of hyperparameter settings via pruning of bad performers, or if it is better to aim for quality of collected results by using repetitions. For this we consider the Successive Halving, Random Search, and Bayesian Optimization algorithms, the latter two with and without repetitions. We apply these to tuning the PPO2 algorithm on the Cartpole balancing task and the Inverted Pendulum Swing-up task. We demonstrate that pruning may negatively affect the optimization and that repeated sampling does not help in finding hyperparameter settings that perform better across random seeds. From our experiments we conclude that Bayesian optimization with a noise robust acquisition function is the best choice for hyperparameter optimization in reinforcement learning tasks.
Mohammadamin Tavakoli, Forest Agostinelli, Pierre Baldi
We introduce SPLASH units, a class of learnable activation functions shown to simultaneously improve the accuracy of deep neural networks while also improving their robustness to adversarial attacks. SPLASH units have both a simple parameterization and maintain the ability to approximate a wide range of non-linear functions. SPLASH units are: 1) continuous; 2) grounded (f(0) = 0); 3) use symmetric hinges; and 4) the locations of the hinges are derived directly from the data (i.e. no learning required). Compared to nine other learned and fixed activation functions, including ReLU and its variants, SPLASH units show superior performance across three datasets (MNIST, CIFAR-10, and CIFAR-100) and four architectures (LeNet5, All-CNN, ResNet-20, and Network-in-Network). Furthermore, we show that SPLASH units significantly increase the robustness of deep neural networks to adversarial attacks. Our experiments on both black-box and open-box adversarial attacks show that commonly-used architectures, namely LeNet5, All-CNN, ResNet-20, and Network-in-Network, can be up to 31% more robust to adversarial attacks by simply using SPLASH units instead of ReLUs.
In order to continuously represent molecules, we propose a generative model in the form of a VAE which is operating on the 2D-graph structure of molecules. A side predictor is employed to prune the latent space and help the decoder in generating meaningful adjacency tensor of molecules. Other than the potential applicability in drug design and property prediction, we show the superior performance of this technique in comparison to other similar methods based on the SMILES representation of the molecules with RNN based encoder and decoder.
Implementing artificial neural networks is commonly achieved via high-level programming languages like Python and easy-to-use deep learning libraries like Keras. These software libraries come pre-loaded with a variety of network architectures, provide autodifferentiation, and support GPUs for fast and efficient computation. As a result, a deep learning practitioner will favor training a neural network model in Python, where these tools are readily available. However, many large-scale scientific computation projects are written in Fortran, making it difficult to integrate with modern deep learning methods. To alleviate this problem, we introduce a software library, the Fortran-Keras Bridge (FKB). This two-way bridge connects environments where deep learning resources are plentiful, with those where they are scarce. The paper describes several unique features offered by FKB, such as customizable layers, loss functions, and network ensembles. The paper concludes with a case study that applies FKB to address open questions about the robustness of an experimental approach to global climate simulation, in which subgrid physics are outsourced to deep neural network emulators. In this context, FKB enables a hyperparameter search of one hundred plus candidate models of subgrid cloud and radiation physics, initially implemented in Keras, to be transferred and used in Fortran. Such a process allows the model's emergent behavior to be assessed, i.e. when fit imperfections are coupled to explicit planetary-scale fluid dynamics. The results reveal a previously unrecognized strong relationship between offline validation error and online performance, in which the choice of optimizer proves unexpectedly critical. This reveals many neural network architectures that produce considerable improvements in stability including some with reduced error, for an especially challenging training dataset.
Weight-sharing is one of the pillars behind Convolutional Neural Networks and their successes. However, in physical neural systems such as the brain, weight-sharing is implausible. This discrepancy raises the fundamental question of whether weight-sharing is necessary. If so, to which degree of precision? If not, what are the alternatives? The goal of this study is to investigate these questions, primarily through simulations where the weight-sharing assumption is relaxed. Taking inspiration from neural circuitry, we explore the use of Free Convolutional Networks and neurons with variable connection patterns. Using Free Convolutional Networks, we show that while weight-sharing is a pragmatic optimization approach, it is not a necessity in computer vision applications. Furthermore, Free Convolutional Networks match the performance observed in standard architectures when trained using properly translated data (akin to video). Under the assumption of translationally augmented data, Free Convolutional Networks learn translationally invariant representations that yield an approximate form of weight sharing.
Dealing with sparse rewards is a longstanding challenge in reinforcement learning. The recent use of hindsight methods have achieved success on a variety of sparse-reward tasks, but they fail on complex tasks such as stacking multiple blocks with a robot arm in simulation. Curiosity-driven exploration using the prediction error of a learned dynamics model as an intrinsic reward has been shown to be effective for exploring a number of sparse-reward environments. We present a method that combines hindsight with curiosity-driven exploration and curriculum learning in order to solve the challenging sparse-reward block stacking task. We are the first to stack more than two blocks using only sparse reward without human demonstrations.
A long standing open problem in the theory of neural networks is the development of quantitative methods to estimate and compare the capabilities of different architectures. Here we define the capacity of an architecture by the binary logarithm of the number of functions it can compute, as the synaptic weights are varied. The capacity provides an upper bound on the number of bits that can be extracted from the training data and stored in the architecture during learning. We study the capacity of layered, fully-connected, architectures of linear threshold neurons with $L$ layers of size $n_1,n_2, \ldots, n_L$ and show that in essence the capacity is given by a cubic polynomial in the layer sizes: $C(n_1,\ldots, n_L)=\sum_{k=1}^{L-1} \min(n_1,\ldots,n_k)n_kn_{k+1}$, where layers that are smaller than all previous layers act as bottlenecks. In proving the main result, we also develop new techniques (multiplexing, enrichment, and stacking) as well as new bounds on the capacity of finite sets. We use the main result to identify architectures with maximal or minimal capacity under a number of natural constraints. This leads to the notion of structural regularization for deep architectures. While in general, everything else being equal, shallow networks compute more functions than deep networks, the functions computed by deep networks are more regular and "interesting".
The unified approach of Feldman and Cousins allows for exact statistical inference of small signals that commonly arise in high energy physics. It has gained widespread use, for instance, in measurements of neutrino oscillation parameters in long-baseline experiments. However, the approach relies on the Neyman construction of the classical confidence interval and is computationally intensive as it is typically done in a grid-based fashion over the entire parameter space. In this letter, we propose an efficient algorithm for the Feldman-Cousins approach using Gaussian processes to construct confidence intervals iteratively. We show that in the neutrino oscillation context, one can obtain confidence intervals 5 times faster in one dimension and 10 times faster in two dimensions, while maintaining an accuracy above 99.5%.
Stephen McAleer, Forest Agostinelli, Alexander Shmakov et al.
A generally intelligent agent must be able to teach itself how to solve problems in complex domains with minimal human supervision. Recently, deep reinforcement learning algorithms combined with self-play have achieved superhuman proficiency in Go, Chess, and Shogi without human data or domain knowledge. In these environments, a reward is always received at the end of the game, however, for many combinatorial optimization environments, rewards are sparse and episodes are not guaranteed to terminate. We introduce Autodidactic Iteration: a novel reinforcement learning algorithm that is able to teach itself how to solve the Rubik's Cube with no human assistance. Our algorithm is able to solve 100% of randomly scrambled cubes while achieving a median solve length of 30 moves -- less than or equal to solvers that employ human domain knowledge.
In a physical neural system, learning rules must be local both in space and time. In order for learning to occur, non-local information must be communicated to the deep synapses through a communication channel, the deep learning channel. We identify several possible architectures for this learning channel (Bidirectional, Conjoined, Twin, Distinct) and six symmetry challenges: 1) symmetry of architectures; 2) symmetry of weights; 3) symmetry of neurons; 4) symmetry of derivatives; 5) symmetry of processing; and 6) symmetry of learning rules. Random backpropagation (RBP) addresses the second and third symmetry, and some of its variations, such as skipped RBP (SRBP) address the first and the fourth symmetry. Here we address the last two desirable symmetries showing through simulations that they can be achieved and that the learning channel is particularly robust to symmetry variations. Specifically, random backpropagation and its variations can be performed with the same non-linear neurons used in the main input-output forward channel, and the connections in the learning channel can be adapted using the same algorithm used in the forward channel, removing the need for any specialized hardware in the learning channel. Finally, we provide mathematical results in simple cases showing that the learning equations in the forward and backward channels converge to fixed points, for almost any initial conditions. In symmetric architectures, if the weights in both channels are small at initialization, adaptation in both channels leads to weights that are essentially symmetric during and after learning. Biological connections are discussed.
Antihydrogen is at the forefront of antimatter research at the CERN Antiproton Decelerator. Experiments aiming to test the fundamental CPT symmetry and antigravity effects require the efficient detection of antihydrogen annihilation events, which is performed using highly granular tracking detectors installed around an antimatter trap. Improving the efficiency of the antihydrogen annihilation detection plays a central role in the final sensitivity of the experiments. We propose deep learning as a novel technique to analyze antihydrogen annihilation data, and compare its performance with a traditional track and vertex reconstruction method. We report that the deep learning approach yields significant improvement, tripling event coverage while simultaneously improving performance by over 5% in terms of Area Under Curve (AUC).
Chase Shimmin, Peter Sadowski, Pierre Baldi et al.
We describe a strategy for constructing a neural network jet substructure tagger which powerfully discriminates boosted decay signals while remaining largely uncorrelated with the jet mass. This reduces the impact of systematic uncertainties in background modeling while enhancing signal purity, resulting in improved discovery significance relative to existing taggers. The network is trained using an adversarial strategy, resulting in a tagger that learns to balance classification accuracy with decorrelation. As a benchmark scenario, we consider the case where large-radius jets originating from a boosted resonance decay are discriminated from a background of nonresonant quark and gluon jets. We show that in the presence of systematic uncertainties on the background rate, our adversarially-trained, decorrelated tagger considerably outperforms a conventionally trained neural network, despite having a slightly worse signal-background separation power. We generalize the adversarial training technique to include a parametric dependence on the signal hypothesis, training a single network that provides optimized, interpolatable decorrelated jet tagging across a continuous range of hypothetical resonance masses, after training on discrete choices of the signal mass.
Random backpropagation (RBP) is a variant of the backpropagation algorithm for training neural networks, where the transpose of the forward matrices are replaced by fixed random matrices in the calculation of the weight updates. It is remarkable both because of its effectiveness, in spite of using random matrices to communicate error information, and because it completely removes the taxing requirement of maintaining symmetric weights in a physical neural system. To better understand random backpropagation, we first connect it to the notions of local learning and learning channels. Through this connection, we derive several alternatives to RBP, including skipped RBP (SRPB), adaptive RBP (ARBP), sparse RBP, and their combinations (e.g. ASRBP) and analyze their computational complexity. We then study their behavior through simulations using the MNIST and CIFAR-10 bechnmark datasets. These simulations show that most of these variants work robustly, almost as well as backpropagation, and that multiplication by the derivatives of the activation functions is important. As a follow-up, we study also the low-end of the number of bits required to communicate error information over the learning channel. We then provide partial intuitive explanations for some of the remarkable properties of RBP and its variations. Finally, we prove several mathematical results, including the convergence to fixed points of linear chains of arbitrary length, the convergence to fixed points of linear autoencoders with decorrelated data, the long-term existence of solutions for linear systems with a single hidden layer and convergence in special cases, and the convergence to fixed points of non-linear chains, when the derivative of the activation functions is included.
Experiments in particle physics produce enormous quantities of data that must be analyzed and interpreted by teams of physicists. This analysis is often exploratory, where scientists are unable to enumerate the possible types of signal prior to performing the experiment. Thus, tools for summarizing, clustering, visualizing and classifying high-dimensional data are essential. In this work, we show that meaningful physical content can be revealed by transforming the raw data into a learned high-level representation using deep neural networks, with measurements taken at the Daya Bay Neutrino Experiment as a case study. We further show how convolutional deep neural networks can provide an effective classification filter with greater than 97% accuracy across different classes of physics events, significantly better than other machine learning approaches.
In a physical neural system, where storage and processing are intimately intertwined, the rules for adjusting the synaptic weights can only depend on variables that are available locally, such as the activity of the pre- and post-synaptic neurons, resulting in local learning rules. A systematic framework for studying the space of local learning rules is obtained by first specifying the nature of the local variables, and then the functional form that ties them together into each learning rule. Such a framework enables also the systematic discovery of new learning rules and exploration of relationships between learning rules and group symmetries. We study polynomial local learning rules stratified by their degree and analyze their behavior and capabilities in both linear and non-linear units and networks. Stacking local learning rules in deep feedforward networks leads to deep local learning. While deep local learning can learn interesting representations, it cannot learn complex input-output functions, even when targets are available for the top layer. Learning complex input-output functions requires local deep learning where target information is communicated to the deep layers through a backward learning channel. The nature of the communicated information about the targets and the structure of the learning channel partition the space of learning algorithms. We estimate the learning channel capacity associated with several algorithms and show that backpropagation outperforms them by simultaneously maximizing the information rate and minimizing the computational cost, even in recurrent networks. The theory clarifies the concept of Hebbian learning, establishes the power and limitations of local learning rules, introduces the learning channel which enables a formal analysis of the optimality of backpropagation, and explains the sparsity of the space of learning rules discovered so far.